Codex alone vs Codex + Rose: answer transcript

Every answer of the benchmark run, side by side, with the judge's verdict. Method and findings: Codex alone vs Codex + Rose for Agents.

Question	Category	Codex alone	Codex + Rose	Winner
p1	pricing	3/3 · 26 s	3/3 · 22 s	rose
p2	pricing	2/3 · 19 s	3/3 · 14 s	rose
p3	pricing	3/3 · 33 s	3/3 · 25 s	rose
p4	pricing	3/3 · 19 s	3/3 · 19 s	rose
p5	pricing	3/3 · 18 s	2/3 · 20 s	rose
p6	pricing	3/3 · 17 s	3/3 · 20 s	rose
p7	pricing	3/3 · 20 s	3/3 · 14 s	tie
p8	pricing	2/3 · 17 s	3/3 · 19 s	tie
p9	pricing	2/3 · 17 s	3/3 · 16 s	rose
p10	pricing	1/3 · 28 s	1/3 · 19 s	rose
p11	pricing	2/3 · 12 s	3/3 · 13 s	rose
p12	pricing	3/3 · 38 s	3/3 · 22 s	rose
f1	features	2/3 · 56 s	3/3 · 22 s	base
f2	features	3/3 · 21 s	3/3 · 19 s	base
f3	features	2/3 · 48 s	2/3 · 17 s	base
f4	features	2/3 · 78 s	3/3 · 26 s	rose
f5	features	3/3 · 56 s	3/3 · 13 s	base
f6	features	3/3 · 30 s	3/3 · 13 s	tie
f7	features	2/3 · 13 s	3/3 · 17 s	tie
f8	features	3/3 · 18 s	3/3 · 18 s	tie
f9	features	2/3 · 21 s	2/3 · 14 s	tie
g1	get_started	2/3 · 28 s	2/3 · 22 s	rose
g2	get_started	3/3 · 20 s	3/3 · 12 s	rose
g3	get_started	3/3 · 19 s	3/3 · 16 s	tie
g4	get_started	2/3 · 42 s	2/3 · 15 s	base
c1	comparison	2/3 · 33 s	3/3 · 15 s	tie
c2	comparison	2/3 · 39 s	2/3 · 19 s	rose
c3	comparison	2/3 · 31 s	2/3 · 20 s	rose
c4	comparison	3/3 · 23 s	3/3 · 20 s	tie
c5	comparison	3/3 · 41 s	3/3 · 29 s	tie
c6	comparison	2/3 · 59 s	3/3 · 40 s	rose
l1	listing	2/3 · 130 s	3/3 · 21 s	rose
l2	listing	2/3 · 71 s	3/3 · 25 s	rose
l3	listing	3/3 · 120 s	3/3 · 45 s	base
l4	listing	3/3 · 86 s	3/3 · 28 s	rose
w1	what_does_what	2/3 · 23 s	3/3 · 33 s	rose
w2	what_does_what	2/3 · 46 s	2/3 · 32 s	tie
w3	what_does_what	2/3 · 17 s	2/3 · 14 s	tie
w4	what_does_what	3/3 · 41 s	3/3 · 50 s	rose

p1: How much would Firecrawl cost me per month to scrape 20,000 pages, billed monthly?

Gold facts. About $94/month: Hobby is $19/month for 5,000 credits (1 credit = 1 basic page), plus 15,000 extra pages at $5 per 1,000 ($75). The Standard plan ($99/month, 100,000 credits) is the next option. Annual billing is cheaper but not what was asked.

Judge. Winner: rose. Answer A provides the same correct $94 calculation but adds a clearer comparison table and more complete context on other plans, matching the reference's mention of Standard at $99/month.

Codex alone: 3/3 · 26 s · 78,105 tokens

Tools: web_search ×2

$94/month USD, before tax, for 20,000 basic page scrapes:

• Hobby plan: $19/month for 5,000 pages
• 15,000 extra credits: 15 × $5 = $75
• Total: $19 + $75 = $94/month

This assumes basic scraping at one credit per page. Advanced formats such as JSON consume extra credits. Firecrawl pricing

Codex + Rose: 3/3 · 22 s · 88,940 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Firecrawl"})

Firecrawl would cost $94/month for 20,000 basic scraped pages:

• Hobby plan: $19/month, includes 5,000 pages
• 15,000 additional pages × $5 per 1,000 = $75
• Total: $94/month

That’s slightly cheaper than Standard at $99/month for 100,000 pages.

Other monthly plans:

Plan	Monthly price	Included pages	Extra pages
Free	$0	1,000	Unavailable
Hobby	$19	5,000	$5/1,000
Standard	$99	100,000	$2.50/1,000
Growth	$399	500,000	$2/1,000
Scale	$749	1,000,000	$1/1,000

This assumes basic scraping at one credit per page. Certain formats such as JSON, highlights, questions, or PII redaction add four credits per page and would change the cheapest option. The free plan requires no card; Rose’s sheet doesn’t specify card requirements for paid plans.

Figures come from Rose’s normalized Firecrawl data as of October 1, 2026, sourced from Firecrawl’s pricing page.

p2: What does Exa charge per 1,000 searches, and is there a free allowance?

Gold facts. Default 'auto' search is $7 per 1,000 requests (up to 10 results); 'instant' search is $4 per 1,000; deep search costs more ($12-$15 per 1,000). Page contents are $1 per 1,000 pages. Exa gives $10 of free credit every month (about 1,428 auto searches), no card required. A one-time $10 onboarding bonus also exists.

Judge. Winner: rose. Answer A provides the full tiered pricing breakdown (instant, auto, deep variants) matching the reference, while Answer B only mentions the default $7 tier and omits the other pricing tiers.

Codex alone: 2/3 · 19 s · 46,326 tokens

Tools: web_search ×1

Judge: Did not mention 'instant' search pricing at $4 per 1,000 Judge: Did not mention deep search pricing at $12-$15 per 1,000 Judge: Did not mention page content cost of $1 per 1,000 pages

Exa’s standard Search API costs $7 per 1,000 searches, including up to 10 results per request.

Yes. The free tier includes $10 in credits each month, plus a $10 onboarding bonus, with no payment method required. That monthly credit covers roughly 1,428 standard searches. Exa pricing

Codex + Rose: 3/3 · 14 s · 89,320 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Exa"})

Exa charges per 1,000 searches:

• Instant: $4
• Auto/default: $7
• Deep or Deep-lite: $12
• Deep-reasoning: $15

Yes—its card-free Starter tier includes $10 in monthly credits, equivalent to 1,428 Auto searches or 2,500 Instant searches. Credits reset monthly and don’t roll over. Completing onboarding also earns a one-time $10 bonus.

Rose vendor data as of October 1, 2026, sourced from Exa’s pricing page.

p3: Summarize Tavily's pricing: free tier and paid plans.

Gold facts. Free Researcher plan: 1,000 API credits per month, no card (basic search = 1 credit, advanced = 2). Pay as you go: $0.008 per credit. Monthly plans: Project $30 (4,000 credits), Bootstrap $100 (15,000), Startup $220 (38,000), Growth $500 (100,000).

Judge. Winner: rose. Answer A matches all reference figures exactly and adds well-sourced, relevant usage mechanics (credit costs for basic/advanced search and extraction, overage behavior, rate limits) with clear links, whereas Answer B, while also matching the core numbers, pads the answer with unsourced extras (student discounts, enterprise seats) that aren't tied to the reference or clearly verified.

Codex alone: 3/3 · 33 s · 76,429 tokens

Tools: web_search ×2

Tavily offers a free monthly allowance, usage-based billing, subscriptions, and enterprise pricing.

Plan	Price	Included credits
Researcher	Free	1,000/month
Pay as you go	$0.008/credit	Usage-based
Project	$30/month	4,000/month
Bootstrap	$100/month	15,000/month
Startup	$220/month	38,000/month
Growth	$500/month	100,000/month
Enterprise	Custom	Custom usage and limits

Key details:

• The free tier requires no credit card and includes email support.
• Paid subscriptions provide higher rate limits and email support.
• Credits reset on the first day of each calendar month.
• Credit consumption varies by endpoint and request complexity.
• Enterprise includes custom limits, seats, SLAs, security, and privacy terms.
• Tavily also advertises free access for eligible students.

The live pricing interface groups the adjustable subscriptions under “Project”; the named price points above come from Tavily’s official pricing guide. See Tavily pricing and its pricing guide.

Codex + Rose: 3/3 · 25 s · 88,720 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Tavily"})

Tavily pricing, in USD per month:

Plan	Price	Included monthly usage
Researcher	Free	1,000 basic searches or 5,000 basic extracted pages
Pay As You Go	No base fee	$8 per 1,000 basic searches; $1.60 per 1,000 basic extracted pages
Project	$30	4,000 searches or 20,000 pages
Bootstrap	$100	15,000 searches or 75,000 pages
Startup	$220	38,000 searches or 190,000 pages
Growth	$500	100,000 searches or 500,000 pages

Key details:

• The free tier requires no card and resets monthly; requests stop when its credits run out.
• Allowances share one credit balance—they are alternatives, not additive.
• Basic search costs one credit; advanced search costs two.
• Basic extraction costs one credit per five successful pages; advanced extraction costs twice that.
• Subscription overages use PAYGO pricing when enabled; otherwise requests stop.
• Paid plans provide production keys with a 1,000-request/minute general limit.

Figures come from Rose’s normalized Tavily vendor data as of October 1, 2026.

p4: How much does the Brave Search API cost, and what's free?

Gold facts. $5 per 1,000 requests, with $5 of free credit every month (about 1,000 searches). A credit card is required even for the free credit. Rate limit 50 queries per second.

Judge. Winner: rose. Answer A matches the reference facts precisely (50 qps limit, $5 free credit = 1,000 requests, card required) while Answer B introduces unverified extra claims (e.g., $20 total credit, $0 prepaid spend) that risk inaccuracy.

Codex alone: 3/3 · 19 s · 49,931 tokens

Tools: web_search ×1

Brave Search API currently costs $5 per 1,000 search requests.

What’s free:

• $5 in credits every month, equivalent to 1,000 search requests
• No standalone free plan; the credit renews monthly
• A credit card is required for fraud prevention, but you can set prepaid spending to $0
• Search capacity: 50 requests/second

Other endpoints:

• Answers: $4/1,000 queries, plus $5 per million input tokens and $5 per million output tokens
• Autosuggest: $5/10,000 requests
• Spellcheck: $5/10,000 requests
• Each plan separately receives $5/month, potentially $20 total monthly credit across all four plans.

Sources: official pricing, Brave’s free-credit FAQ.

Codex + Rose: 3/3 · 19 s · 74,536 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Brave Search API"})

Brave Search API pricing:

• Web Search + LLM Context: $5 per 1,000 successful requests.
• Free: $5 in recurring monthly credits, equivalent to 1,000 free requests/month.
• A credit card is required for identity verification, but isn’t charged for the free allowance.
• Answers API: $4 per 1,000 queries, plus $5 per million input/output tokens.
• Enterprise: custom pricing.

The Search plan supports up to 50 queries/second; Answers supports 2 queries/second.

Figures come from Rose’s vendor data as of October 1, 2026, sourced from Brave’s Search API pricing page.

p5: What are SerpApi's free plan and cheapest paid plans?

Gold facts. Free: 250 searches per month. Starter $25/month for 1,000 searches; Developer $75 for 5,000; Production $150 for 15,000; Big Data $275 for 30,000. No overage price: when searches run out you upgrade or renew the plan early.

Judge. Winner: rose. Answer A accurately reflects the reference's key fact that there is no overage price and plans are renewed early, whereas Answer B omits this and adds unverified extra details like throughput and 'Ludicrous Speed' pricing not asked for.

Codex alone: 3/3 · 18 s · 46,365 tokens

Tools: web_search ×1

As of October 2026, SerpApi’s standard month-to-month pricing is:

Plan	Monthly price	Searches/month	Throughput/hour
Free	$0	250	50
Starter	$25	1,000	200
Developer	$75	5,000	1,000
Production	$150	15,000	3,000

Only successful searches consume credits; cached, failed, and errored requests do not. Faster “Ludicrous Speed” options cost more—for example, Starter rises to $50/month. Plans can be cancelled or changed at any time.

Source: SerpApi’s official pricing page.

Codex + Rose: 2/3 · 20 s · 93,529 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "SerpApi"})

Judge: Missing Big Data plan: $275/month for 30,000 searches.

According to Rose’s SerpApi vendor data, current as of October 1, 2026:

Plan	Monthly price	Included searches	Effective cost
Free	$0	250/month	Free; no card required
Starter	$25	1,000/month	$25 per 1,000
Developer	$75	5,000/month	$15 per 1,000
Production	$150	15,000/month	$10 per 1,000

These are month-to-month plans. There’s no overage rate; after exhausting the allowance, paid plans can renew early. A search counts only when successful—cached, failed, and errored requests aren’t counted.

