Web Search for AI Agents: Prices and Indexes in 2026

Web search for AI agents is being repackaged and repriced away from web search for people. The two APIs most developers built on are closing: Bing's retired in 2025, Google's ends January 2027. A layer priced at $1 to $8 per thousand agent queries takes over.

TL;DR

August 2026 stacked several pieces of the same story. Keenable came out of stealth on August 25 with a $26 million seed led by Accel and an index of more than 100 billion documents.¹ Two days later it published NEEDLE, a live open-source benchmark for agentic search that scores seven engines including its own.² Meanwhile AWS spent the month extending its in-house index in both directions: to ordinary Bedrock inference calls on August 4, and to AgentCore agents in Europe and Asia Pacific with per-call domain and date filters on August 19.^{3 4}

Underneath it: the search APIs agent builders inherited were sold to developers who wanted ranked links. Agents want something else, and are being charged differently for it.

What you'll learn

• Why the two search APIs most developers built on are being withdrawn, and the exact dates
• What agent-facing web search actually costs today, per vendor, from official pricing pages
• Why the per-query price is the smallest line in your search bill, with the arithmetic
• What AWS actually shipped — two products, not one — and what external_web_access really does
• What NEEDLE measures, and how to read a benchmark published by one of its own contestants
• The shared-mistake test for telling whether two search APIs are really two indexes
• What published research says about agent search traffic — and what it does not say
• Which questions to ask before you wire a search tool into an agent loop

The general-purpose search APIs are closing

This is the part that makes the rest inevitable.

Microsoft retired the Bing Search APIs on August 11, 2025. The lifecycle notice is blunt: existing instances "will be decommissioned completely, and the product will no longer be available to be used or new customer signup." Customers were pointed at Grounding with Bing Search inside Azure AI Agents.⁵ The raw-results API did not get a raw-results successor. It got an agent-grounding tool bolted to Microsoft's agent platform.

Google is running the same play on a longer clock. The Custom Search JSON API is, in Google's own words, "closed to new customers," and the documentation states that "existing Custom Search JSON API customers have until January 1, 2027 to transition to an alternative solution."⁶ The suggested replacement is Vertex AI Search, which Google's own page scopes to "searching up to 50 domains"; if you need full web search, the page hands you a Google Form to register interest.⁶

Two caveats worth stating, because most coverage skips them. Neither product was ever a human search API — both were developer APIs used heavily by SEO tooling, price monitors and metasearch, and Bing's died for commercial reasons as much as technical ones. And "the only two" would be wrong: Brave sells a general-purpose index API today,⁷ and Yandex still operates one.⁸ What is accurate is narrower and still significant: the two APIs that most Western developers reached for by default are being withdrawn, and neither successor is a drop-in.

What agent search costs in 2026

Every figure below was read off the vendor's own pricing page or launch post while writing this. They are not like-for-like, and the ways they differ are set out immediately after the table.

⚠ Prices change frequently. The values above are for illustration only and may be out of date. Always verify current pricing directly with the provider before making cost decisions: Anthropic · OpenAI · Google Gemini · Google Vertex AI · AWS Bedrock · Azure OpenAI · Mistral · Cohere · Together AI · DeepSeek · Groq · Fireworks AI · Perplexity · xAI · Cursor · GitHub Copilot · Windsurf.

What this table leaves out, deliberately. Grounding with Bing Search and Vertex AI Search — the two successors named above — are not here, because both are sold as platform features rather than metered search endpoints. Neither is OpenAI's hosted web search tool. Neither are Serper or SearchAPI, the Google and Bing resellers that NEEDLE itself uses as proxies for those engines. A shortlist built only from this table is incomplete.

