Exa 3.0 tops the Artificial Analysis Search Index, but DeepSeek and Octen undercut it on price

Exa announced Exa 3.0, which it calls its largest improvement to search quality. It says Exa is now #1 on every benchmark it tested, from public evals to verticals such as coding and GTM intelligence. Replies on X also cite a larger embedding model for retrieval and "Dynamic Highlights" that cut tokens by about 95% versus full text. Those two details come from replies, not from Exa's own post.

Exa's launch table: Exa 3.0 against Perplexity, Brave and Parallel on eight benchmarks. @ExaAILabs

#What the independent data shows

The Artificial Analysis Search Index covers 27 products from 14 providers. It averages DeepSearchQA, BrowseComp and AA-Omniscience, all run with GPT-5.6 Luna (medium) as the answer model. Its latest data is dated Oct 19, 2026.

• Exa Search (auto, dynamic highlights) and Perplexity Search (medium) both score 80. Exa leads on DeepSearchQA (86 vs 81). Perplexity leads on BrowseComp (87 vs 83).
• Exa Search (lite, dynamic highlights) scores 79. Its search cost is $31.20 per 1,000 tasks, against $53.17 for the top Exa configuration.
• Octen Search (highlights) scores 77 at $9.07 search cost per 1,000 tasks. It is the fastest, at 14.9s per task.
• DeepSeek Search (summary) scores 69, with search cost of $8.47 per 1,000 tasks. Its model cost per 1,000 tasks is $23.24, the highest of any provider in the table.
• The model-only baseline scores 33, so search adds roughly 36 to 47 points across providers.

Search Index against total cost per task. Exa lite and Octen sit on the Pareto line; the default view shows 12 of 27 providers. Artificial Analysis

In replies, an Artificial Analysis account said DeepSeek is about 80% cheaper per query than Exa Lite. It said the gap shrinks to about 30% per task, because weaker results led to about 1.5x more searches. It also said DeepSeek scores about 10 points lower on quality.

#DeepSeek's new search API

DeepSeek's API docs describe a Search/Fetch endpoint at /v1/tools/search. It returns ranked results with snippets, longer model-oriented summaries, optional page content, dates and an authority score. Several X replies put its price at $0.75 per 1,000 queries. That figure comes from replies, not from the docs page supplied here.

#Caveats

• Exa's claim covers the benchmarks Exa chose. Many replies pointed out that its charts left out some competitors, such as Octen, DeepSeek and Tavily.
• On the Artificial Analysis index, Exa's lead over Perplexity is a tie at the headline score.
• Time per task for the top Exa configuration is 45.3s, against 26.5s for Perplexity medium.

#Why it matters

Search quality is a main bottleneck for agents. The market is splitting into a high-quality tier (Exa, Perplexity) and a cheaper tier (Octen, DeepSeek). Independent data shows that per-query price is not the whole cost, because worse results mean more searches and more model tokens. Teams choosing a search API should compare cost per completed task, not cost per query.