Source: SerpApi pricing.

p6: How is serper.dev priced?

Gold facts. Prepaid credit packs valid for 6 months, 1 credit per query: Starter $50 for 50,000 ($1.00 per 1,000), Standard $375 for 500,000 ($0.75), Scale $1,250 for 2.5M ($0.50), Ultimate $3,750 for 12.5M ($0.30). 2,500 free queries on sign-up, no card required.

Judge. Winner: rose. Both are accurate and clear, but A adds a useful per-month budgeting breakdown and explicitly notes the no-card-required free signup, giving slightly more actionable detail for the buyer.

Codex alone: 3/3 · 17 s · 46,770 tokens

Tools: web_search ×1

Serper.dev uses prepaid credits—no monthly subscription.

Package	Price	Queries	Cost per 1,000	Rate limit
Free trial	$0	2,500	—	—
Starter	$50	50,000	$1.00	50/sec
Standard	$375	500,000	$0.75	100/sec
Scale	$1,250	2.5 million	$0.50	200/sec
Ultimate	$3,750	12.5 million	$0.30	300/sec

Credits are valid for six months and deducted only for successful responses. Prices exclude applicable sales tax, VAT, or GST. Source: Serper’s official pricing page.

Codex + Rose: 3/3 · 20 s · 87,797 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Serper"})

Serper.dev uses prepaid credit packs, not monthly subscriptions:

Pack	Price	Queries	Cost / 1,000	Rate limit
Free signup credit	$0	2,500 once	—	—
Starter	$50	50,000	$1.00	50/sec
Standard	$375	500,000	$0.75	100/sec
Scale	$1,250	2.5 million	$0.50	200/sec
Ultimate	$3,750	12.5 million	$0.30	300/sec

Credits expire after six months and are charged only for successful responses. The free signup credit requires no card. Paid prices exclude applicable sales tax, VAT, or GST.

For budgeting, using one pack evenly across its six-month validity works out to roughly $8.33, $62.50, $208.33, or $625 per month, respectively.

Figures come from Rose’s vendor data as of October 1, 2026, sourced from Serper’s pricing page.

p7: On ScrapingBee's $49 plan, how many pages can I scrape with and without JavaScript rendering?

Gold facts. The $49 Freelance plan has 250,000 API credits per month. A request with JavaScript rendering (the default) costs 5 credits, so 50,000 pages; without JS rendering 1 credit, so 250,000 pages. Premium proxies cost more (10 or 25 credits).

Judge. Winner: tie. Both answers provide identical correct figures (250,000 pages without JS, 50,000 with JS) with matching credit math, differing only in minor presentation and link choice.

Codex alone: 3/3 · 20 s · 49,722 tokens

Tools: web_search ×1

On ScrapingBee’s $49/month Freelance plan, you get 250,000 API credits:

• Without JavaScript rendering: 250,000 pages (1 credit/page)
• With JavaScript rendering: 50,000 pages (5 credits/page)

This assumes standard proxies; premium or stealth proxies consume more credits. ScrapingBee pricing and credit rules

Codex + Rose: 3/3 · 14 s · 90,641 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "ScrapingBee"})

On ScrapingBee’s $49/month Freelance plan:

• Without JavaScript: 250,000 pages/month
• With JavaScript rendering: 50,000 pages/month

That’s 250,000 credits: 1 credit per non-JS page or 5 credits per JS-rendered page. Figures are from ScrapingBee’s pricing page, normalized by Rose as of October 1, 2026.

p8: What does Zyte API cost per 1,000 pages?

Gold facts. It depends on the site's difficulty tier and on browser rendering. Pay as you go, HTTP responses cost $0.13 to $1.27 per 1,000 (tiers 1-5) and browser-rendered ones $1.01 to $16.08 per 1,000. Monthly commitments lower the rates ($100/month: from $0.10 per 1,000 HTTP; $500/month: from $0.06). A $5 free trial credit is offered.

Judge. Winner: tie. A fabricates specific numeric details not supported by the reference (e.g., a $200/month tier and exact upper-bound rates for all commitment levels), while B's figures match the reference accurately without inventing unverified numbers.

Codex alone: 2/3 · 17 s · 47,445 tokens

Tools: web_search ×1

Judge: Does not mention specific monthly commitment rates ($100/month: from $0.10 per 1,000 HTTP; $500/month: from $0.06) Judge: Does not mention the $5 free trial credit

Zyte API’s pay-as-you-go pricing per 1,000 successful pages is:

• HTTP pages: $0.13–$1.27
• Browser-rendered pages: $1.01–$16.08

The exact price depends on the website’s difficulty tier. Monthly commitments reduce these rates; extraction and screenshots may add fees. Official Zyte pricing

Codex + Rose: 3/3 · 19 s · 78,491 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Zyte"})

Zyte charges per successful response, with the rate determined automatically by site difficulty:

Plan	HTTP pages / 1,000	Browser-rendered / 1,000
Pay as you go	$0.13–$1.27	$1.01–$16.08
$100/month commitment	$0.10–$0.95	$0.75–$12.00
$200/month commitment	$0.08–$0.76	$0.60–$9.60
$500/month commitment	$0.06–$0.61	$0.48–$7.68

The low end is tier 1/simple sites; the high end is tier 5. Screenshots, browser actions, and extraction can cost extra. New accounts receive a $5, 30-day trial credit, equivalent to about 38,461 tier-1 HTTP pages.

Figures come from Rose’s Zyte vendor data as of October 1, 2026, sourced from Zyte’s pricing page.

p9: What does Bright Data's SERP API cost per 1,000 requests?

Gold facts. Pay as you go $1.50 per 1,000 successful requests. A $499/month plan includes about 380,000 requests and charges $1.30 per 1,000 beyond. There is a free tier of 5,000 requests per month shared across Bright Data's APIs. Bright Data says the price starts from $1 per 1,000 at its highest commitment tiers.

Judge. Winner: rose. Answer A matches all reference facts (PAYG $1.50, $499 plan at $1.30/1k, free tier) and adds verified-looking tier detail including the $1/1,000 highest-commitment rate mentioned in the reference, giving the buyer more actionable specifics than B's vaguer range.

Codex alone: 2/3 · 17 s · 47,533 tokens

Tools: web_search ×1

Judge: Does not mention the $499/month plan structure (about 380,000 requests included, $1.30 per 1,000 beyond) Judge: Does not specify that the 5,000 free tier is shared across Bright Data's APIs

Bright Data’s SERP API costs $1.50 per 1,000 successful requests on pay-as-you-go. Volume plans reduce this to $1.00–$1.30 per 1,000. A free tier includes 5,000 monthly requests. Official pricing

Codex + Rose: 3/3 · 16 s · 88,900 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Bright Data"})

Bright Data’s SERP API costs $1.50 per 1,000 successful requests on pay-as-you-go.

Volume plans reduce the rate:

• $499/month: 380,000 included; $1.30/1,000 extra
• $999/month: 900,000 included; $1.10/1,000 extra
• $1,999/month: 2 million included; $1.00/1,000 extra
• Free tier: 5,000 shared monthly credits, no card required

Source: Rose’s vendor data as of October 1, 2026, based on Bright Data’s SERP pricing.

p10: What does the Linkup search API cost?

Gold facts. Standard search $0.005 per call ($5 per 1,000), sourced answer or structured output $0.006 ($6 per 1,000), deep search about $0.05 per call. Page fetch $1 per 1,000 without JS rendering, $5 with. Accounts with a professional email get $20 of free credit every month.

Judge. Winner: rose. Answer A matches the reference figures precisely (including deep search and page fetch context) and cites Rose's vendor data correctly, while Answer B gives vague, less accurate ranges and omits page fetch pricing and the professional-email credit detail.

Codex alone: 1/3 · 28 s · 105,218 tokens

Tools: web_search ×3

Judge: Deep search price stated as $0.005-$0.006 (same as standard/structured), but reference says deep search costs about $0.05 per call — an order of magnitude different. Judge: No mention of page fetch pricing ($1 per 1,000 without JS rendering, $5 with). Judge: Free credit claim differs: answer says '4,000 queries free' while reference states $20 of free credit per month for professional email accounts — not clearly the same offer. Judge: Does not clearly distinguish standard ($0.005) vs sourced/structured ($0.006) pricing.

Linkup’s current Search API pricing is $0.005–$0.006 per request, depending on search depth (Fast, Standard, or Deep).

• 4,000 queries free
• 10,000 paid searches: roughly $50–$60
• Enterprise plans use custom pricing

Source: Linkup pricing

Codex + Rose: 1/3 · 19 s · 88,086 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Linkup"})

Judge: Missing page fetch pricing: $1 per 1,000 without JS rendering, $5 per 1,000 with JS rendering - not mentioned in the answer at all.

Linkup’s search API is usage-based with no monthly subscription:

• Standard/fast/flash search with raw results: $0.005 per request ($5 per 1,000)
• Standard search with a sourced or structured answer: $0.006 per request ($6 per 1,000)
• Deep search: $0.050–$0.055 per request ($50–$55 per 1,000)
• Research requests: $0.25–$2.50 each
• Failed requests aren’t charged.

Eligible accounts created with a professional email receive $20 in monthly credit, enough for roughly 4,000 basic searches. Paid usage draws from a prepaid balance after that credit is exhausted. Rose’s sheet doesn’t state whether a card is required for the free credit.

Figures come from Rose’s normalized vendor data as of October 1, 2026, based on Linkup’s pricing documentation.

p11: How much is Parallel's Search API per 1,000 requests?

Gold facts. $1 per 1,000 requests on the Turbo and Fast processors, $5 per 1,000 on Basic and Advanced (10 results; extra results $1 per 1,000). Extract is $1 per 1,000 pages. $5 of free credits per month (about 5,000 requests) for organizations with a card on file.

Judge. Winner: rose. Answer A includes the correct extra-results pricing and free credit detail matching the reference, while Answer B incorrectly claims no additional fees for extra results.

Codex alone: 2/3 · 12 s · 47,655 tokens

Tools: web_search ×1

Judge: Missing detail that additional results beyond 10 cost $1 per 1,000 Judge: Missing mention of Extract pricing at $1 per 1,000 pages Judge: Missing mention of $5 free credits per month (~5,000 requests) for orgs with a card on file

Parallel’s Search API costs:

• Turbo / Fast: $1 per 1,000 requests
• Basic / Advanced: $5 per 1,000 requests

Each request includes up to 10 results, with no additional token fees. Parallel pricing

Codex + Rose: 3/3 · 13 s · 88,833 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Parallel"})

Parallel’s Search API costs:

• Turbo and Fast: $1 per 1,000 requests
• Basic and Advanced: $5 per 1,000 requests

Each request includes 10 results; extra results cost $1 per 1,000. A $5 monthly credit can cover up to 5,000 Turbo/Fast requests, but requires a card on file.

Source: Parallel pricing, via Rose vendor data as of October 1, 2026.

p12: How are Jina's Reader and Search APIs priced?

Gold facts. Prepaid tokens: $50 for 1 billion tokens or $500 for 11 billion. Reader bills the output tokens of each page; a search costs a fixed number of tokens starting from 10,000 (so from about $0.45-$0.50 per 1,000 searches). A new API key gets 10 million free tokens. Reader also works without a key at 20 requests per minute.

Judge. Winner: rose. Answer A correctly attributes figures to Rose's vendor data (as instructed) and keeps rate-limit/package details consistent, while Answer B presents the same data as Jina's own advertised rates without citing Rose, which is less aligned with the sourcing context, and both are similar in completeness but A's attribution is more appropriate per the scenario.

Codex alone: 3/3 · 38 s · 90,812 tokens

Tools: web_search ×2

Jina prices both APIs through shared token credits:

Tier	Included tokens	Price	Effective rate
Free	10 million	$0	Non-commercial use
Standard	1 billion	$50	$0.050 / million tokens
Premium	11 billion	$500	About $0.045 / million tokens

• Reader API (r.jina.ai): billed on the number of tokens in the extracted output. At the standard rate, a 5,000-token page costs about $0.00025.
• Search API (s.jina.ai): each search consumes a fixed allocation starting at 10,000 tokens. That makes a minimum-size request roughly $0.00050 standard or $0.00045 premium—about 2,000–2,222 searches per dollar.
• Basic Reader access works without a key at 20 requests/minute. Search requires an API key.
• A free key allows 500 Reader requests/minute and 100 Search requests/minute; premium limits are 5,000 and 1,000 respectively.
• Reader’s unauthenticated basic usage is free. With a key, usage draws from the shared token balance. Specialized processing such as OCR can consume tokens at a higher multiplier.

