What Each Exa Search Tier Actually Buys

Exa sells web search as a ladder whose rungs differ in kind as well as in price. Its search endpoint offered six types on 5 August 2026, and this run was charged from $7 per thousand requests for the three quickest to $15 for deep reasoning. An answer endpoint returns one cited reply for $5 per thousand, and an asynchronous agent runs at five fixed effort levels priced from $0.012 to $1.00 a run. On 5 August 2026 the same eight research questions went to fifteen of those configurations, and on these questions the higher agent efforts mainly broadened source coverage while the mean score rose by a tenth of a point.

The questions came from live research work: archival sources for two wartime economies, disclosures of how AI hardware serves large models, rental and API prices, a copyright question, and current flagship model identifiers. Each ran once through all six search types, the three deep types with an output schema (the mode in which they write an answer), the answer endpoint and all five agent efforts, which made 120 calls. Claude Fable 5 issued every call and graded every output on a five part rubric written before grading began, checking about twenty load bearing claims against primary pages, and Exa's charges for the run totalled $13.97.

The agent ladder barely climbs. Minimal effort averaged 4.80 with a median latency of about fifty seconds, low effort averaged 4.58 at just over twice the price, and medium, high and extra high each averaged 4.90. The extra money bought reach, with the extra high runs issuing as many as 120 searches where minimal issued three. On no question did the minimal run assert something a higher effort run refuted, although twenty checked claims cannot rule out an error the runs shared. The grading model's report makes the minimal run its default research call and keeps the highest efforts for a source inventory or a genuinely adversarial question. Beyond those uses, it says, the highest effort bought nothing the minimal run did not have.

Synthesis is where price and quality come apart. The deep type with a schema, priced at the same $0.012 as the minimal agent run, averaged 3.02 with a worst question of 1.4. It gave $0.20 per million tokens as the cache write price for Claude Sonnet 5 when the vendor listed $2.50 for writes and $0.20 for reads. It also stated that no serving disclosure existed for Google's Ironwood TPU although Google had published some in July, and described Gemini 1.5 Flash and Claude 3.5 Haiku as the cheap models of August 2026. Deep reasoning with the same schema cost $0.003 more and averaged 4.83, and the answer endpoint averaged 3.77 with headline errors on the two questions that needed the most care.

The clearest separation came from a price that older articles got wrong. Google's pricing page listed on demand rates for its Ironwood AI accelerator when the run checked it, while some July articles said no public rate card existed. Every agent run returned the page's figures, correct across five columns, and the deep lite and deep synthesis modes and the answer endpoint repeated the older articles, which fits agents that read pages at question time, though the retrieval path was not instrumented. Among the types that return pages, fast averaged 4.33 against 4.03 for automatic and 3.73 for deep, at the same or lower price.

Each configuration answered each of the eight questions once, and one model issued and graded every call, so a different grader or question mix could reorder close means. The rubric scores accuracy at zero for a materially wrong checked claim, but the grades gave such errors one or two points, and applying it strictly lowers four weak configurations slightly. That puts the answer endpoint just below deep page search, 3.650 against 3.675, and leaves the agent ladder and the fast search lead unchanged. Prices are what this run was charged on 5 August 2026 and type names are Exa's from that date, and the Exa search tool connected to the calling agent had no tier selector, so the runs called the API directly. The rubric, the per question grades and the error log are in the companion method and per question results.