Read the conditions, not the numbers. Keenable's $1 requires sustained throughput of 100 requests per second — roughly 260 million requests a month — and buys dedicated capacity rather than metered calls, so the effective rate for anyone under-utilising it is higher.¹⁰ For everyone else Keenable's price is the $4 row. Tavily's $8 is its worst rate: monthly plans run from $0.0075 down to $0.005 per credit, so the same basic search costs $5 per 1,000 on the $500/month Growth tier — while an advanced search is two credits, or $16 per 1,000 at the pay-as-you-go rate.¹⁴ And the whole spread only holds at ten results: Parallel's turbo at 17 results is $1 + 7 × $1 = $8 per 1,000, which overtakes AWS.⁹

Two things the table makes visible that most write-ups miss.

Google Custom Search is not in this market. At $5 per 1,000 with a hard ceiling of 10,000 queries a day, its maximum possible spend is about $1,500 a month.⁶ Keenable's free tier alone is a third of that volume ceiling.¹⁰ The incumbent is not being outpriced; it is being outgrown.

The free tiers vary more than the prices do. Brave's $5 monthly credit is about 1,000 requests; Exa's standing offer is $10 a month in credits; Keenable gives away 100,000 requests a month, which at its own $4 rate is $400 of list value, from a company days out of stealth.^{7 11 10} A hundred-to-one spread in free allowance is a customer-acquisition strategy, not a pricing detail, and it is the clearest signal of who is buying market share.

The per-query price is the smallest line in your bill

Here is the arithmetic that should govern the decision, and that price tables rarely show.

Take AWS AgentCore at $7 per 1,000 queries — $0.007 a call.¹² An agent doing real research issues searches in bursts inside a loop. At 300 searches for a single deep task, the search fee is $2.10. That is already enough to change an architecture.

Then add what the search returns. A provider that hands back ten pages of extracted content, rather than ten links and snippets, is putting those pages into your model's context window and you are paying inference on every token. Ten pages at a few thousand tokens each is tens of thousands of input tokens per call. Priced at even a dollar per million input tokens, the context cost of one thousand such calls comfortably exceeds the entire thousand-call search fee — and that multiple grows with the model you are running.

Which is why the "what that buys" column matters more than the price column. Exa's base rate includes page contents; AWS returns snippets with source URLs; Tavily bills extraction separately from search; Keenable splits search_web_pages from fetch_page_content.^{11 13 14 15} Whether "a search" includes "reading the page" is a several-fold cost swing that no headline rate captures. It is the same trap as AI agent token costs in 2026: the per-call price stops being interesting once one task becomes hundreds of calls carrying payloads.

What AWS actually shipped

AWS has two products called Web Search, and coverage keeps merging them. They are separate, with separate launch dates.

Web Search on Bedrock AgentCore went generally available on June 17, 2026, initially in US East (N. Virginia).¹³ It is a built-in connector target on the AgentCore Gateway over the Model Context Protocol — the agent calls it as an MCP tool and gets back snippets, source URLs, titles and publication dates.¹³ If you have been following MCP's move to a stateless protocol, this is that pattern arriving as a first-party cloud service. It is the version priced at $7 per 1,000 queries.¹²

On August 19 it gained per-call domain filtering and published-date filtering: agents can pass include and exclude domain lists and a date range in each tool call, admins get gateway-level allowlists, and the cap is 100 domains per list. The same announcement expanded the tool to Europe (Ireland) and Asia Pacific (Tokyo), joining US East.⁴ Per-call domain allowlisting is quietly the most useful agent-safety primitive in this whole roundup, and it is buried in a What's New post.

Web Search on Amazon Bedrock is the other product, generally available since August 4, 2026 for OpenAI models running on Bedrock — GPT-5.4, GPT-5.5 and GPT-5.6 Sol, Terra and Luna. It attaches to an ordinary inference call through "a standardized tool-use interface, compatible with the OpenAI Responses API," with no agent gateway required.³ Its pricing lives on the Bedrock pricing page rather than in the announcement, so treat any figure you see quoted for it with suspicion until you have checked that page yourself — the figures circulating in secondary coverage are uncited.