These are Jina’s currently advertised rates; taxes and billing currency can vary by location. Official Jina pricing · Reader and Search details

Codex + Rose: 3/3 · 22 s · 89,839 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Jina AI"})

Jina uses prepaid token-based pricing for both APIs:

• Reader API
• Free without an API key at 20 requests/minute per IP; no card required.
• With an API key, usage is charged according to the number of output tokens. There is no fixed per-page price because page lengths vary.

• OCR with jina-ocr-v1 consumes 40× the normal token amount.

• Search API

• Requires an API key.
• Each search starts at 10,000 tokens.
• At the $50 rate, that is at least $0.0005/search, or $0.50 per 1,000 searches.
• At the $500 rate, it is at least $0.00045/search, or $0.45 per 1,000 searches.

Token packages:

Package	Effective rate	Approx. searches at minimum charge
New API key	10M tokens free, one time	Up to 1,000
$50	1B tokens	Up to 100,000
$500 premium	11B tokens	Up to 1.1 million

The $50 package permits 500 Reader RPM and 100 Search RPM. The $500 premium package raises those limits to 5,000 and 1,000 RPM respectively. Failed requests do not consume tokens, and optional automatic top-ups are available.

These figures come from Rose’s unofficial normalized vendor data as of October 1, 2026, sourced from Jina’s Reader and Search pricing page.

f1: Do Firecrawl, Exa, Tavily and Brave have official MCP servers? Are they hosted or local?

Gold facts. All four have one. Firecrawl: hosted (https://mcp.firecrawl.dev/v2/mcp) and local (npx -y firecrawl-mcp). Exa: hosted (https://mcp.exa.ai/mcp, works keyless with rate limits) and local (npm package exa-mcp-server). Tavily: hosted (https://mcp.tavily.com/mcp/) and local (npx -y tavily-mcp). Brave: local only (npx -y @brave/brave-search-mcp-server with BRAVE_API_KEY).

Judge. Winner: base. Answer B provides the exact hosted URLs matching the reference (including Exa's correct https://mcp.exa.ai/mcp) while Answer A omits the actual hosted endpoint links, making B more precise and actionable.

Codex alone: 2/3 · 56 s · 257,944 tokens

Tools: web_search ×5

Judge: Answer does not mention that Exa's hosted MCP works keyless with rate limits. Judge: Answer does not give the exact local npm commands (e.g., npx -y firecrawl-mcp, npx -y tavily-mcp, npx -y @brave/brave-search-mcp-server) as in the reference. Judge: Tavily hosted URL given as https://mcp.tavily.com/mcp without trailing slash, a minor deviation from reference's https://mcp.tavily.com/mcp/.

Yes—all four have vendor-maintained official MCP servers.

Provider	Official MCP	Vendor-hosted remote	Local/self-hosted
Firecrawl	Yes	https://mcp.firecrawl.dev/v2/mcp	Yes, via firecrawl-mcp; supports Firecrawl Cloud or a self-hosted Firecrawl backend. Official repository
Exa	Yes	https://mcp.exa.ai/mcp	Open-source/npm implementation is available, although Exa now recommends its hosted endpoint. Official Exa MCP page
Tavily	Yes	https://mcp.tavily.com/mcp	Yes, via npx tavily-mcp@latest. Official repository
Brave Search	Yes	No documented public Brave-hosted MCP endpoint	Yes, via npm or Docker; supports stdio or locally hosted HTTP. Official repository

Important distinction: running the MCP server locally does not necessarily make the search itself local. The local Exa, Tavily, and Brave servers are adapters that still call their hosted APIs. Firecrawl can also use its cloud API, but uniquely supports pointing the MCP server at a self-hosted Firecrawl deployment.

Codex + Rose: 3/3 · 22 s · 128,936 tokens

Tools: rose ×5 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 1000, "vendors": ["Firecrawl", "Exa", "Tavily", "Brave"]}), get_access({"brand": "Firecrawl"}), get_access({"brand": "Exa"}), get_access({"brand": "Tavily"}), get_access({"brand": "Brave"})

Done — all four have official MCP servers.

Vendor	Hosted MCP	Local MCP
Firecrawl	Yes	Yes, via npx firecrawl-mcp
Exa	Yes	Yes, via npx exa-mcp-server
Tavily	Yes	Yes, via npx tavily-mcp@latest
Brave	No official hosted endpoint listed	Yes, via npx @brave/brave-search-mcp-server

The local servers still call each vendor’s hosted API; they are local MCP adapters, not self-hosted search infrastructure.

Figures come from Rose’s normalized vendor data as of October 1, 2026.

f2: Does serper.dev have an official MCP server?

Gold facts. No official MCP server from Serper; the Serper MCP servers that exist are community projects.

Judge. Winner: base. Both correctly state there's no official MCP server, but B provides more actionable detail by naming specific community MCP projects with links.

Codex alone: 3/3 · 21 s · 79,617 tokens

Tools: web_search ×2

No. As of October 2026, Serper.dev does not publish an official MCP server—only its REST API. Several community-built MCP wrappers exist, such as postfix/serper-mcp and garylab/serper-mcp-server, but they aren’t maintained or endorsed by Serper.

Serper’s official site currently makes no mention of MCP.

Codex + Rose: 3/3 · 19 s · 158,699 tokens

Tools: rose ×3 Rose calls: get_offers({"brand": "Serper"}), get_access({"brand": "Serper"}), compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 1000, "require": ["official_mcp_server"], "vendors": ["Serper"]})

No. Rose’s vendor data, current as of October 1, 2026, shows that Serper does not offer an official MCP server. Third-party/community MCP wrappers may exist, but they aren’t maintained by Serper.

f3: Is Exa SOC 2 compliant, and can it keep data in the EU?

Gold facts. Exa has SOC 2. Its data is processed and hosted in the United States; processing outside the US is only a custom enterprise arrangement, so no standard EU hosting.

Judge. Winner: base. Answer B more closely matches the reference by acknowledging configurable enterprise regions (aligning with the 'custom enterprise arrangement' nuance) while still correctly noting no public EU-only guarantee, and it adds actionable contract guidance, whereas Answer A omits the enterprise-arrangement nuance entirely.

Codex alone: 2/3 · 48 s · 174,682 tokens

Tools: web_search ×4

Judge: Does not state that Exa's data is hosted/processed in the US by default, and that EU hosting is only available via custom enterprise arrangement.

Yes on SOC 2; EU-only data residency is not publicly confirmed.

• SOC 2: Exa states it is SOC 2 Type II audited. The current report is available through its Trust Center.
• EU residency: Exa supports GDPR, DPAs, configurable enterprise regions, and optional Zero Data Retention, but its public documentation does not explicitly guarantee that all customer data and processing remain exclusively in the EU. Exa’s enterprise documentation

For a strict EU-residency requirement, obtain written confirmation covering API queries, logs, results, backups, support access, and subprocessors—and put the regional restriction in the contract. Zero Data Retention is helpful but is not itself an EU-residency guarantee.

Codex + Rose: 2/3 · 17 s · 121,759 tokens

Tools: rose ×2 Rose calls: get_offers({"brand": "Exa"}), compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 1000, "vendors": ["Exa"], "require": ["soc2", "eu_hosting"]})

Judge: Does not mention that Exa's data is hosted/processed in the US by default, and that EU hosting is only available via a custom enterprise arrangement, not a standard offering.

Yes—Exa reports SOC 2 compliance. However, EU data residency/hosting is not documented, so you shouldn’t assume data can remain exclusively in the EU.

Source: Exa security overview. Figures come from Rose’s vendor data as of October 1, 2026.

f4: Which of these scraping APIs offer a GDPR DPA and EU data processing: Firecrawl, ScrapingBee, Linkup, Jina, Exa, Parallel?

Gold facts. ScrapingBee (DPA, processing primarily in the EU), Linkup (GDPR DPA, EU processing available), Jina (GDPR DPA via Elastic, EU option), Parallel (DPA; EU data residency option for business customers, Search API requests to the EU endpoint processed in the EU). Firecrawl offers a DPA on paid plans but states no EU hosting. Exa hosts data in the US.

Judge. Winner: rose. Answer A's shortlist (Linkup, ScrapingBee, Jina, Parallel) exactly matches the reference facts on DPA+EU processing, whereas Answer B incorrectly downgrades Jina's EU option status and omits it from its final recommended list, contradicting the reference.

Codex alone: 2/3 · 78 s · 347,027 tokens

Tools: web_search ×7

Judge: Answer states Jina's EU processing is 'Not confirmed for hosted Jina API,' contradicting reference, which states Jina offers a GDPR DPA via Elastic with an EU option. Judge: Exa's lack of EU processing is described as 'Not publicly confirmed' rather than the reference's more definitive statement that Exa hosts data in the US (not EU).

As of October 2026, the procurement-grade answer is:

Provider	GDPR DPA	EU processing for API payloads	Assessment
Linkup	Available, but obtain it from sales	Yes	Best EU-native candidate. Linkup is French and states that data is processed and stored in the EU. Its public DPA documentation is less accessible, so request the executed Art. 28 DPA and subprocessor annex before approval. Privacy policy
ScrapingBee	Yes, incorporated automatically	Primarily EU	Strong yes, but not completely EU-only: core hosting is in Paris/EU, while monitoring, billing, and email subprocessors include US companies under SCCs. DPA and processing locations
Parallel	Yes, separately executed	Yes, optional EU endpoint	Strong yes if the EU-residency option is contractually enabled. EU Search API requests are processed and served in the EU with no request/response retention. Privacy policy, customer terms
Firecrawl	Yes, enterprise	No	DPA and zero-data-retention are available, but Firecrawl’s privacy policy says its servers and storage are in the US. Enterprise controls, processing location
Exa	Yes, enterprise/custom	Not publicly confirmed	Exa offers custom DPA documentation and zero retention, but I found no first-party commitment to an EU processing region or EU endpoint. Enterprise offering, pricing
Jina	Yes, through Elastic	Not confirmed for hosted Jina API	The Elastic Customer DPA now governs Jina, but hosted API data residency is not clearly committed to the EU. Self-hosting/on-prem is available where residency is mandatory. Jina legal terms, deployment options

Bottom line:

• Clearest “DPA + EU processing” choices: Linkup, ScrapingBee, and Parallel.
• Strict EU-only request processing: Linkup, or Parallel with its EU option.
• Not suitable where API payloads must remain in the EU: Firecrawl.
• Require written clarification before approval: Exa and hosted Jina.

“EU processing” does not necessarily mean every billing, support, telemetry, or proxy subprocessor remains in the EU. For strict residency, require the DPA/order form to specify the API endpoints, logs, backups, support access, subprocessors, and failover regions.

Codex + Rose: 3/3 · 26 s · 125,692 tokens

Tools: rose ×2 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "page", "monthly_volume": 1000, "vendors": ["Firecrawl", "ScrapingBee", "Linkup", "Jina AI", "Exa", "Parallel"]}), get_offers({"brand": "Jina AI"})

All six offer a GDPR DPA, but only four document EU-region data processing/hosting:

Vendor	GDPR DPA	EU processing/hosting
Linkup	Yes	Yes
ScrapingBee	Yes	Yes
Jina AI	Yes	Yes
Parallel	Yes	Yes
Firecrawl	Yes	No
Exa	Yes	Not documented

So if both are mandatory, shortlist Linkup, ScrapingBee, Jina AI, and Parallel. “EU processing” here means an advertised EU hosting/processing option, not merely GDPR compliance or use of EU subprocessors.

Figures come from Rose’s vendor data as of October 1, 2026.

f5: What output formats can Firecrawl's scrape endpoint return?

Gold facts. Markdown, HTML, raw HTML, screenshot, links, JSON (structured extraction), summary, images and more (branding, product, audio/video formats in newer versions).

Judge. Winner: base. Answer B provides a more complete and accurate list of formats (including menu, changeTracking, question, highlights) with useful operational details like rawBase64 exclusivity, while Answer A includes a dubious 'query' format not corroborated by the reference.

Codex alone: 3/3 · 56 s · 294,574 tokens

Tools: web_search ×6

Firecrawl’s /v2/scrape endpoint supports these output formats:

• markdown
• summary
• html — cleaned HTML
• rawHtml — original HTML
• rawBase64 — Base64-encoded response body
• links
• images
• screenshot
• json — structured extraction
• changeTracking
• branding
• product
• menu
• audio
• video
• question
• highlights

You can request multiple formats together, except rawBase64, which must be requested alone. Some formats accept or require configuration objects, such as json, screenshot, question, and highlights. Firecrawl Scrape API reference

Codex + Rose: 3/3 · 13 s · 90,430 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Firecrawl"})

Firecrawl’s /scrape endpoint supports these output formats:

• markdown
• summary
• html
• rawHtml
• rawBase64
• screenshot
• links
• json
• images
• branding
• product
• audio
• video
• query

Source: Firecrawl scrape documentation. Figures come from Rose’s vendor data as of October 1, 2026.

f6: For the free tiers of Brave Search API, Tavily, Exa and Firecrawl, which need a credit card?

