{"slug": "exa-claims-the-top-spot-on-eight-ai-search-benchmarks-with-exa-3-0", "title": "Exa claims the top spot on eight AI search benchmarks with Exa 3.0", "summary": "Exa released Exa 3.0 on October 19, claiming first place on all eight AI search benchmarks it tested, including 91.8% on BrowseComp and 88.2% on DSQA, ahead of Perplexity, Brave Search and Parallel. Exa said the gains came from training a larger embedding model on more compute and more data than any prior version, and it published the results against a benchmark suite it built itself. The company, founded about five years ago by Will Bryk and Jeff Wang, closed a $250 million Series C in May 2026 led by Andreessen Horowitz at a $2.2 billion valuation, and its API serves more than 5,000 companies and over 400,000 developers.", "body_md": "*91.8% on BrowseComp. 88.2% on DSQA. Exa says its newest engine tops Perplexity, Brave and Parallel across the board, and it built the benchmarks too.*\n\nExa released Exa 3.0 on October 19. The company posted the results on X as \"our largest ever improvement to search quality.\" It says the new version ranks first on all eight benchmarks it tested, spanning public retrieval evals and vertical tasks like coding search and go-to-market intelligence. On BrowseComp, a multi-hop retrieval test, Exa 3.0 scored 91.8%. On DSQA it hit 88.2%. The rest of the sheet: 56.1% on WideSearch, 42.0% on ATLAS, 75.5% on WebCode, and 68.0% on FinSearchComp, plus 15.4 on Company FindAll and 7.6 on People FindAll.\n\nThose numbers mean little without a point of comparison, which is exactly why Exa published them against Perplexity, Brave Search and Parallel, the three rivals it names directly in its own benchmark suite. Exa says the gains came from training a larger embedding model on more compute and more data than any version it has shipped before. That's a meaningful admission on its own: the company isn't claiming a clever new technique, it's claiming it out-spent and out-scaled the field on the thing that actually moves retrieval accuracy.\n\nExa was founded about five years ago by Will Bryk and Jeff Wang on a simple bet. General search engines like Google were built for people clicking blue links, not for software that needs to fetch and reason over facts with no human in the loop. That bet has paid off. The company closed a $250 million Series C in May 2026 led by Andreessen Horowitz, with QuantumLight also participating, at a $2.2 billion valuation - roughly 3.1 times the $700 million mark it set at its Series B just eight months earlier, in September 2025, according to Exa's own funding announcement. Growth like that in under a year tells you where investors think AI infrastructure spend is actually heading: not just toward bigger models, but toward the plumbing those models need to stay grounded in current facts.\n\nExa's API now serves more than 5,000 companies and over 400,000 developers, including HubSpot, monday.com, Cursor, OpenRouter and Cognition. Those are not casual integrations. Coding assistants like Cursor and workflow platforms like HubSpot need search results accurate enough for a model to act on without a human double-checking the citation. That's a much higher bar than ranking a page for a person who can scroll past the bad results.\n\n[Exa says its new search engine for AI agents beats every rival on eight benchmarks](https://startupfortune.com/exa-says-its-new-search-engine-for-ai-agents-beats-every-rival-on-eight-benchmarks/)\n\nExa Labs, backed by a16z at a $2.2 billion valuation, unveiled a cheaper Exa Lite tier at $4 per 1,000 requests alongside the October 18 release, part of a broader push to undercut Perplexity on agent search costs. - [AI search engine beats Google for agent queries](https://startupfortune.com/exa-says-its-new-search-engine-for-ai-agents-beats-every-rival-on-eight-benchmarks/) - [Exa 3.0 embedding model outperforms all rival benchmarks](https://startupfortune.com/exa-says-its-new-search-engine-for-ai-agents-beats-every-rival-on-eight-benchmarks/)\n\nThat's also where the competitive picture gets interesting. Perplexity, the best-known name in AI-native search, has raised $1.72 billion and carries a reported $20.4 billion valuation, dwarfing Exa's. But Perplexity is built primarily as a consumer answer engine, a chat interface competing with Google itself. Parallel is a closer match to what Exa is doing, and it was founded by former Twitter chief executive Parag Agrawal. It raised a $100 million Series B in April 2026 at a $2 billion valuation and pitches itself explicitly as web infrastructure for agents, not people. Exa landing ahead of Parallel on its own benchmark sheet is the kind of claim that invites skepticism precisely because the two companies are chasing the same customer.\n\nAlongside the model upgrade, Exa made Dynamic Highlights generally available. The feature pulls the relevant tokens out of a set of pages instead of dumping full page text at a model. Exa says it cuts token usage by 95% compared with returning raw page content, while improving answer quality. For any company running search calls at scale inside an agent loop, that's not a nice-to-have. Token costs compound fast when an agent is firing off dozens of searches to complete one task, so a 95% cut in what gets passed to the model is a direct line to lower inference bills.\n\nNone of this settles the argument that benchmark leaderboards in AI search are becoming as gameable as the SEO rankings they're meant to replace. Companies run their own evals, pick their own comparison set, and publish the results that favor them. Exa at least open-sourced its query sets and grading code on GitHub for outside verification, which is more transparency than most vendors offer. Whether Exa 3.0 actually holds the top spot once independent evaluators run their own tests is the question that will decide if developers standardize on it, or just note the claim and keep testing alternatives themselves.\n\n**Also read:** [Satya Nadella says every AI model should be treated as an insider threat](https://startupfortune.com/satya-nadella-says-every-ai-model-should-be-treated-as-an-insider-threat/) • [Stripe Will Buy OpenRouter, the AI Traffic Cop, for $7.5 Billion](https://startupfortune.com/stripe-will-buy-openrouter-the-ai-traffic-cop-for-75-billion/) • [Nscale's SEC filing confirms a $44.6 billion Anthropic compute deal](https://startupfortune.com/nscales-sec-filing-confirms-a-446-billion-anthropic-compute-deal/)\n\n*This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](https://startupfortune.com/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/exa-claims-the-top-spot-on-eight-ai-search-benchmarks-with-exa-3-0", "canonical_source": "https://startupfortune.com/exa-claims-the-top-spot-on-eight-ai-search-benchmarks-with-exa-30/", "published_at": "2026-10-20 00:10:06+00:00", "updated_at": "2026-10-20 00:17:25.779658+00:00", "lang": "en", "topics": ["ai-search", "artificial-intelligence", "ai-infrastructure", "ai-agents"], "entities": ["Exa", "Exa 3.0", "Perplexity", "Brave Search", "Parallel", "Will Bryk", "Jeff Wang", "Andreessen Horowitz"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/exa-claims-the-top-spot-on-eight-ai-search-benchmarks-with-exa-3-0", "markdown": "https://wpnews.pro/news/exa-claims-the-top-spot-on-eight-ai-search-benchmarks-with-exa-3-0.md", "text": "https://wpnews.pro/news/exa-claims-the-top-spot-on-eight-ai-search-benchmarks-with-exa-3-0.txt", "jsonld": "https://wpnews.pro/news/exa-claims-the-top-spot-on-eight-ai-search-benchmarks-with-exa-3-0.jsonld"}}