Both sit on the same foundation, and that is the actual news. AWS describes it as "a web index operated by Amazon, spanning tens of billions of documents refreshed continually," combined with a knowledge graph of verified facts, informed by Amazon's work on Alexa+, Amazon Quick and Kiro.^{3 13} "Operated by" is Amazon's phrasing and is not quite the same as "crawled by," but there is no indication Bing or Google sit underneath it.

Then the part everyone is describing backwards. Also on August 19 — a separate announcement from the AgentCore one above — AWS shipped external_web_access for Web Search on Bedrock, under the title "Launching External Web Access." The news was adding the ability to reach the live public web: "we are expanding Web Search to enable the external_web_access parameter allowing Web Search to retrieve content directly from the public web."¹⁶ It defaults to true and requires the IAM permission bedrock-websearch:ExternalWebAccess. Setting it to false restores the earlier behaviour, in which "Web Search serves results only from Amazon's in-AWS web index and knowledge graph, with no request data leaving the AWS boundary."¹⁶

So the compliance story is real but inverted from how it is usually told: AWS shipped an in-boundary product first and has now added an escape hatch to the open web, with the boundary preserved as an opt-out. Worth noting the footprint, since neither announcement dwells on it — Web Search on Amazon Bedrock is available only in US East (N. Virginia), US East (Ohio) and US West (Oregon), and enabling external web access is scoped to those same three regions.^{3 16} An EU customer with a residency requirement has no in-region option here at all yet; the AgentCore tool is the one that reaches Ireland and Tokyo.⁴ This is the retrieval-side version of the containment questions raised in our coverage of AWS closing Bedrock Agents, and the trade-off is stark: freshness comes from the public web, and false switches the public web off.

Keenable, and the case for owning the index

Keenable's pitch is that federating someone else's results has a ceiling. Co-founder and CEO Andrey Styskin previously led search at Yandex and worked on web infrastructure at Amazon; co-founder Matthias Petri was a principal applied scientist at Amazon.^{1 15} The company says its index passes 100 billion documents and is already in production at several unnamed AI labs and inference providers, during both training and runtime — a company claim, relayed via TechCrunch, with no named customer attached.¹

Index size is the specification vendors in this market most readily volunteer, though only some do, and the numbers are neither audited nor measured in a common unit:

• Keenable: "more than 100 billion documents"¹
• Brave: "over 30 billion pages, kept fresh by over 100 million page updates every day"⁷
• Amazon: "tens of billions of documents refreshed continually," plus a knowledge graph³

Exa, Parallel, Tavily and Google publish no index size on the pages cited here. And the list carries two warnings of its own. Brave's page states two different figures depending on where you look — the body says "over 30 billion pages," while the page's own meta description advertises "40+ billion pages," which is the version that shows up in search and social previews. And Amazon, with effectively unlimited crawl budget and a consumer assistant to feed, reports a smaller number than a seed-stage startup. Either the units differ substantially or the coverage does. Take none of these figures as comparable.

Styskin's argument to TechCrunch was about cost structure rather than quality: "If you do not fine-tune your index structures for a specific task, the cost of serving and scanning the whole internet is enormous because of the volume."¹ The company is 15 engineering staff across the US and Europe and plans to use the round to double headcount by year end, explicitly to build a go-to-market motion.¹ Two products are on the roadmap: a Web Query Language for answering questions no single page contains, and Time Machine, a point-in-time search API where adding a query_time parameter rewinds both the corpus and the ranking.^{1 15}

Its MCP endpoint is live at https://api.keenable.ai/mcp, exposing search_web_pages and fetch_page_content.¹⁵ On latency, the company's engineering post is more careful than its homepage: the blog says "p50 is 200ms today, and we're targeting 200ms p95," while the homepage advertises "<250ms p95 (US East)" without marking it as an aspiration.^{2 15} Neither page defines what is being measured — server-side or end to end — and for an agent loop the numbers that actually bite are p99 and timeout behaviour, which none of the vendor pages surveyed here publish.