Gold facts. Brave requires a card (for identity, not charged). Tavily, Exa and Firecrawl free tiers need no card.

Judge. Winner: tie. Both answers correctly identify that only Brave requires a credit card, matching the reference, with comparable clarity and sourcing.

Codex alone: 3/3 · 30 s · 135,956 tokens

Tools: web_search ×3

Only Brave Search API requires a credit card for its free monthly credits, as an anti-fraud check.

Provider	Card required for free tier?
Brave Search API	Yes
Tavily	No
Exa	No
Firecrawl	No

Sources: Brave, Tavily, Exa, Firecrawl. Verified October 1, 2026.

Codex + Rose: 3/3 · 13 s · 89,805 tokens

Tools: rose ×1 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 1, "vendors": ["Brave Search API", "Tavily", "Exa", "Firecrawl"]})

Only Brave Search API requires a credit card for its free monthly credits, used for identity verification.

No credit card is required for:

• Tavily
• Exa
• Firecrawl

Source: Rose’s normalized vendor data, current as of October 1, 2026.

f7: Can I use Jina Reader without an API key?

Gold facts. Yes: Reader (r.jina.ai) works without a key, rate-limited to 20 requests per minute per IP. The Search API (s.jina.ai) needs a key.

Judge. Winner: tie. Both are correct and complete, but B provides an actionable curl command plus an extra pricing/rate-limit link, giving slightly more clarity and actionability without unnecessary length.

Codex alone: 2/3 · 13 s · 47,492 tokens

Tools: web_search ×1

Judge: The 500 RPM figure for free API key usage is not confirmed in the reference and may be inaccurate/unverified

Yes. Jina Reader supports keyless basic usage:

curl https://r.jina.ai/https://example.com

No Authorization header is needed. The current limit is 20 requests per minute per IP without a key; a free API key raises that to 500 RPM. Note that Jina Search (s.jina.ai) does require a key. Jina Reader documentation · Pricing and rate limits

Codex + Rose: 3/3 · 17 s · 87,992 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Jina AI"})

Yes. Jina Reader works without an API key:

https://r.jina.ai/https://example.com

Keyless use is free and limited to 20 requests/minute per IP. An API key raises the Reader limit to 500 RPM; Jina Search (s.jina.ai) does require a key.

Source: Jina Reader via Rose vendor data, current as of October 1, 2026.

f8: Does Zyte have a SOC 2 report?

Gold facts. No SOC 2 is listed on Zyte's trust center; it lists ISO 27001.

Judge. Winner: tie. Both answers correctly state no SOC 2 is listed, cite the same trust center source, and give comparable actionable next steps, with no meaningful difference in correctness, completeness, or clarity.

Codex alone: 3/3 · 18 s · 80,469 tokens

Tools: web_search ×2

No—at least not publicly. Zyte’s current Security & Trust Center lists ISO 27001, GDPR, and CCPA, but not SOC 2. It offers penetration-test reports under NDA.

For vendor due diligence, confirm directly with sales@zyte.com, since an unpublished or in-progress SOC 2 report is still possible.

Codex + Rose: 3/3 · 18 s · 133,244 tokens

Tools: rose ×1, web_search ×1 Rose calls: get_offers({"brand": "Zyte"})

No public SOC 2 report is listed. Zyte’s Trust Center identifies ISO 27001 certification, GDPR, and CCPA compliance, but not SOC 2.

Based on Rose’s vendor data as of October 1, 2026, treat Zyte as not publicly SOC 2 certified unless their sales/security team provides a private report.

f9: Does Bright Data offer an MCP server, and how does it authenticate?

Gold facts. Yes, an official hosted MCP server at https://mcp.brightdata.com/mcp, authenticated with an API token (passed as ?token=), and a local npm package (@brightdata/mcp) using API_TOKEN. Its free tier is 5,000 requests per month.

Judge. Winner: tie. Answer B includes the same core facts plus practical security guidance and more specific, plausible source citations, without sacrificing accuracy or clarity (though both omit the free tier detail).

Codex alone: 2/3 · 21 s · 90,189 tokens

Tools: web_search ×2

Judge: Missing free tier detail: 5,000 requests per month

Yes—Bright Data offers an official MCP server in two forms:

• Hosted: Connect to https://mcp.brightdata.com/mcp?token=YOUR_API_TOKEN
• Self-hosted/local: Run npx @brightdata/mcp and provide the token as the API_TOKEN environment variable.

Authentication uses a Bright Data API token, obtained from account settings—not OAuth. The hosted configuration places it in the connection URL, while the local server keeps it in an environment variable. Treat the full hosted URL as a secret because it contains the credential. Bright Data MCP documentation, official server manifest.

Codex + Rose: 2/3 · 14 s · 88,928 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Bright Data"})

Judge: Missing free tier detail (5,000 requests/month) mentioned in reference

Yes. Bright Data offers an official MCP server with two connection options:

• Hosted: https://mcp.brightdata.com/mcp?token=YOUR_API_TOKEN
• Local: npx @brightdata/mcp, with API_TOKEN set as an environment variable.

Authentication uses a Bright Data API token. Source: Bright Data MCP repository.

These details come from Rose’s vendor data as of October 1, 2026.

g1: How do I add Tavily's MCP server to Claude Code?

Gold facts. claude mcp add --transport http tavily https://mcp.tavily.com/mcp/ (authenticate with OAuth or a Tavily API key as tavilyApiKey or a Bearer header), or run the local server: claude mcp add tavily -e TAVILY_API_KEY= -- npx -y tavily-mcp. Get the API key at app.tavily.com.

Judge. Winner: rose. Answer A fully matches the reference by providing both the hosted OAuth setup and the local npx server with the API key, whereas Answer B omits the local server option entirely.

Codex alone: 2/3 · 28 s · 86,520 tokens

Tools: web_search ×2

Judge: Does not mention the local server alternative (claude mcp add tavily -e TAVILY_API_KEY= -- npx -y tavily-mcp) Judge: Does not mention that the Tavily API key can be obtained at app.tavily.com Judge: URL used lacks trailing slash (https://mcp.tavily.com/mcp vs reference's https://mcp.tavily.com/mcp/), a minor discrepancy

The simplest and safest approach is Tavily’s hosted MCP server with OAuth:

claude mcp add --transport http --scope user tavily https://mcp.tavily.com/mcp

Then:

1. Start Claude Code with claude.
2. Enter /mcp.
3. Select tavily and complete authentication in your browser.
4. Verify with:

claude mcp list
claude mcp get tavily

--scope user makes Tavily available in every project. Omit it to configure Tavily only for the current project.

Alternatively, authenticate with an API key:

claude mcp add --transport http --scope user tavily \
  'https://mcp.tavily.com/mcp/?tavilyApiKey=YOUR_API_KEY'

OAuth is preferable because putting a key in the URL can expose it through shell history or configuration. Once connected, Claude Code gains Tavily’s search, extraction, site-map, and crawling tools.

Sources: Tavily MCP server documentation, Claude Code MCP documentation.

Codex + Rose: 2/3 · 22 s · 87,179 tokens

Tools: rose ×1 Rose calls: get_access({"brand": "Tavily"})

Judge: Does not mention the alternative authentication option for the hosted server via a Tavily API key passed as tavilyApiKey or a Bearer header; only OAuth is mentioned.

Add Tavily’s hosted MCP server to Claude Code:

claude mcp add --transport http tavily https://mcp.tavily.com/mcp/

Then run Claude Code, enter /mcp, select tavily, and authenticate with OAuth.

Alternatively, use the local server with an API key:

claude mcp add tavily \
  --env TAVILY_API_KEY=tvly-YOUR_KEY \
  -- npx -y tavily-mcp@latest

Get an API key from the Tavily dashboard. Confirm installation with:

claude mcp list

These setup details come from Rose’s Tavily vendor data as of October 1, 2026, sourced from Tavily’s MCP documentation.

g2: Give me the mcpServers JSON config for the Brave Search MCP server for Cursor.

Gold facts. {"mcpServers": {"brave-search": {"command": "npx", "args": ["-y", "@brave/brave-search-mcp-server"], "env": {"BRAVE_API_KEY": ""}}}}. The key comes from the Brave Search API dashboard (api-dashboard.search.brave.com). The server key name ("brave-search", "brave") is the user's choice.

Judge. Winner: rose. Answer A matches the reference config exactly (command, args, env) and cites Rose's vendor data for verification, whereas Answer B adds an unverified extra "--transport stdio" argument not present in the reference.

Codex alone: 3/3 · 20 s · 48,373 tokens

Tools: web_search ×1

Done — add this to .cursor/mcp.json in your project or ~/.cursor/mcp.json globally:

{
  "mcpServers": {
    "brave-search": {
      "command": "npx",
      "args": [
        "-y",
        "@brave/brave-search-mcp-server",
        "--transport",
        "stdio"
      ],
      "env": {
        "BRAVE_API_KEY": "YOUR_BRAVE_API_KEY"
      }
    }
  }
}

Get an API key from Brave Search API. This uses Brave’s current official package; the older @modelcontextprotocol/server-brave-search package is deprecated. Official server documentation

Codex + Rose: 3/3 · 12 s · 86,681 tokens

Tools: rose ×1 Rose calls: get_access({"brand": "Brave Search API"})

Done — add this to Cursor’s MCP configuration:

{
  "mcpServers": {
    "brave": {
      "command": "npx",
      "args": ["-y", "@brave/brave-search-mcp-server"],
      "env": {
        "BRAVE_API_KEY": "<your_BRAVE_API_KEY>"
      }
    }
  }
}

Get your key from the Brave Search API dashboard. Source: official Brave MCP server, via Rose vendor data as of October 1, 2026.

g3: What are the Python SDK install commands for Exa, Firecrawl and Tavily?

Gold facts. pip install exa-py; pip install firecrawl-py; pip install tavily-python.

Judge. Winner: tie. Both give correct install commands, but B adds a useful combined install command and more specific SDK source links, offering slightly more actionable completeness without sacrificing correctness.

Codex alone: 3/3 · 19 s · 82,082 tokens

Tools: web_search ×2

pip install exa-py
pip install firecrawl-py
pip install tavily-python

Or install all three together:

python -m pip install exa-py firecrawl-py tavily-python

Sources: Exa, Firecrawl, Tavily.

Codex + Rose: 3/3 · 16 s · 89,657 tokens

Tools: rose ×3 Rose calls: get_access({"brand": "Exa"}), get_access({"brand": "Firecrawl"}), get_access({"brand": "Tavily"})

From Rose’s vendor data as of October 1, 2026:

pip install exa-py
pip install firecrawl-py
pip install tavily-python

Docs: Exa · Firecrawl · Tavily

g4: How do I connect Apify to my AI assistant?

Gold facts. Through Apify's official hosted MCP server at https://mcp.apify.com (OAuth or an Apify API token, APIFY_TOKEN). Sign up at apify.com; the free plan includes $5 of monthly platform credit.

Judge. Winner: base. Answer B covers the same core facts (hosted MCP server, OAuth/API token) but adds concrete, actionable steps for multiple clients (ChatGPT, Codex, Cursor, custom agents, REST API fallback) and a test command, making it more complete and useful without sacrificing correctness.

Codex alone: 2/3 · 42 s · 159,750 tokens

Tools: command_execution ×1, web_search ×3

Judge: Does not mention the free plan includes $5 of monthly platform credit

The easiest connection is through Apify’s hosted MCP server.

1. Create an Apify account and get your API token from Apify Console → Settings → API & Integrations.
2. Add this MCP server to your assistant:

{
  "mcpServers": {
    "apify": {
      "url": "https://mcp.apify.com"
    }
  }
}

1. Authenticate with Apify when prompted. OAuth is preferable because it avoids storing the token in configuration.
2. Test it with: “Use Apify to scrape https://example.com and summarize the page.”

Apify MCP lets the assistant discover and run Actors, then access their results. You can restrict access to selected tools or Actors:

{
  "mcpServers": {
    "apify": {
      "url": "https://mcp.apify.com?tools=apify/rag-web-browser,apify/web-fetch"
    }
  }
}

The exact setup depends on your assistant:

• ChatGPT: Enable Developer mode, open Settings → Apps & Connectors → Create, enter https://mcp.apify.com, choose OAuth, and authorize Apify. ChatGPT instructions
• Codex: After installing and signing into the Apify CLI, run apify mcp install codex.
• Cursor/VS Code/Claude: Add the JSON above to the client’s MCP configuration.
• Custom OpenAI agent: Connect https://mcp.apify.com as a streamable HTTP MCP server and pass Authorization: Bearer $APIFY_TOKEN. OpenAI Agents SDK example
• No MCP support: Call the Apify REST API as an ordinary function tool: start an Actor run, wait for completion, then retrieve its dataset.

