Critic catch: Agentic Search triples document-QA accuracy while cutting tokens

What happened

Mistral shipped Agentic Search, a retrieval layer that gives a model five tool-like actions — search, open, navigate, read, grep — so it works through a document set iteratively instead of retrieving once and answering. The benchmark numbers are the story: on FinanceBench (368 SEC filings, 150 questions) Mistral Medium 3.5's accuracy went from 26.7% to 86%, roughly a tripling, while token usage fell 23.9% and P90 latency fell from 255s to 154s. Better and cheaper and faster is rare enough to be worth flagging as a capability result rather than a feature announcement — the usual shape is a quality gain bought with tokens. (Mistral)

Sources 1

Mistral mistral.ai unrated

Credibility ratings combine two public datasets: Lin et al. 2023, PNAS Nexus 2(9):pgad286 (CC BY 4.0) for the domain-quality score, and Wikipedia's perennial-sources list (CC BY-SA 4.0) for the reliability tier. A rating describes an outlet's general record, never this particular article. Unrated means absent from those datasets — not a judgement.