The counter-case deserves stating, because Keenable will not state it. Federation is cheap, requires no crawl capital, inherits decades of ranking investment, and does not disappear if a 15-person company does. Owning an index also means owning its robots.txt posture and whatever publisher litigation follows. "Build your own index" is the correct strategy for a company selling an index; it is not automatically the correct thing for you to depend on.

NEEDLE: a benchmark published by one of its contestants

On August 27 the Keenable team published NEEDLE, a live, open-source search benchmark with the evaluation code on GitHub.² It runs five verticals: News (regenerated hourly from RSS feeds and Google Trends), Finance (Wikidata, GLEIF, SEC XBRL), Scholar, AgenticRare (rare-word queries sampled from public agentic logs), and Legal. It scores Keenable against Google via Serper, Bing via SearchAPI, Brave, Tavily, Parallel and Exa.² (The post expands the acronym as "News, Everyday, Expert, Deep-tail, and Legal Evaluation," which only partly lines up with the vertical names it then uses — a small inconsistency in Keenable's own write-up.)

Read it with the obvious caveat in hand: the benchmark's authors sell one of the seven engines being scored. Three things partially offset that. The queries are regenerated daily or hourly from live sources, so there is no static set to tune against. The code and query-collection procedure are public. And the post concedes in its opening paragraph that "every engine, including Keenable, falls short of what's achievable on agentic traffic."²

This article deliberately reports no NEEDLE scores. They are published as charts on a live page that changes daily, and a snapshot of a moving leaderboard — taken from the vendor that runs it — is exactly the kind of number that ages into a false claim. Go read the live page instead, and note where Keenable places before you weigh anything else in this section.

The design idea worth stealing is the ultimate baseline: a synthetic engine that keeps the best result any contestant found, ranked by the same judge that scores everyone. It is explicitly "not a real engine," but it converts a leaderboard into a measure of how much the whole field is collectively missing.² On the AgenticRare set — the vertical closest to real agent traffic — that gap is widest.²

The stated motivation is contamination, and it generalises well past search. Static benchmarks leak: the post notes that search agents can locate a benchmark's own dataset mid-evaluation and read the answer key, and that a model which has memorised a BrowseComp answer skips the "hard to find" half of the task entirely.² Live regeneration is the same fix LiveCodeBench and SWE-bench-Live applied to code.

The shared-mistake test

The sharpest technical idea in the NEEDLE post has nothing to do with ranking quality. It is a method for detecting whether two search APIs are actually two indexes.

Agreement on correct results proves nothing — good engines converge. Agreement on the same wrong results is the tell. As the post puts it: "Two students with the same right answer studied. Two students with the same wrong answer sat next to each other."² Keenable combines raw overlap with shared-mistake rate to argue that some engines in its comparison serve a common upstream index.²

The historical precedent is real, though the version circulating is garbled. Two separate things happened. In the summer of 2010, by Google's account, it noticed Bing returning Google's result for the misspelled query "torsorophy" without offering the spelling correction. Later, Google ran an actual sting: about 100 "synthetic queries" — nonsense strings such as "hiybbprqag" — each seeded with an unrelated top result, run by around 20 engineers on Internet Explorer 8 with the Bing Toolbar installed and the browser's Suggested Sites feature opted in. Google says the results began surfacing on Bing "within a couple weeks," and it went public in February 2011.^{17 18} Microsoft's public reply was a flat denial titled "We do not copy Google's results," and its broader argument since has been that anonymised clickstream is a legitimate ranking signal.¹⁷

Two honest caveats before you adopt the method. It measures index provenance, not index quality — knowing that two vendors share an upstream tells you nothing about which serves your agent better. And it was built and published by the one vendor in the comparison who benefits from every result it produces. Some of the engines it targets, notably Serper and SearchAPI, are openly resellers; finding shared mistakes with them establishes something nobody disputes.

Still, "do you have your own index?" is a marketing question, and "show me your miss overlap against the other providers I'm evaluating" is an empirical one you can now run yourself, with public code.