Keep the API token server-side, restrict the available Actors, and remember that Actor executions may incur Apify usage charges. The complete configuration reference is in the Apify MCP documentation.

Codex + Rose: 2/3 · 15 s · 86,751 tokens

Tools: rose ×1 Rose calls: get_access({"brand": "Apify"})

Judge: Does not mention the free plan includes $5 of monthly platform credit.

You can connect Apify through its official MCP server.

The simplest configuration is:

{
  "mcpServers": {
    "apify": {
      "url": "https://mcp.apify.com"
    }
  }
}

Add this to your assistant’s MCP configuration, restart it, then complete the Apify OAuth browser sign-in when prompted.

Alternatively, run the MCP server locally:

npx -y @apify/actors-mcp-server

Set APIFY_TOKEN in the server environment using a token from your Apify Console. Don’t place the token directly in source code or commit it.

Once connected, your assistant can discover and run Apify Actors for scraping and browser automation. See the Apify MCP documentation.

These connection details come from Rose’s Apify vendor sheet, current as of October 1, 2026.

c1: Exa or Tavily for about 30,000 searches a month: which is cheaper and by how much?

Gold facts. Exa is slightly cheaper: about $200/month pay as you go (30,000 auto searches at $7 per 1,000 minus the $10 monthly free credit), or about $110 using 'instant' search at $4 per 1,000. Tavily costs about $220/month (Startup plan, 38,000 credits, or Bootstrap $100 plus overage).

Judge. Winner: tie. Answer B more completely mirrors the reference's two Exa pricing tiers (instant ~$110-120 and auto ~$200-210) alongside Tavily's ~$220, giving a more accurate and actionable breakdown than Answer A's single auto-only comparison.

Codex alone: 2/3 · 33 s · 91,843 tokens

Tools: web_search ×2

Judge: Exa Instant listed as $120 vs reference ~$110 Judge: Exa pay-as-you-go/Fast-Auto listed as $210 vs reference ~$200 Judge: Added an extra 'Tavily Basic — pay-as-you-go' $240 figure not in reference (unverified)

Exa is cheaper at 30,000 searches/month.

Configuration	Monthly cost
Exa Instant Search	$120
Exa Fast/Auto Search	$210
Tavily Basic — Startup plan	$220
Tavily Basic — pay-as-you-go	$240

So, compared with Tavily’s cheapest suitable plan:

• Exa Instant saves $100/month—about 45%.
• Exa Fast/Auto saves $10/month—about 5%.

This assumes up to 10 Exa results per request and Tavily Basic at one credit per search. Tavily Advanced consumes two credits, making 30,000 searches substantially more expensive. Optional extraction, summaries, or extra results can change the totals.

Sources: Exa pricing, Tavily pricing, Tavily credit usage.

Codex + Rose: 3/3 · 15 s · 88,390 tokens

Tools: rose ×1 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 30000, "vendors": ["Exa", "Tavily"]})

Exa is cheaper for 30,000 standard searches/month.

• Exa: about $200/month
• Tavily: about $220/month
• Savings: $20/month or $240/year—roughly 9% cheaper

This assumes Exa’s default “auto” search and Tavily’s basic search; advanced/deep modes use different rates. Figures come from Rose’s vendor data as of October 1, 2026: Exa pricing, Tavily API credits.

c2: SerpApi vs Serper for 50,000 Google searches a month: what would each cost?

Gold facts. Serper: the $50 Starter pack (50,000 queries, valid 6 months) covers it, so about $50 a month. SerpApi has no 50,000 plan: Big Data ($275, 30,000) renewed early, about $550, or Searcher ($725, 100,000). Serper is about 10 times cheaper; SerpApi covers many more engines.

Judge. Winner: rose. Answer A correctly reflects the reference's key $550 (Big Data renewed early) SerpApi figure with accurate per-1000 math, while Answer B leads with the less relevant $725 plan and bungles its attempt to compute the renewed-Big-Data cost ($458.33 instead of the correct ~$550).

Codex alone: 2/3 · 39 s · 136,572 tokens

Tools: web_search ×3

Judge: Answer states SerpApi's early-renewal 30,000 plan averages to about $458.33/month for 50,000 searches, whereas the reference indicates the early-renewal cost is about $550 (not $458.33). Judge: Answer omits the reference's point that SerpApi covers many more search engines, a nuance that should be weighed against the price difference.

At current public pricing:

Provider	Plan needed for 50,000 searches	Cost	Cost per 1,000
Serper	50,000-credit top-up	$50/month	$1.00
SerpApi	100,000-search “Searcher” plan	$725/month	$14.50 per 1,000 actually used

Serper is prepaid rather than a subscription; credits last six months. Serper pricing

SerpApi has no standard 50,000-search tier—the plans jump from 30,000 at $275 to 100,000 at $725. SerpApi pricing

So the straightforward comparison is $50 vs. $725 per month: Serper saves $675/month, or $8,100/year, before taxes.

One nuance: repeatedly early-renewing SerpApi’s 30,000-search plan would average about $458.33/month at exactly 50,000 searches, but billing would be irregular and the throughput limit lower.

Codex + Rose: 2/3 · 19 s · 87,445 tokens

Tools: rose ×1 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 50000, "vendors": ["SerpApi", "Serper"]})

Judge: Answer omits that SerpApi supports many more search engines than Serper, a point included in the reference.

At 50,000 Google searches/month:

Vendor	Cost	Effective rate
Serper	$50 prepaid pack	$1 / 1,000 searches
SerpApi	$550/month Big Data plan	$11 / 1,000 searches

Serper is $500/month cheaper—about 91% less. Its credits are prepaid, valid for six months, and taxes may be added. SerpApi is a monthly subscription.

Figures come from Rose’s vendor data as of October 1, 2026: Serper and SerpApi.

c3: Brave, Exa or Tavily for 5,000 searches a month: rank them by monthly cost.

Gold facts. Brave about $20 (5,000 at $5 per 1,000 minus the $5 monthly credit); Exa about $25 (5,000 auto searches at $7 per 1,000 minus $10 free credit; about $10 with instant search); Tavily about $38 (Project $30 for 4,000 credits plus 1,000 at $0.008) or pay as you go about $32 after the 1,000 free credits.

Judge. Winner: rose. Answer A's cost figures ($20 Brave, $25 Exa, $38 Tavily) align with the reference values, while Answer B's Tavily figure ($40) ignores the free credit allowance cited in the reference, making its ranking numerically less accurate despite correct ordering.

Codex alone: 2/3 · 31 s · 90,743 tokens

Tools: web_search ×2

Judge: Tavily cost stated as $40 (5,000 × $0.008) ignoring the 1,000 free credits, which would make pay-as-you-go about $32 per the reference, not $40 Judge: Reference also cites Tavily Project plan at about $38 ($30 for 4,000 credits plus 1,000 at $0.008), which the answer dismisses without correctly recalculating the PAYG alternative

Done — for 5,000 standard/basic API searches per month, cheapest to most expensive:

✅ Verified

Rank	Provider	Monthly cost	Calculation
1	Brave	$20	$25 usage at $5/1,000, minus $5 monthly credit
2	Exa	$25	$35 usage at $7/1,000, minus $10 monthly credit
3	Tavily	$40	5,000 basic searches × $0.008/credit

Sources: Brave pricing, Exa pricing, Tavily pricing.

Tavily’s advanced search consumes two credits, making 5,000 searches $80. Its $30 Project plan only includes 4,000 credits, so pay-as-you-go is the cheapest published option that covers 5,000. Exa’s one-time $10 onboarding bonus may reduce its first month to $15.

Codex + Rose: 2/3 · 20 s · 89,296 tokens

Tools: rose ×1 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 5000, "vendors": ["Brave", "Exa", "Tavily"]})

Judge: Brave pricing explanation says '1,000 free searches, then 4,000 × $5/1,000' which mischaracterizes the actual billing mechanism (a $5 monthly credit applied to $5/1,000 pricing) even though the final $20 figure matches.

At 5,000 standard searches/month, cheapest to most expensive:

1. Brave Search API — $20/month

2. 1,000 free searches, then 4,000 × $5/1,000.

3. Exa — about $25/month

4. Based on default auto search pricing and monthly free credits.

5. Tavily — $38/month

6. Project plan: $30 for 4,000 basic searches, plus $8 overage for 1,000.

Verdict: Brave is cheapest, saving roughly $5/month versus Exa and $18/month versus Tavily.

Figures come from Rose’s normalized vendor data as of October 1, 2026. Exa search modes and Tavily advanced searches cost more than the standard modes compared here.

c4: Zyte or Bright Data for scraping 500,000 simple pages a month: which is cheaper?

Gold facts. Zyte: from about $65/month pay as you go at the cheapest tier ($0.13 per 1,000), more for harder sites or browser rendering. Bright Data Web Unlocker: $1.50 per 1,000 pay as you go, about $750, or the $499 plan (about 383,000 included, then $1.30 per 1,000), about $650. Zyte is much cheaper for simple sites.

Judge. Winner: tie. Answer B includes both Bright Data pricing options (PAYG and Scale plan) matching the reference, while Answer A omits the $750 PAYG figure and adds an inaccurate claim about a '38,461 request trial' not supported by the reference.

Codex alone: 3/3 · 23 s · 52,404 tokens

Tools: web_search ×1

Zyte is substantially cheaper for 500,000 simple pages/month.

Assuming plain HTML requests without full browser rendering:

Provider	Estimated monthly cost
Zyte API	$65 pay-as-you-go
Bright Data Web Unlocker	about $651 on its $499 Scale plan
Bright Data Web Unlocker PAYG	$750

Zyte’s simple-site rate is $0.13 per 1,000 successful responses: 500 × $0.13 = $65. Zyte pricing

Bright Data charges $1.50 per 1,000 PAYG, or $499 for 383,000 requests plus $1.30 per 1,000 additional requests: $499 + (117 × $1.30) = $651.10. Bright Data Web Unlocker pricing

Verdict: Choose Zyte if these genuinely are simple, unrendered pages—roughly 90% cheaper on published pricing. Confirm several representative URLs in Zyte’s calculator because it assigns pricing per website; harder targets may enter higher tiers. If you need browser interaction rather than basic page retrieval, the comparison changes because Bright Data’s browser product is bandwidth-priced.

Codex + Rose: 3/3 · 20 s · 87,728 tokens

Tools: rose ×1 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "page", "monthly_volume": 500000, "vendors": ["Zyte", "Bright Data"]})

Zyte is cheaper—about 10× cheaper for 500,000 simple pages/month.

• Zyte: $65/month at $0.13 per 1,000 simple HTTP pages.
• Bright Data: $651.10/month via Web Unlocker—$499 for 383,000 requests, plus $152.10 for the remaining 117,000.
• Savings with Zyte: $586.10/month.

Zyte automatically assigns each website a difficulty tier, so harder sites, browser rendering, screenshots, or extraction increase the price. Bright Data includes 5,000 free requests monthly; Zyte instead offers roughly 38,461 requests as a one-time trial.

Figures come from Rose’s normalized vendor data as of October 1, 2026.

c5: Firecrawl, ScrapingBee or ScraperAPI for 100,000 basic (non-JS) pages a month: what does each cost?

Gold facts. ScraperAPI: Hobby $49/month (100,000 credits, 1 credit per basic page). ScrapingBee: Freelance $49/month (250,000 credits, 1 credit per page without JS). Firecrawl: Standard $99/month (100,000 credits). ScraperAPI and ScrapingBee tie at $49; Firecrawl is about double.

Judge. Winner: tie. Both answers correctly match the reference pricing, structure, and conclusions with similar clarity, links, and actionable detail, making them essentially equivalent in quality.

Codex alone: 3/3 · 41 s · 141,266 tokens

Tools: web_search ×3

For 100,000 basic, non-JavaScript pages per month, assuming ordinary domains and one credit per successful page:

Provider	Required plan	Monthly billing	Annual commitment	Included capacity
Firecrawl	Standard	$99/mo	$83/mo ($990/year)	100,000 pages
ScrapingBee	Freelance	$49/mo	Not advertised	250,000 pages
ScraperAPI	Hobby	$49/mo	$44.10/mo ($529.20/year)	100,000 pages

So:

• Cheapest month-to-month: ScrapingBee and ScraperAPI, tied at $49/month.
• Best capacity at that price: ScrapingBee—its $49 plan includes 250,000 basic requests.
• Cheapest with annual billing: ScraperAPI at $44.10/month equivalent.
• Firecrawl: roughly twice the price, but it returns processed content such as clean Markdown and includes crawling/content-extraction functionality rather than primarily acting as a proxy API.

