Eighteen months ago, choosing a web search API for an agent was straightforward: use Bing, or use a Google SERP reseller. Both of those defaults are now gone or diminished, and roughly $600 million of venture funding has arrived in their place.
The two things that broke
Microsoft retired the Bing Web Search API on 11 August 2025. Not deprecated with a compatible successor — retired, taking Web, Image, News, Video, Entity, Autosuggest, Spell Check and Custom Search with it. API key creation had been silently disabled that March, weeks before any announcement. The designated migration path, Grounding with Bing Search inside Azure AI Foundry, is a platform commitment rather than a change of base URL, which is why most teams went elsewhere.
Brave ended its perpetual free developer tier in February 2026, replacing it with a $5 monthly credit. The index — over 40 billion pages, genuinely independent of Google and Bing — is unchanged and still excellent. But Brave’s free tier was how a great many projects started, and a project that was free is now a line item.
The money that arrived
| Company | Raised | Valuation |
|---|---|---|
| Exa | $85M Series B (Sep 2025) → $250M (May 2026, a16z) | $2.2B |
| Parallel Web Systems | $230M across three rounds | $2.0B |
| Tavily | $25M, then acquired by Nebius for $275M (Feb 2026) | — |
| Firecrawl | $14.5M Series A | — |
| Linkup | $10M seed (Feb 2026) | — |
Parallel is the most interesting arrival by thesis: founded by Parag Agrawal, the former Twitter CEO, building its own index explicitly for agents rather than for people. It reports over 100,000 developers and customers including Clay, Harvey, Notion and Opendoor.
What this changed about the argument
In 2024, “we run our own index” was a complete pitch. It is not one now. Exa, Parallel and Brave all crawl and index the web themselves, with billions in valuation behind them. Owning an index has become the entry requirement rather than the edge — if a vendor’s pitch stops there, it is a 2024 pitch.
Three things still separate providers:
Price. The field runs from about $0.30 to $35 per 1,000 queries, and the headline rates are not comparable — credits, per-request-plus-token fees, processor tiers and surcharges all mean different things. Normalising on one plain search returning ten results is the only way to make a table meaningful, and we publish that normalisation with a source link on every figure.
Graph traversal. Every provider ranks documents. As of 2026, essentially none let an agent traverse relationships between documents across the open web. Ask ChatGPT about this and it will tell you the graph is typically private and application-specific — you build it over your own corpus with GraphRAG or Neo4j. That gap is real, and it is where we spend our effort.
Coverage transparency. You can ask any of these APIs for a URL and get nothing back. None of them will tell you whether the page was rejected, dropped, or never crawled. For anything auditable, that silence is the problem.
What to do about it
If you are choosing now, the honest advice is to work out which category you actually need before comparing prices. Own-index APIs, platform grounding, retrieval layers, SERP resellers and extraction tools have genuinely different failure modes, and choosing across categories on price is how teams end up rebuilding six months later. We wrote that up as how to choose a web search API.
And assume at least one more of these events happens to whoever you pick. In eighteen months this category saw an API retired, a free tier withdrawn and a vendor acquired. Keep retrieval behind an interface, and do not scatter provider-specific parameters through your codebase.
Every figure in this post is sourced on our landscape page, with a link to the vendor’s own pricing page or funding announcement and the date we read it. We sell one of the products discussed here — weigh the framing accordingly, and check the numbers.