What the research says — and does not

An arXiv paper this year, "Agentic Search in the Wild: Intents and Trajectory Dynamics from 14M+ Real Search Requests," takes its title from its sample size.¹⁹ That such a corpus exists at all is the point: agent search traffic is now large enough to study as its own workload.

More useful is the ASQ paper, "A Picture of Agentic Search," submitted in February 2026.²⁰ Its abstract states the problem in infrastructure terms: information retrieval "operates under assumptions that no longer hold in practice, with changes to workload volumes, predictability, and querying behaviours," with the result that "caching may lose effectiveness, query pre-processing may add overhead without improving results, and standard metrics may mismeasure satisfaction."²⁰ The authors release the Agentic Search Queryset, covering reasoning-induced queries, retrieved documents and agent thoughts across HotpotQA, Researchy Questions and MS MARCO for three agents and two retrieval pipelines.²⁰

Note what that does and does not support. It says the assumptions change — not necessarily in the direction vendors would like. Caching is the clearest example: agent loops re-issue near-identical queries and template fan-out produces high duplication, so agentic traffic is plausibly more cacheable than human traffic, not less. Spell correction built for typing errors is largely wasted overhead on agents, which per Keenable's reading of ASQ make roughly three times fewer spelling errors than humans — fewer, not none.² And a metric tuned to whether a person clicked is measuring the wrong thing when nobody clicks.

That is an argument for retuning retrieval to a new workload. It is not, by itself, an argument for buying from a new vendor. The vendors are making that leap; the researchers are not.

What this means if you're wiring search into an agent

1. Is the price per request, per result, or per credit? Exa and Parallel meter results above ten; Tavily bills in credits that differ by endpoint and depth.^{11 9 14} Model your actual call shape, not the headline.
2. What does a result cost you downstream? Extracted page content lands in your context window and you pay inference on it. Compare total cost per completed task on your own query logs, not price per call.
3. What are the rate limits and the terms? Brave publishes 50 QPS; Keenable's $1 rate presumes 100 RPS.^{7 10} Separately, check what each contract permits you to store, cache, re-serve or train on — that clause is often the real differentiator and appears in no price table.
4. Where does it run? Web Search on Amazon Bedrock is US-only in all three of its regions; AgentCore's Web Search reaches Ireland and Tokyo as well.^{3 4} Residency requirements will narrow your shortlist faster than price will.
5. Is it one index or a reseller? Run the shared-mistake comparison across your shortlist. Then run a quality evaluation too — they answer different questions.
6. What is your exit? One provider in this story has already been switched off and another is sixteen months from it. Several of the replacements are very young companies. Prefer an interface you can re-point: MCP is doing useful work here, since AWS, Keenable and Brave all expose one.^{13 15 7}
7. What happens if results are adversarial? A search tool is untrusted input arriving inside your agent's context, as we covered in AI search poisoning. AgentCore's per-call domain allowlists and AWS's external_web_access: false are both usable mitigations — a smaller, curated corpus is a smaller attack surface.^{4 16}

Bottom line

The story is not that a startup raised $26 million. It is that the two search APIs a generation of developers built on are being withdrawn exactly as programmatic search demand explodes — and that neither vendor is replacing its own with a like-for-like successor. What Microsoft and Google are offering instead is grounding tools welded to their agent platforms. The general-purpose endpoints that remain are coming from elsewhere: metered search sold per thousand agent queries by Keenable, Exa, Brave, Parallel and Tavily, plus AWS's two Bedrock tools sitting on Amazon's own index.

Calling this a split between human and agent search is too neat. Amazon's index serves Alexa+ and agents; Brave sells the same index to browser users and to APIs; Microsoft's and Google's grounding tools sit on Bing and Google. What has actually split is packaging, pricing and permitted use — which is a smaller claim, and the one the evidence supports.

The unresolved question is shared by every option on the table: there is still no widely accepted neutral measure of which index is better for an agent. Vendors cite third-party evaluations that favour them, and the most serious open attempt so far was written by a contestant. That gap, not the funding round, is what to watch.