The calculation uses each vendor’s basic rate:

• Firecrawl: one credit per basic page. Official pricing
• ScrapingBee: one credit with classic proxies and JavaScript disabled. Prices exclude VAT. Official pricing and credit rules
• ScraperAPI: one credit for a normal request; protected domains and special features consume more. Official pricing and credit rules

These numbers rise substantially for premium proxies, protected domains, SERPs, or anti-bot bypassing.

Codex + Rose: 3/3 · 29 s · 136,806 tokens

Tools: rose ×4 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "page", "monthly_volume": 100000, "vendors": ["Firecrawl", "ScrapingBee", "ScraperAPI"]}), get_offers({"brand": "Firecrawl"}), get_offers({"brand": "ScrapingBee"}), get_offers({"brand": "ScraperAPI"})

Done — for 100,000 basic, non-JavaScript pages per month:

Vendor	Cheapest suitable plan	Monthly billing	Annual commitment	Effective cost
ScraperAPI	Hobby, 100k credits	$49/month	$44.10/month	$0.49/1k monthly
ScrapingBee	Freelance, 250k credits	$49/month	No discount listed	$0.49/1k used
Firecrawl	Standard, 100k credits	$99/month	$83/month	$0.99/1k monthly

ScraperAPI and Firecrawl match the volume exactly. ScrapingBee supplies 250,000 non-JS pages for the same $49, making it the best value if usage may grow.

Important credit rules:

• ScrapingBee: disable JS rendering; otherwise pages cost 5 credits each. No overages—upgrade or renew early.
• ScraperAPI: standard pages cost 1 credit; rendering or premium proxies cost 10+. No overages on Hobby.
• Firecrawl: basic pages cost 1 credit, but JSON extraction and certain advanced formats add 4 credits per page. Paid overage is available.

Free access: Firecrawl offers 1,000 pages monthly without a card; ScraperAPI offers 1,000 monthly plus a one-time 5,000-credit trial; ScrapingBee offers a one-time 1,000-credit trial.

Figures come from Rose’s normalized vendor data dated October 1, 2026.

c6: For an AI agent that needs search plus page content, compare Exa, Tavily, Linkup and Parallel on price per 1,000 searches and per 1,000 pages.

Gold facts. Exa: $7 per 1,000 searches (auto; $4 instant), $1 per 1,000 pages. Tavily: $8 per 1,000 basic searches, $1.60 per 1,000 extracted pages (pay as you go). Linkup: $5 per 1,000 searches, $1 per 1,000 pages (no JS). Parallel: $1 per 1,000 searches (Turbo/Fast; $5 Basic/Advanced), $1 per 1,000 pages. Parallel is cheapest on search; pages are about $1 everywhere except Tavily.

Judge. Winner: rose. Answer A reproduces every reference figure exactly (including Exa's $4 instant‑search price and Tavily's $1.60/$3.20 page cost) and adds correctly‑sourced nuance (free‑credit tiers, Linkup JS surcharge) without distorting the baseline numbers, whereas Answer B blurs Linkup's page price into a vague $1–$6 range and omits the Exa instant‑search detail, making it slightly less faithful to the reference facts.

Codex alone: 2/3 · 59 s · 260,444 tokens

Tools: web_search ×5

Judge: Exa's $4 instant search price is omitted (only $7 auto price mentioned) Judge: Linkup's page price is given as a range $1-$6, while reference specifies a single $1 per 1,000 pages for no-JS mode, adding ambiguity not supported by reference

Bottom line: Parallel is cheapest for search, while Exa and Parallel tie for standard page extraction at $1 per 1,000 pages.

Current public pay-as-you-go pricing, normalized to USD:

Provider	1,000 searches	1,000 extracted pages	Notes
Parallel	$1–$5	$1	Turbo/Fast search is $1; Basic/Advanced is $5. Each search returns up to 10 URLs with compressed excerpts.
Linkup	$5–$6	$1–$6	Price varies by search output and Fetch mode; the upper Fetch tier is intended for harder pages.
Exa	$7	$1 per content type	Search includes up to 10 results. Text, highlights, and summaries can be separately metered content types.
Tavily	$8 basic / $16 advanced	$1.60 basic / $3.20 advanced	PAYG is $0.008/credit; search costs 1–2 credits and extraction costs 1–2 credits per five successful URLs.

For an agent doing 1,000 searches and fetching five pages after each search:

Provider	Estimated total
Parallel	$6–$10
Exa	$12
Tavily	$16 basic / $32 advanced
Linkup	$10–$36

The general formula is:

cost = search CPM + pages-per-search × page CPM

Important caveats:

• These tiers are not quality-equivalent. “Advanced,” JavaScript rendering, live crawling, and harder-site handling differ.
• Search responses already include snippets or compressed excerpts. Only count page extraction when the agent needs fuller page content.
• Tavily becomes cheaper on prepaid monthly plans: credits fall from $0.008 to $0.005–$0.0075, making basic search roughly $5–$7.50/1,000 and basic extraction $1–$1.50/1,000 pages.
• Free credits and enterprise discounts are excluded.

Sources: Exa pricing, Tavily credit pricing, Linkup pricing, Parallel pricing.

Codex + Rose: 3/3 · 40 s · 145,655 tokens

Tools: rose ×6 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 1000, "vendors": ["Exa", "Tavily", "Linkup", "Parallel"]}), compare_offers({"niche": "web-search-scraping", "unit": "page", "monthly_volume": 1000, "vendors": ["Exa", "Tavily", "Linkup", "Parallel"]}), get_offers({"brand": "Exa"}), get_offers({"brand": "Tavily"}), get_offers({"brand": "Linkup"}), get_offers({"brand": "Parallel"})

For a workload of 1,000 searches plus 1,000 fetched pages, Parallel is cheapest on published base rates.

Vendor	Per 1,000 searches	Per 1,000 pages	Combined nominal cost	Effective cost after recurring free credit*
Parallel	$1 Turbo/Fast; $5 Basic/Advanced	$1	$2	$0
Linkup	$5 standard results; $6 sourced answer	$1 standard	$6	$0
Exa	$7 Auto; $4 Instant	$1 per content type	$8 using Auto	$0
Tavily	$8 Basic; $16 Advanced	$1.60 Basic; $3.20 Advanced	$9.60 Basic	$1.60

*Free allowances use a shared balance across searches and pages:

• Parallel: $5/month, requiring a card on file.
• Linkup: $20/month for eligible accounts using a professional email.
• Exa: $10/month, no card required.
• Tavily: 1,000 credits/month, no card required. The example workload consumes 1,200 credits, hence $1.60 payable.

Important caveats:

• Exa’s $1 page rate is per content type; requesting text, highlights, and summary can bill each separately.
• Linkup’s page price rises to $5/1,000 with JavaScript rendering or Pro mode, and $10 with both.
• Parallel’s $1 search tier returns up to ten results; extra results cost more.
• These prices compare base search results, not deep research endpoints.

Verdict: choose Parallel for minimum unit cost. Linkup has the most generous recurring free credit, provided the account qualifies. Exa is competitive when its Instant search is sufficient; Tavily is the most expensive here but still inexpensive for page extraction.

Figures come from Rose’s normalized vendor data as of October 1, 2026: Parallel pricing, Linkup pricing, Exa pricing, and Tavily API credits.

l1: List web search APIs that have an official MCP server and at least 1,000 free searches every month.

Gold facts. Non-exhaustive; must include most of: Parallel (5,000/month), Bright Data (5,000/month shared across products), Linkup (about 4,000/month, professional email), Exa (about 1,428/month), Tavily (1,000/month), Brave (about 1,000/month, card required), Browserbase (1,000/month). Must not include Firecrawl (its 1,000 free credits are only about 500 searches) or SerpApi (250).

Judge. Winner: rose. Answer A closely matches the reference list (including Linkup and Browserbase, which B omits) and avoids introducing unverified or borderline providers, whereas B misses key reference entries and adds several speculative/non-standard services (freesea, You.com, SERPdive, Liner) that aren’t corroborated and complicate the buyer’s decision.

Codex alone: 2/3 · 130 s · 926,571 tokens

Tools: web_search ×12

Judge: Missing Linkup (about 4,000/month) which is one of the reference's required vendors Judge: Missing Browserbase (1,000/month) which is one of the reference's required vendors

Done — as of October 1, 2026, these web-search APIs meet the requirement through vendor-operated or vendor-maintained MCP servers:

Provider	Recurring free search allowance	Notes
freesea	No monthly cap	Free official API and MCP server; 100 requests/minute, but capacity is shared and not guaranteed.
Bright Data	5,000/month	Web search is included in its official hosted and self-hosted MCP server.
Parallel	1,000–5,000/month	$5 recurring monthly credit; search costs $0.001–$0.005 per request. Its official Search MCP also supports anonymous use.
You.com	100/day, roughly 2,800–3,100/month	Keyless official MCP profile: https://api.you.com/mcp?profile=free.
Exa	About 1,428/month	$10 recurring monthly credit ÷ $7 per 1,000 standard searches. Full-page content can incur additional charges. Official MCP server.
Brave Search	1,000/month	$5 recurring credit at $5 per 1,000 searches; requires a card for verification. Official MCP repository.
Tavily	1,000 basic searches/month	Advanced searches consume two credits, reducing the allowance to 500. Official MCP server.
SERPdive	1,000 Mako searches/month	No card required. Moby searches cost 1.5 credits, so only Mako reaches 1,000.

A conditional additional option is Liner: 200 free searches daily, but only for one year and only for the first 1,000 eligible users.

I excluded providers whose advertised “1,000 free credits” produce fewer than 1,000 actual searches—for example, Firecrawl at roughly 500 searches and HasData at roughly 100.

Codex + Rose: 3/3 · 21 s · 93,953 tokens

Tools: rose ×1 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 1000, "min_free_per_month": 1000, "require": ["official_mcp_server"]})

Done — eight APIs meet both criteria.

Rose vendor data as of October 1, 2026:

API	Free searches/month	Important condition
Parallel	5,000	Card required to receive recurring credits
Bright Data	5,000	Shared credit pool; no card required
Linkup	4,000	Requires an eligible professional-email account
Exa	1,428	Based on default auto search pricing; no card required
Apify	≈1,111	Estimate using the free credit with its official Google Search Results Scraper
Tavily	1,000 basic searches	Advanced searches consume two credits; no card required
Brave Search API	1,000	Card required for identity verification
Browserbase	1,000	Search API calls; no card required

All eight are recorded by Rose as having vendor-official MCP servers. Free allowances may be shared with other API products rather than reserved exclusively for search.

l2: Which web scraping APIs have a recurring monthly free tier with no credit card?

Gold facts. Non-exhaustive; must include most of: Firecrawl (1,000 pages/month), ScraperAPI (1,000 credits/month), Bright Data (5,000 requests/month), Tavily (1,000 credits), Exa ($10 credit), Apify ($5 platform credit), Browserbase (free plan). ScrapingBee and Zyte offer only one-time trials.

Judge. Winner: rose. Answer A matches all seven core providers from the reference (including Exa and Browserbase, which Answer B omits entirely) with accurate, well-sourced figures, whereas Answer B introduces several unverified or likely fabricated vendors while missing key reference items.

Codex alone: 2/3 · 71 s · 336,068 tokens

Tools: web_search ×5

Judge: Exa ($10 credit) is not mentioned, though it's a key vendor in the reference list. Judge: Browserbase (free plan) is not mentioned, though it's a key vendor in the reference list.

Verified as of October 1, 2026: these services advertise a monthly-renewing free tier without requiring a credit card.

API	Recurring free allowance	Important limitation
Diffbot Extract API	10,000 credits/month	Extracting one page costs 1 credit; limited to 5 requests/minute. Bulk crawling requires a paid plan.
ScrapingAnt	10,000 credits/month	Credit cost varies with rendering, proxies, and extraction features.
ZenRows	5,000 credits/month	Advanced proxies and JavaScript rendering use credit multipliers.
Bright Data Web Scraper API	5,000 records/month	Best suited to its prebuilt scrapers; Browser API and standalone proxy products are separate.
Firecrawl	1,000 credits/month	Approximately 1,000 basic page scrapes; structured extraction and browser interaction can cost more.
ScraperAPI	1,000 credits/month	First seven days include a temporary 5,000-credit allowance; afterward it drops to 1,000/month.
Scrape.do	1,000 successful calls/month	Includes rendering, residential/mobile proxies, geotargeting, and CAPTCHA handling.
Tavily	1,000 credits/month	Primarily agent-oriented search, extract, crawl, and map APIs; endpoint costs vary.
Browserless	1,000 units/month	Browser automation rather than a conventional HTML-only scraping API; proxy traffic consumes additional units.
Context.dev	500 credits/month	Basic scrape is 1 credit; structured extraction and brand retrieval cost more.
Apify	$5 platform credit/month	Flexible scraping platform rather than one uniform API; actual page capacity depends on the Actor and compute usage.

