Mistral's new Agentic Search feature lifts AI accuracy on SEC filings from 26.7% to 86% by letting models navigate documents instead of just retrieving chunks.
Picture a financial analyst three hours into a 10-K, hunting for one number: the company's effective tax rate in Q3, buried in a footnote on page 94. That afternoon of scrolling, cross-referencing, and double-checking is exactly the workflow Mistral says its new Agentic Search can take over, and take over well. On FinanceBench, a benchmark built from 150 questions asked against 368 real SEC filings averaging 147 pages each, the feature moved AI accuracy from 26.7% to 86%, a threefold jump, according to Mistral's announcement published August 20.
Here's the mechanical difference. Most AI search today is one-shot retrieval-augmented generation: the system grabs a handful of text chunks that look relevant and asks the model to answer from those alone. If the real answer is in a table the retriever missed, or requires flipping to a different document to verify, the model is stuck with what it was handed. Agentic Search instead gives the model five tools that behave like a file system: search, open, navigate, read, and grep. The model can decide a first search wasn't good enough, open a specific document, jump to a specific page or table, and read it directly, the way an actual analyst would, before it answers.
The numbers hold up on a second, harder benchmark too. On OfficeQA Pro, a set of numeric questions against scanned U.S. Treasury Bulletins spanning nearly 89,000 pages, accuracy climbed from 6.3% to 51.9%. And the gains aren't just accuracy for its own sake. Mistral reports the full search-and-navigate loop cut p90 latency on FinanceBench from 255 seconds to 154 seconds, a 39.6% reduction, while using up to a third fewer tokens than a search-only agent loop. The company tested this against its own Mistral Medium 3.5 model and against a third-party model, Z.ai's GLM-5.2, and got the same pattern on both, which is the detail that matters most for buyers: the retrieval layer isn't locked to one model family.
Agentic Search ships now through Mistral Search Toolkit, for teams building their own retrieval pipelines, and through Libraries, which is built into both Mistral Studio and Vibe so nobody has to stand up that pipeline themselves.
For a business leader, the interesting part isn't the benchmark score, it's who gets to use this without hiring a search engineer. According to Mistral's pricing page, Libraries indexing runs $1 per million tokens with calls at a penny apiece, and document OCR runs $3 per 1,000 pages. That's the infrastructure cost of turning a filing cabinet's worth of contracts or filings into something an AI agent can actually search and verify against, not just skim. On the product side, Vibe's Pro tier is $14.99 a month, so a compliance team or a small legal shop can point an agent at their own document set without waiting on an engineering roadmap. That's the shift: document intelligence that used to require a RAG pipeline someone had to design and tune now shows up as a checkbox inside a chat product.
The honest caveat is that 86% is not 100%. On financial filings, that's still roughly one answer in seven landing wrong if nobody checks it, and if the number is going into an audit workpaper or a regulatory filing, an unverified 86% is not a substitute for a human sign-off, it's a first draft with a better hit rate than before. Mistral says as much itself: these results are "floors, not ceilings," meaning they're the out-of-the-box performance before anyone tunes chunking or ranking for their specific document set, which also means the headline numbers assume you haven't done any of that tuning yet. And Mistral's own guidance is that simple, short documents where the answer sits in the first retrieved chunk don't need any of this. If your search problem is "find the right paragraph in a five-page policy doc," a one-shot RAG setup is still the right tool, and paying for navigation you don't need is just cost with no upside.
The line between "AI can chat with your documents" and "AI can be trusted to find and verify the one number that goes in the filing" just moved, and it moved because of five unglamorous verbs: search, open, navigate, read, grep.