Summary: Exa’s April 22 Highlights update returns query-relevant excerpts instead of full pages, and its own unverified evaluations claim 500-character excerpts match the first 8,000 characters on SimpleQA with 16x fewer tokens. The company also claims 4k-character highlights outperform 32k-character full text. Pricing is not disclosed, and results still need testing on a team’s own queries.
Anyone who has pushed an agent through several search rounds has run the same math: each round shoves an entire page into context, the answer takes up a handful of sentences, and everything else is navigation, footers, and boilerplate repeated yet again. As rounds stack up, token spend and latency climb right along with them. Exa names this problem context bloat and argues it is the central tax agentic search pays. Its Highlights update, announced on April 22, swaps “read the whole page” for “read only the passages that matter to this query.” And the claim concerns reading, not ranking: search still hands back the same pages — what changes is how much of each page your agent has to swallow.
Three figures first — and how to take them
Three sets of figures anchor the post. Before reading them, pin down what they are: every one is an evaluation Exa ran on itself, verified by nobody outside the company, and “some evals” never names a benchmark. Take them as vendor claims rather than results anyone has reproduced.
| Claim | Content |
|---|---|
| Matching accuracy | On SimpleQA, Highlights spanning 500 characters reach the accuracy of the page’s first 8,000 characters |
| Fewer tokens | The same comparison spends 16x fewer tokens; Exa separately claims higher-quality results on some evals for ~94% fewer tokens |
| Roomier budgets | Highlights of 4k characters outperform 32k characters of full text |
The SimpleQA pairing is the most concrete of the three: excerpts of 500 characters reach the same accuracy as the first 8,000 characters of raw page content while spending 16x fewer tokens. The “~94%” number stays at marketing-summary altitude — no baseline detail exists to check it against. Row three (4k over 32k) tells you the edge survives once you loosen the budget. What builders should carry away is the direction, not the digits: even under aggressive compression, excerpt quality held. That is the real message of the announcement.
How it works: computed per request, never cached, done inside 100 milliseconds
Highlights is an in-house method Exa built to pull the most relevant excerpts out of web pages, and three design choices define it. One: it recomputes on every request and keeps no cache — the same page surfaces different passages for different queries, and Exa picked relevance over reuse. Two: a single run finishes inside 100 milliseconds, quick enough to live at every step of an agent loop without giving up responsiveness. Three: according to Exa, the model emerged from multiple rounds of research that canvassed a wide range of architectures and training recipes before the current approach, and the company reports marked recent gains on technical coding docs (API references, SDK docs) and long-context documents (specs, research papers).
The widest gap opens on long documents. On long pages under small context budgets, Highlights score 60% at 500 characters where competing methods manage 6%: handed documents drowning in boilerplate, the baselines barely ever find the right passage, while Highlights push the answer passage straight to the front. That is also why multi-turn search gains the most — savings land on every round, and rounds multiply them.
More than a public API: the retrieval base under Exa’s own agent products
The public Highlights API is only part of the story. Exa’s own agentic endpoints — /answer, Deep, and Websets — run on Highlights as their retrieval substrate: at every iteration of those agent loops, what gets read is highlights, not raw page content. Exa goes further and claims its agentic products are Pareto dominant against competing systems on latency, cost, and quality. Self-reported once more — but the logic holds together: build token efficiency into the retrieval layer and every downstream product inherits the speed and cost wins, with no compression bolted on afterward.
Limits and trade-offs to settle before adopting
Four things deserve straight answers before you adopt. One: every figure comes from Exa itself with no third-party verification, and the “some evals” stay anonymous — prove the conclusion on your own query set first. Two: pricing and billing go unmentioned in the post; the cost model is yours to pin down. Three: ecosystem lock-in. Highlights ships as part of Exa’s API, so teams on other search engines or self-built retrieval gain nothing directly — the realistic moves are switching providers or training a comparable model yourself. Four: excerpting throws context away by construction — recall on an eval is not recall on your domain, and the agent cannot tell which details the excerpt dropped.
What builders should do
Three calls you can act on today. First, treat excerpts as the interface your agent actually reads through, not as a finishing flourish: delivering only the relevant passages at retrieval time costs an order of magnitude less than taking full pages and filtering afterward, and it preserves the context window for genuine reasoning. Second, wire up long-document workloads first — API references, SDK docs, and papers sit exactly where the 60%-versus-6% gap appears and where the token bill stings worst. Third, if you already live inside Exa’s ecosystem, Highlights is close to a free upgrade; if you do not, the post still leaves you a training target: teach the retrieval layer to extract passages for a specific query rather than hand back a fixed-length page head.