Mistral shipped Agentic Search on August 20, 2026, replacing one-shot document retrieval with a navigable loop the model drives itself. On FinanceBench, a set of 150 questions over 368 SEC filings, Mistral Medium 3.5 went from 26.7 percent accuracy to 86 percent, a gain of 47.3 percentage points. P90 latency fell from 255 seconds to 154 seconds at the same time, which is the unusual part: the model does more steps and still finishes faster.
The idea is small and the consequence is not. Instead of embedding a query, pulling the top chunks, and hoping the answer is in them, the model gets a set of file-like operations and works the document the way a person would.
The five operations
Mistral's announcement describes a toolset that will look familiar to anyone who has used a terminal.
search finds relevant documents using existing indexes. open opens a specific document. navigate moves to a page, section, or region. read retrieves content at a given location. grep finds patterns inside an open document.
That mapping is deliberate. Classic retrieval-augmented generation treats a corpus as a bag of chunks and asks an embedding model to guess which chunks matter before the reasoning model has seen anything. Agentic Search inverts the order: the model reads a little, decides what it still needs, and goes and gets it. The Agentic Search guide covers the loop in detail, and the Search Toolkit docs cover how to wire it into an application.
What the benchmarks show
Mistral published results on two document sets, and the pattern holds across both. FinanceBench is a public benchmark of questions over real SEC filings, which makes it a reasonable proxy for the kind of long, structured corpus a studio or agency actually keeps.
| Benchmark | Corpus | Model | Gain |
|---|---|---|---|
| FinanceBench | 368 SEC filings, 150 questions | Mistral Medium 3.5 | 26.7% to 86% (+47.3pp) |
| FinanceBench | 368 SEC filings, 150 questions | GLM-5.2 | +52.6pp |
| OfficeQA Pro | 696 Treasury Bulletins, 133 questions | GLM-5.2 | 6.3% to 51.9% (+45.6pp) |
| OfficeQA Pro | 696 Treasury Bulletins, 133 questions | Mistral Medium 3.5 | +27.1pp |
Two details deserve attention. First, the largest FinanceBench gain came from GLM-5.2, a model Mistral does not make. The loop is not a Mistral-model trick; it improves whatever model is driving it. Second, token usage on OfficeQA Pro dropped by up to 33.7 percent. Reading selectively costs fewer tokens than stuffing a context window with chunks that turn out to be irrelevant.
The latency result follows from the same logic. A P90 of 255 seconds falling to 154 seconds while the model takes more turns means the old path was spending its time processing bulk context, not thinking.
Why this matters if you build with documents
Most creator and studio workflows that touch a document corpus are quietly RAG pipelines: a brand guideline library, a back catalogue of scripts, a client's brief archive, a folder of contracts. The failure mode is always the same. The retriever misses, the model answers confidently from the wrong chunk, and you only notice when a client does.
An agentic loop changes the failure mode rather than eliminating it. When the model can open, navigate, and grep, a miss on the first search is recoverable, because the next step is another search rather than a hallucinated answer. That is why the accuracy jumps are so large on long, structured documents like filings and bulletins, where the answer exists in exactly one place and chunk-based retrieval either finds it or does not.
It also lowers the amount of pipeline you own. Chunking strategy, embedding model choice, and re-ranking are three tuning surfaces that mostly disappear when the model navigates the source directly. Anyone who has spent a weekend tuning chunk overlap will recognise the trade.
What it costs you in exchange
An iterative loop is not free, and the honest version of this story names the trade. The model makes multiple tool calls per question instead of one retrieval, which means more round trips and more places for a run to stall. On a corpus small enough to fit comfortably in context, the loop is overhead you do not need.
It also moves the debugging surface. With classic RAG, a wrong answer is usually a retrieval problem you can inspect by looking at which chunks came back. With an agentic loop, a wrong answer is a sequence of decisions, and understanding it means reading a trace rather than a ranked list. Teams that have instrumented their retrieval carefully will find the first few weeks less legible than what they replaced.
The published numbers argue the trade is worth it on long structured documents, where token usage fell by up to 33.7 percent and latency dropped despite the extra turns. They do not argue it for every corpus, and Mistral does not claim they do.
Trying it without rebuilding anything
1. Start from the starter app. Mistral published a search starter app on GitHub. Clone it and point it at a folder of your own documents before you touch your production pipeline.
2. Use your worst queries as the test set. Do not benchmark on questions your current RAG setup already answers. Pull the queries where it returned the wrong section, which is the exact failure the loop is designed to fix.
3. Watch token usage, not just accuracy. The OfficeQA Pro result showed up to 33.7 percent fewer tokens. If your corpus is large, that number may matter more to your bill than the accuracy delta, particularly alongside the falling frontier prices we covered this week.
4. Check it against a structured document. The published gains come from filings and bulletins, which have headings, sections, and consistent structure. A corpus of unstructured transcripts is a different shape, and you should verify rather than assume the same lift.
Where it runs
Agentic Search is available through the Mistral Search Toolkit and Libraries, and is built into Studio and Vibe. Mistral describes it as generally available rather than a preview. That places it alongside the broader move toward tool-shaped retrieval that also shows up in MCP-driven local agent tooling and in Mistral's own recent open releases such as Shieldstral.
Frequently asked questions
What is Agentic Search?
It is Mistral's retrieval system that gives a model five file-like operations, search, open, navigate, read, and grep, so it can iteratively find and verify information across documents instead of relying on a single retrieval step.
How is it different from normal RAG?
Standard retrieval-augmented generation embeds a query, pulls top-matching chunks, and passes them to the model in one shot. Agentic Search lets the model read a little, decide what it still needs, and fetch again, so an initial miss is recoverable rather than fatal.
How much does accuracy improve?
On FinanceBench, Mistral Medium 3.5 rose from 26.7 percent to 86 percent, a 47.3 point gain, and GLM-5.2 gained 52.6 points. On OfficeQA Pro, GLM-5.2 went from 6.3 percent to 51.9 percent.
Does it only work with Mistral models?
No. The largest published FinanceBench gain came from GLM-5.2, which Mistral does not build. The loop improves whatever model drives it.
Is it slower than one-shot retrieval?
Not in the published results. P90 latency on FinanceBench fell from 255 seconds to 154 seconds, and token usage on OfficeQA Pro dropped by up to 33.7 percent, because selective reading avoids processing bulk irrelevant context.
How do I try it?
Use the Mistral Search Toolkit and Libraries, which are built into Studio and Vibe, or clone the search starter app on GitHub and point it at your own document folder.