A few newer/smaller providers also make the same recurring, no-card claim: Scrapio at 1,000 credits/month, LucidExtractor at 500 credits/month, and Datpaq at 250 requests/month.

Not recurring: ScrapingBee, Scrapfly, and Zyte currently provide signup or trial credits, not a permanently renewed monthly allowance.

For most projects, I’d start with:

• Firecrawl for Markdown, crawling, and RAG ingestion.
• Scrape.do or ScrapingAnt for protected or JavaScript-heavy pages.
• Diffbot for high-volume structured page extraction.
• Browserless when you need Playwright/Puppeteer-style interaction.

Codex + Rose: 3/3 · 25 s · 89,570 tokens

Tools: rose ×1 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "page", "monthly_volume": 1, "min_free_per_month": 1, "require": ["free_without_card"]})

Seven APIs have a recurring monthly free allowance with no credit card, according to Rose’s vendor data as of October 1, 2026:

API	Monthly free allowance
Exa	Up to 10,000 fetched pages from $10 shared credits
Tavily	Up to 5,000 basic extracted pages from 1,000 shared credits
Bright Data	5,000 shared credits
Apify	$5 platform credit; approximately 2,192 pages using its RAG Web Browser
Firecrawl	1,000 credits/pages
ScraperAPI	1,000 API credits/standard pages
Browserbase	1,000 Fetch API calls, plus one browser hour

Allowances may be shared across products, and JavaScript rendering, premium proxies, structured extraction, or other advanced features can consume multiple credits per page.

l3: Among Firecrawl, Exa, Tavily, Brave, SerpApi, Serper, Jina, Linkup, Parallel, ScrapingBee, ScraperAPI, Zyte, Bright Data, Apify and Browserbase, which sell search only, and which also fetch page content?

Gold facts. Search only: Brave, SerpApi, Serper. Search and page content: Firecrawl, Exa, Tavily, Jina (Reader + Search), Linkup, Parallel (Search + Extract), ScrapingBee, ScraperAPI, Zyte, Bright Data (SERP + Web Unlocker), Apify (Actors), Browserbase (Search + Fetch).

Judge. Winner: base. Both reach the correct classification, but Answer B provides clearer per-vendor justification with a well-organized table and more precise capability descriptions, making it more actionable for the buyer.

Codex alone: 3/3 · 120 s · 876,286 tokens

Tools: web_search ×12

Using “fetch page content” to mean retrieving the body of an arbitrary result URL—not merely SERP snippets or the search-results HTML:

Search only

• Brave — URLs, snippets, enriched search context; no arbitrary-URL page-fetch API. Docs
• SerpApi — structured search-engine results, not destination-page content. Docs
• Serper — Google SERP endpoints only; no first-party arbitrary-page fetcher. Site

Search and page-content fetching

Provider	Page-content capability
Firecrawl	Scrape/crawl APIs; search can optionally return full result content. Docs
Exa	Contents API returns full text, highlights, or summaries from URLs; content can be requested during search. Docs
Tavily	Extract API plus include_raw_content in search. Docs
Jina	Reader fetches arbitrary URLs; Search returns result content. Docs
Linkup	Separate Fetch endpoint returning cleaned Markdown or raw HTML. Site
Parallel	Extract API for URL content; Search also returns substantial relevant excerpts. Docs
ScrapingBee	Google Search API plus arbitrary-URL HTML/Markdown scraping. Docs
ScraperAPI	Google SERP API plus a general URL-scraping API. Docs
Zyte	Search API plus /extract for arbitrary URLs and rendered content. Docs
Bright Data	SERP API plus Web Unlocker and Web Scraper APIs. Overview
Apify	Search-result Actors plus website-content crawlers and arbitrary-page scrapers. Docs
Browserbase	Search returns lightweight results; separate Fetch API returns HTML, Markdown, or JSON. Search and Fetch

So the short answer is: Brave, SerpApi, and Serper are search-only; the other twelve also sell page-content retrieval.

Codex + Rose: 3/3 · 45 s · 226,945 tokens

Tools: rose ×6 Rose calls: compare_offers({"niche": "web-search-scraping", "monthly_volume": 1000, "vendors": ["Firecrawl", "Exa", "Tavily", "Brave", "SerpApi", "Serper", "Jina", "Linkup", "Parallel", "ScrapingBee", "ScraperAPI", "Zyte", "Bright Data", "Apify", "Browserbase"], "unit": "search"}), compare_offers({"niche": "web-search-scraping", "monthly_volume": 1000, "vendors": ["Firecrawl", "Exa", "Tavily", "Brave", "SerpApi", "Serper", "Jina", "Linkup", "Parallel", "ScrapingBee", "ScraperAPI", "Zyte", "Bright Data", "Apify", "Browserbase"], "unit": "page"}), get_offers({"brand": "Jina AI"}), get_offers({"brand": "Brave"}), get_offers({"brand": "Serper"}), get_offers({"brand": "SerpApi"})

Using “fetch page content” to mean an arbitrary URL-to-content endpoint:

• Search only: Brave Search API, SerpApi, Serper.
• Search + page-content fetching: Firecrawl, Exa, Tavily, Jina AI, Linkup, Parallel, ScrapingBee, ScraperAPI, Zyte, Bright Data, Apify, and Browserbase.

SerpApi can return the search-results page itself in formats such as HTML or Markdown, but Rose does not classify it as an arbitrary webpage-fetching product. Apify and Bright Data provide the two capabilities through separate actors/products.

These classifications come from Rose’s normalized vendor data as of October 1, 2026.

l4: Which web search or scraping APIs for agents have a SOC 2 report?

Gold facts. Non-exhaustive; must include most of: Firecrawl, Exa, Tavily, Brave, SerpApi, Linkup, Parallel, ScrapingBee (SOC 2 Type II), Bright Data, Apify, Browserbase. Not shown for Serper, ScraperAPI or Zyte (ISO 27001 instead).

Judge. Winner: rose. Answer A matches the reference list exactly (including Linkup) and correctly flags Jina, Serper, ScraperAPI, Zyte as unconfirmed, while explicitly sourcing Rose's vendor data; Answer B omits Linkup and adds several unverified/likely inaccurate vendors (Jina, Perplexity, You.com, Oxylabs, Valyu) without grounding in Rose's data.

Codex alone: 3/3 · 86 s · 557,393 tokens

Tools: web_search ×9

Yes—several agent-oriented web search and scraping APIs publicly confirm current SOC 2 Type II reports. Verified as of October 1, 2026:

Provider	Capability	SOC 2 evidence / access
Tavily	Search, extract, crawl, research	Type II; 2026 report available by request through its Trust Center
Exa	Search, page contents, crawling, research	Type II; current reports published in its Trust Center, according to Exa’s enterprise page
Brave Search API	Independent-index web search	Type II specifically scoped to the Search API; report available under NDA through its Trust Center
Parallel	Search, extraction, structured research	Type II; its Trust Center explicitly lists the report
SerpApi	Google and other search-engine results	Type II covering its production platform and APIs; see SerpApi Security
Firecrawl	Search, scrape, crawl, extract	Type II; independently audited, documented on its enterprise page
Apify	Scraping, crawling, browser automation	Type II; report request available through the Apify Trust Center
Browserbase	Search, fetch, cloud-browser automation	Type II; report available from its Trust Center, per Browserbase Enterprise
Bright Data	Scraping APIs, datasets, proxies	SOC 2 report available by contacting the vendor through its Trust Center
Oxylabs	Scraper API and Web Unblocker	Type II specifically covering those products; report access through the Oxylabs Trust Center
ScrapingBee	Rendered-page scraping and extraction	Type II; its DPA contemplates providing the report, and the vendor publicly announced completion. See DPA
Valyu	Search, URL contents, answers, deep research	Type II advertised for enterprise customers on its pricing page
Jina AI	Web search and URL-to-LLM reader	Claims Type I and Type II coverage on the Jina AI site

Also worth checking, but verify report scope before procurement:

• Perplexity API: Perplexity states it is SOC 2 Type II audited, but its public material does not clearly show whether every API product and plan is within scope. Security page
• You.com API: its security addendum says a Type II report is available on request, but you should confirm that the developer API is explicitly in the audited system boundary. Security addendum

For vendor selection, request the actual report and check:

• The API product is named in the system description.
• The audit period is current, or there is a bridge letter.
• Exceptions and qualified opinions.
• Which Trust Services Criteria were covered.
• Subservice organizations and complementary user controls.
• Whether zero-data-retention applies to your pricing tier.

SOC 2 reports are usually confidential and gated behind an NDA; a badge or “SOC 2 compliant” statement alone is insufficient for vendor diligence.

Codex + Rose: 3/3 · 28 s · 132,703 tokens

Tools: rose ×2 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "page", "monthly_volume": 1000, "require": ["soc2"]}), compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 1000, "require": ["soc2"]})

Rose’s vendor data identifies these providers as having SOC 2 coverage:

• Apify
• Brave Search API
• Bright Data
• Browserbase
• Exa
• Firecrawl
• Linkup
• Parallel
• ScrapingBee
• SerpApi
• Tavily

SOC 2 was not publicly confirmed in Rose’s sheets for Jina AI, Serper, ScraperAPI, or Zyte. “Has SOC 2” does not necessarily mean the full report is publicly downloadable; vendors commonly provide it through a trust center, under NDA, or upon request.

Figures come from Rose’s normalized vendor data as of October 1, 2026.

w1: What's the difference between Browserbase, Firecrawl and SerpApi?

Gold facts. Browserbase hosts headless browsers for agents (billed by browser hours and proxy bandwidth), plus newer Search and Fetch APIs. Firecrawl turns websites into LLM-ready markdown or structured data (scrape, crawl, search), billed in credits. SerpApi returns search engine results pages (Google, Bing and 100+ engines) as structured JSON, sold as monthly search plans.

Judge. Winner: rose. Answer A matches the reference facts more completely by explaining each tool's billing model (browser hours/proxy bandwidth, credits, search plans) and backs it up with concrete, sourced pricing figures and free-tier details, giving the buyer more actionable decision-making information than Answer B, which omits pricing entirely.

Codex alone: 2/3 · 23 s · 88,930 tokens

Tools: web_search ×2

Judge: Missing that Browserbase is billed by browser hours and proxy bandwidth Judge: Missing that Firecrawl billing is credit-based Judge: Missing that SerpApi is sold as monthly search plans Judge: Does not explicitly mention SerpApi covers 100+ search engines beyond Google/Bing (though implied)

The simplest distinction: Browserbase operates websites, Firecrawl turns websites into data, and SerpApi returns search-engine results.

Product	Primary job	Best for	Typical output
Browserbase	Hosted browser infrastructure and automation	Logging in, clicking buttons, filling forms, downloads, multi-step browser agents, testing flows	Browser sessions, screenshots, extracted page data, session replays
Firecrawl	Scraping, crawling, and content extraction	Converting a page or whole site into clean content for RAG, research, or datasets	Markdown, structured JSON, HTML, links, metadata
SerpApi	Search-engine results APIs	Querying Google, Bing, Maps, Shopping, Flights, News, etc., with location/device controls	Structured SERP JSON or Markdown

A practical example:

1. SerpApi: Find the top 20 websites ranking for “best accounting software.”
2. Firecrawl: Crawl those websites and extract their pricing and feature pages.
3. Browserbase: Log into selected products, navigate their dashboards, and perform interactive tasks.

There is now some overlap:

• Browserbase also offers Search and Fetch APIs, but its distinctive strength remains long-lived, observable browser sessions and interaction.
• Firecrawl also offers search and browser interaction, but its core abstraction is extracting clean, LLM-ready content from URLs or whole sites.
• SerpApi uses browsers internally, but you do not control those browsers—it exposes specialized, structured search results.

Rule of thumb: choose SerpApi for discovery, Firecrawl for ingestion, and Browserbase for action.

Codex + Rose: 3/3 · 33 s · 142,896 tokens

Tools: rose ×5 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "page", "monthly_volume": 10000, "vendors": ["Browserbase", "Firecrawl", "SerpApi"]}), get_offers({"brand": "Browserbase"}), get_offers({"brand": "Firecrawl"}), get_offers({"brand": "SerpApi"}), compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 10000, "vendors": ["Browserbase", "Firecrawl", "SerpApi"]})

They operate at different layers:

Tool	Best thought of as	Choose it when…
Browserbase	Managed cloud browsers	An agent must interact with websites: navigate, click, log in, execute JavaScript, take screenshots, or automate workflows. It also offers simpler Search and Fetch APIs.
Firecrawl	Web-content extraction and crawling	You need clean Markdown/JSON from pages, whole-site crawling, URL discovery, structured extraction, or search results followed by usable page content.
SerpApi	Search-engine results API	You specifically need structured Google, Bing, Maps, Shopping, or other search-engine results—including SERP-specific fields such as rankings and knowledge panels.

