Exa 3.0 Claims First Place on 8 Search Benchmarks — All Run by Exa

Exa says its 3.0 search API outranks Perplexity, Brave and Parallel on all 8 benchmarks it tested — and every one of those benchmarks was run by Exa itself. The company reported the results on October 20, 2026, alongside two shipping changes: a larger, undisclosed-size embedding model and a new cheaper search type, Exa Lite, priced at $4 per 1,000 requests. For developers building AI agents, the story is less about the leaderboard and more about the price sheet: Exa's premium Auto tier costs $7 per 1,000 requests while all three rivals list $5.

The stakes are real. Agent frameworks now spend heavily on retrieval, and context-window tokens often cost more than the search calls themselves. Exa claims its Dynamic Highlights feature — now generally available — cuts token consumption by 95% compared with full page text while improving answer quality, at no extra charge. That claim, like the benchmark numbers, is self-reported; no rival has published matching results.

What does Exa 3.0 actually change?

Exa 3.0 upgrades retrieval quality across every search type offered through the company's hosted API. Exa runs its own crawler and maintains an independent index covering more than 2 trillion URLs and serving 100 billion pages. It is not a wrapper on Google or Bing results — a structural distinction that matters for providers competing on index quality rather than reseller margins.

Exa attributes the gains to three changes:

• A larger embedding model, trained with more compute and data than any earlier Exa model. The parameter count is not disclosed.
• Dynamic Highlights, now generally available. The model selects tokens across all results rather than page by page, allocating more tokens to better pages and dropping repeated information.
• Billions of new documents, adding coverage of products, local businesses, flights, events, sports and weather.

The company lists configurable search latency of 450ms to 1 second. There is no self-hosted option. The undisclosed embedding model size means buyers cannot independently assess the retrieval upgrade; they can only measure its output.

How do the benchmark numbers hold up?

Exa compared each provider's highest-scoring single-step endpoint — Exa 3.0 Auto, Perplexity Web, Brave LLM Context and Parallel Advanced — inside one agentic harness. The agent was GPT-6 Luna, with up to 25 search rounds and 10 results per search. Evaluation ran in October 2026.

The headline scores:

Benchmark	Exa 3.0 Auto	Perplexity	Brave	Parallel
BrowseComp (accuracy)	91.8%	91.5%	83.5%	63.1%
DSQA (F1)	88.2%	87.6%	78.6%	82.8%
WideSearch (F1)	56.1%	54.0%	54.8%	53.7%
ATLAS	42.0%	37.2%	28.2%	24.8%
WebCode	75.5%	69.9%	68.0%	64.5%
FinSearchComp	68.0%	63.4%	57.6%	66.0%
Company FindAll (mean entities)	15.4	14.2	13.3	9.7
People FindAll (mean entities)	7.6	6.8	3.8	3.8

The margins vary widely, and the widest ones deserve the most scrutiny. ATLAS and WebCode show the clearest leads, at 4.8 and 5.6 points over Perplexity respectively. But ATLAS is a benchmark Exa published itself on October 8, 2026 — roughly two weeks before this release. On BrowseComp, Exa's lead over Perplexity is 0.3 points. DSQA and WideSearch leads are each under 1.5 points.

In plain terms: the two benchmarks where Exa wins biggest are either home turf or unreplicated, and the narrowest margins may not survive a different harness or agent model.

What is Exa Lite, and what does it cost?

Exa Lite is the new budget search type, set with type: "lite" on the /search endpoint. It runs on the same index as Exa Auto; the difference is that Auto spends more compute on candidate generation, highlights and ranking. Lite supports query, result count, highlights, text, categories, domain and date filters, location and moderation. Unsupported parameters return a 400 error naming the field — so some existing pipelines will need Auto regardless.

On cost, Exa plots agentic search score against cost per task. Lite scores about 79 at roughly $0.045 per task; Auto scores about 80.5 at roughly $0.065. Exa says the pair sets a new Pareto frontier (the set of options where no competitor is simultaneously cheaper and better) among tested providers. Again, Exa drew that chart from Exa-run data.

Exa now recommends three search types — lite, auto and deep — while fast, instant and deep-reasoning remain as legacy options that still work.

The practical takeaway: Lite at $4 per 1,000 requests is the cheapest option on the table, but it is not a drop-in replacement for Auto in every pipeline.

What's in the expanded index?

Exa states it prioritized personal AI assistant use cases. The technical page lists 1.27 billion ecommerce listings with prices and availability, plus more than 211 million local businesses. Real-time news, sports, weather and market data join flights, events, transit and podcasts.

The Exa Library structures this data into searchable collections:

• 320M people pages, 33M company pages, 500M research paper pages
• 19M SEC filings, 10M patents, 9M court opinions

That structured-entity coverage is where Exa's FindAll numbers — 15.4 companies and 7.6 people versus Parallel's 9.7 and 3.8 — show the most daylight. The gap suggests genuine index depth for entity-heavy queries, though the harness remains Exa's own.

Where does the claim break down?

Every promise here comes with an asterisk. The specific weaknesses:

• Self-run evals. Exa picked the harness, the endpoints and the agent model. ATLAS is Exa's own benchmark.
• Thin margins. Several leads are under 2 points and may not survive other harnesses.
• Lite limits. Unsupported parameters trigger a 400 error, forcing Auto for some workflows.
• Higher Auto list price. $7 per 1,000 requests versus $5 for all three rivals.
• Opaque model details. Embedding model size and architecture are not disclosed.

The pattern is familiar in AI search: vendors benchmark themselves, lead with the widest deltas, and leave replication to everyone else. None of the three rivals has published results on this harness. Perplexity, Brave and Parallel each hold a $1,000-request price advantage over Exa Auto, which the benchmark table does not capture.

Exa's release is best read as a strong claim awaiting independent verification. Developers weighing a switch should run their own query sets before moving production traffic — the cheapest way to settle a self-reported leaderboard is to reproduce it.