If you have run an agent through multiple rounds of search, you know the arithmetic: every round stuffs a full page into context, the answer occupies a few sentences, and the rest is navigation, footers, and repeated boilerplate. Token costs and latency scale up with each round. Exa calls this context bloat and argues it is the core tax on agentic search. The Highlights update announced on April 22 replaces “read the whole page” with “read only the passages that matter to this query.” The claim is about reading, not ranking: search still returns the same pages, and what changes is how much of each page your agent has to ingest.
Three numbers, and how to read them
Exa’s post leads with three sets of figures. Be clear about their nature first: all of them are Exa-reported evaluation results with no third-party verification, and “some evals” names no specific benchmark. Treat them as vendor data, not reproducible conclusions.
| Claim | Content |
|---|---|
| Accuracy equivalence | On SimpleQA, 500 characters of Highlights match the accuracy of the page’s first 8,000 characters |
| Token savings | That comparison uses 16x fewer tokens; Exa also reports higher-quality results for ~94% fewer tokens on some evals |
| Larger budgets | 4k characters of Highlights beat 32k characters of full text |
The most concrete pair is the SimpleQA one: 500 characters of excerpts match the accuracy of the first 8,000 characters of raw page content, at 16x fewer tokens. The “~94%” claim sits at marketing-summary level — there is not enough baseline detail to compare it against anything. The third row (4k beats 32k) shows the advantage holds when you loosen the budget. For builders, the direction matters more than the digits: excerpt quality does not collapse under aggressive compression, and that is the core message of the announcement.
Mechanism: per-request, uncached, under 100 milliseconds
Highlights is Exa’s in-house method for extracting the most relevant excerpts from web pages, and three design decisions stand out. First, it runs for every request and is not cached: the same page yields different passages under different queries, and Exa chooses relevance over reuse. Second, each run completes in under 100 milliseconds — fast enough to sit inside every step of an agent loop without sacrificing responsiveness. Third, Exa says the model took multiple rounds of research, exploring a wide range of architectures and training recipes before the current approach, and reports significant recent improvements on technical coding docs (API references, SDK docs) and long-context documents (specs, research papers).
Long documents are where the gap is widest. On long pages at small context budgets, the comparison against competing methods is 60% versus 6% at 500 characters: baseline methods barely locate the correct passage inside boilerplate-heavy documents, while Highlights put the answer passage up front. This is also why multi-turn search benefits disproportionately — every round saves, and the savings multiply with rounds.
Beyond the public API: the substrate behind Exa’s agentic products
Highlights is not only the public Highlights API. Exa’s own agentic endpoints — /answer, Deep, and Websets — use Highlights as their retrieval substrate: every iteration of those agent loops reads highlights rather than raw page content. Exa further claims its agentic products are Pareto dominant on latency, cost, and quality against competing systems. That too is self-reported, but the structure is plausible: once the retrieval layer is token-efficient, downstream products inherit both speed and cost advantages without after-the-fact compression.
Trade-offs and limits
Be honest about four things before adopting. First, every number is Exa-reported with no third-party verification, and “some evals” is unnamed — validate on your own query set before trusting the conclusion. Second, the post discloses nothing about pricing or billing; confirm the cost model yourself. Third, ecosystem lock-in: Highlights is part of Exa’s API. Teams on other search engines or self-built retrieval get no direct benefit; the options are switching providers or training a comparable model. Fourth, any excerpting discards context by construction — recall on an eval is not recall on your domain, and the agent has no way to know what the excerpt left out.
Advice for builders
Three judgments you can act on. First, treat excerpts as the agent’s reading interface, not decoration: feeding only the relevant passages at retrieval time is an order of magnitude cheaper than filtering after receiving full pages, and it leaves the context window for actual reasoning. Second, prioritize long-document workloads — API references, SDK docs, and papers are where the 60%-versus-6% gap lives and where the token bill hurts most. Third, if you are already inside Exa’s ecosystem, Highlights is nearly a free upgrade; if you are not, the post still hands you a training target: your retrieval layer should learn to extract passages for a specific query, not return a fixed-length page head.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.