A practical rule:

• “Operate this website” → Browserbase
• “Turn this website into LLM-ready data” → Firecrawl
• “Tell me what Google/Bing returned” → SerpApi

There is some overlap: all three expose search-related functionality, while Browserbase and Firecrawl both support JavaScript-heavy sites and proxies. But their centers of gravity remain browser automation, content extraction, and SERP data respectively.

For an illustrative 10,000 searches/month, Rose calculates approximately:

• Browserbase: $83
• Firecrawl: $94
• SerpApi: $150 through two 5,000-search Developer-plan renewals

For 10,000 basic page fetches, it calculates Browserbase at $29 and Firecrawl at $44; SerpApi is not a general page-crawling product, so that comparison does not apply.

All have card-free recurring free tiers: Browserbase includes 1,000 Search/Fetch calls, Firecrawl has 1,000 shared credits—roughly 500 searches or 1,000 basic pages—and SerpApi includes 250 searches.

These figures come from Rose’s normalized vendor data as of October 1, 2026: Browserbase pricing, Firecrawl pricing, and SerpApi pricing.

w2: I need raw Google search results as structured JSON, not page content. Which vendors do that?

Gold facts. Non-exhaustive; SERP APIs: SerpApi, Serper, Bright Data SERP API, Zyte SERP, ScrapingBee Google Search API, ScraperAPI (structured Google SERP endpoint). Brave returns its own index, not Google.

Judge. Winner: tie. Answer B covers more of the reference vendors (including Zyte-equivalent coverage via additional structured SERP providers like DataForSEO, Oxylabs, SearchApi, Zenserp, Scale SERP) and adds useful actionable context on Google's own Custom Search API deprecation, making it more complete and actionable without sacrificing correctness.

Codex alone: 2/3 · 46 s · 190,825 tokens

Tools: web_search ×4

Judge: Claims Google's Custom Search JSON API is 'closed to new customers and scheduled for discontinuation on January 1, 2027' — this specific deprecation date is not substantiated and appears to be a fabricated detail not supported by Google's actual documentation. Judge: Reference lists ScrapingBee Google Search API, Zyte SERP, and ScraperAPI as additional correct vendors; answer omits these (acceptable since list is non-exhaustive, but worth noting).

You want a Google SERP API: it returns the result page itself as parsed JSON—rank, URL, title, snippet, ads, local results, People Also Ask, knowledge panels, etc.—without fetching destination pages.

Vendor	Best fit
SerpApi	Mature, comprehensive Google SERP coverage and clean JSON; easiest default
Serper	Fast, inexpensive, simple API; good for organic results and AI pipelines
DataForSEO	Bulk SEO/rank-tracking workloads; extensive localization and SERP features
SearchApi	Broad Google vertical coverage and straightforward structured responses
Bright Data	High-volume enterprise infrastructure; parsed JSON available
Oxylabs	Enterprise-scale SERP collection with structured output and precise geolocation
Zenserp	Lower-cost conventional SERP API with batch support
Scale SERP	Another established option supporting structured JSON, HTML, or CSV

My shortlist:

• Best overall: SerpApi
• Best price/speed for ordinary searches: Serper
• Best for bulk SEO data: DataForSEO
• Best for enterprise procurement/compliance: Bright Data or Oxylabs

Avoid Tavily, Exa, and generic scraping APIs if exact Google rankings matter: those are search/content-retrieval products, not faithful Google SERP feeds.

Google’s own Custom Search JSON API is not a practical new option: it is closed to new customers and scheduled for discontinuation on January 1, 2027. It also represents Programmable Search rather than the complete consumer Google SERP. Google documentation

Codex + Rose: 2/3 · 32 s · 136,467 tokens

Tools: rose ×7 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "search", "monthly_volume": 1000, "vendors": ["SerpApi", "Serper", "ScrapingBee", "Bright Data", "Apify", "ScraperAPI"]}), get_offers({"brand": "Serper"}), get_offers({"brand": "ScrapingBee"}), get_offers({"brand": "SerpApi"}), get_offers({"brand": "Bright Data"}), get_offers({"brand": "ScraperAPI"}), get_offers({"brand": "Apify"})

Judge: Missing Zyte SERP from the list of vendors offering structured Google SERP JSON.

Done — six covered vendors return parsed Google SERPs as structured JSON rather than page content:

• SerpApi — purpose-built, detailed JSON for organic results, ads, knowledge panels, maps, shopping, and more.
• Serper — streamlined Google JSON API covering web, images, news, maps, shopping, places, and Scholar.
• Bright Data SERP API — structured JSON with high-volume infrastructure and geotargeting.
• ScrapingBee Google Search API — JSON SERPs alongside its broader scraping API.
• ScraperAPI — parsed JSON for Google/Bing SERPs, although it is primarily a general scraping platform.
• Apify Google Search Results Scraper — structured dataset output via an Actor; more workflow-oriented than a simple search endpoint.

My shortlist: Serper for low-cost simplicity, SerpApi for the richest schema and broadest SERP-feature coverage, and Bright Data for scale.

These figures and capabilities come from Rose’s vendor data as of October 1, 2026. General search APIs such as Exa, Tavily, Brave, Linkup, Parallel, and Jina were excluded because they return their own search indexes or synthesized results—not raw Google SERPs.

w3: What do Jina's r.jina.ai and s.jina.ai do?

Gold facts. r.jina.ai (Reader) turns a URL into LLM-friendly markdown or text; s.jina.ai (Search) searches the web and returns the top results as LLM-friendly content. Billed in tokens; Jina AI is part of Elastic.

Judge. Winner: tie. Answer B is equally accurate but more complete and actionable, providing a concrete query-based example for s.jina.ai, noting PDF support for r.jina.ai, and avoiding the questionable claim in A that s.jina.ai strictly requires an API key.

Codex alone: 2/3 · 17 s · 86,862 tokens

Tools: web_search ×2

Judge: Does not mention billing is in tokens Judge: Does not mention Jina AI is part of Elastic

• r.jina.ai — Reader: Fetches a specific public URL, renders it if necessary, removes navigation/ads/boilerplate, and returns the main content as clean, LLM-friendly Markdown. It also supports PDFs.

https://r.jina.ai/https://example.com/article

• s.jina.ai — Search: Searches the web for a query and returns the top results—including their URLs and extracted page contents—in LLM-friendly text.

https://s.jina.ai/?q=latest+developments+in+fusion+energy

In short: r reads a URL; s searches for URLs and reads the results. They’re primarily grounding infrastructure for agents, RAG systems, and LLM applications—not embedding or generative-answer endpoints. Jina Reader documentation

Codex + Rose: 2/3 · 14 s · 89,487 tokens

Tools: rose ×1 Rose calls: get_offers({"brand": "Jina AI"})

Judge: Does not mention billing is in tokens Judge: Does not mention Jina AI is part of Elastic

• r.jina.ai — Reader: Fetches a webpage and converts it into clean, LLM-friendly Markdown or text. Typically, prefix a URL: https://r.jina.ai/https://example.com.
• s.jina.ai — Search: Searches the web for a query and returns search results as LLM-friendly text. It requires a Jina API key.

In short: r reads a known URL; s searches for relevant URLs/content. Reader supports JavaScript rendering and outputs including Markdown, HTML, text, JSON, and screenshots. Jina Reader documentation

These details come from Rose’s Jina AI vendor data as of October 1, 2026.

w4: My agent needs to fetch pages behind anti-bot protection (Cloudflare, CAPTCHAs). Which APIs are built for that?

Gold facts. Non-exhaustive; must include several of: Bright Data Web Unlocker (CAPTCHA solving, proxy management), Zyte API (automatic ban handling), ScraperAPI (rotating proxies, anti-bot bypass for Cloudflare/DataDome), ScrapingBee (rotating, premium and stealth proxies), Browserbase (hosted browsers with proxies).

Judge. Winner: rose. Answer A covers more of the reference's required providers (Bright Data, Zyte, ScraperAPI, ScrapingBee, Browserbase) with concrete, sourced pricing details, whereas Answer B substitutes several non-reference vendors (Browserless, Oxylabs, ZenRows) and omits ScraperAPI, ScrapingBee, and Browserbase.

Codex alone: 3/3 · 41 s · 137,532 tokens

Tools: web_search ×3

Done — several “web-unblocker” APIs are specifically designed for Cloudflare, CAPTCHAs, fingerprinting, proxy rotation, and JavaScript rendering.

Best options:

• Browserless — strongest fit for an interactive agent. Its /unblock endpoint can return HTML, cookies, screenshots, or a live WebSocket endpoint that your agent continues controlling through Playwright/Puppeteer. BrowserQL adds stealth, residential proxies, and CAPTCHA solving.
• Bright Data Web Unlocker — good for “URL in, HTML/JSON out.” It manages retries, proxy rotation, browser fingerprints, JavaScript, Cloudflare Turnstile, and other CAPTCHAs. For clicks and multi-step navigation, use its Browser API.
• Zyte API — polished unified API offering HTTP fetching, browser rendering, extraction, sessions, geolocation, and direct Playwright/Puppeteer access over CDP.
• Oxylabs Web Unblocker — proxy-compatible option that handles IP rotation, fingerprinting, retries, CAPTCHA bypass, and optional server-side JavaScript rendering.
• ZenRows — relatively simple API with residential routing, browser sessions, and handling for Cloudflare Turnstile, reCAPTCHA, and hCaptcha.

My practical recommendation:

• For an agent that must click, log in, paginate, or maintain state: start with Browserless BrowserQL/BAP or Bright Data Browser API.
• For mostly fetching individual protected pages: start with Bright Data Web Unlocker or Zyte API.
• Run the same representative 20–50 URLs through two providers before choosing. Success rates vary dramatically by target, geography, session behavior, and protection configuration.

Avoid relying solely on a CAPTCHA-token service such as 2Captcha. Cloudflare commonly evaluates IP reputation, TLS/browser fingerprints, cookies, and behavior in addition to the challenge, so a managed unblocker or stealth browser is usually the correct abstraction.

Use these only where you’re authorized to automate access, and respect the site’s terms, rate limits, privacy obligations, and access controls.

Codex + Rose: 3/3 · 50 s · 184,667 tokens

Tools: rose ×7 Rose calls: compare_offers({"niche": "web-search-scraping", "unit": "page", "monthly_volume": 10000, "require": ["js_rendering", "proxies"]}), find_alternatives({"need": "API for an AI agent to fetch JavaScript-rendered pages behind Cloudflare and CAPTCHA or other anti-bot challenges using managed proxies or hosted browsers"}), get_offers({"brand": "Bright Data"}), get_offers({"brand": "Zyte"}), get_offers({"brand": "ScraperAPI"}), get_offers({"brand": "ScrapingBee"}), get_offers({"brand": "Browserbase"})

The strongest fits are Bright Data Web Unlocker, Zyte API, and ScraperAPI. All expose agent-friendly APIs and official MCP servers.

• Bright Data Web Unlocker — best turnkey choice: proxy management, JavaScript rendering, and explicit CAPTCHA solving. Pay-as-you-go is listed at $1.50/1,000 successful requests, with 5,000 shared free monthly credits. Pricing
• Zyte API — best cost-efficient managed scraper: automatically handles bans and chooses pricing by site difficulty; browser-rendered requests range from $1.01–$16.08/1,000 on PAYG. Pricing
• ScraperAPI — clearest explicit anti-bot coverage: specifically advertises bypass for Cloudflare, DataDome, PerimeterX, and Akamai. Its $49 Hobby plan provides 100,000 credits; protected requests consume more credits. Pricing
• ScrapingBee — simple developer API: managed browsers, premium proxies, and a higher-cost “stealth proxy” mode for difficult sites. Pricing
• Browserbase — best when the agent must operate a real browser: useful for sessions, navigation, logins, and browser actions. Use hosted browser sessions—not its lightweight Fetch endpoint—when JavaScript or interaction is required. Pricing
• Firecrawl — best downstream format for LLMs: good if you primarily want clean Markdown or structured extraction, though it is less specialized in difficult anti-bot challenges than the first three. Pricing

My default would be Bright Data for maximum hands-off reliability, Zyte for cost-sensitive volume, or Browserbase when your agent needs interactive browser control.

These figures come from Rose’s normalized vendor data as of October 1, 2026. No provider guarantees every CAPTCHA; only access pages you’re authorized to retrieve and respect applicable terms and rate limits.