Mistral released OCR 4.1 today, an update to its document OCR model, and the announcement is unusual for what it doesn't contain: no benchmark table, no OlmOCRBench score, no head-to-head comparison against a rival. Just six named, specific fixes — the kind that only shows up when real customers feed real, messy documents through a production pipeline rather than a benchmark set.
The six fixes are all long-tail correctness, not capability
Bounding boxes now align to the content they mark, with no nested images and no drift — a defect that doesn't show up in an accuracy percentage but absolutely shows up when a downstream system tries to crop an image from a coordinate the model reported. Reference pages keep their structure, returning one box per citation instead of collapsing a numbered list into a single undifferentiated block — exactly the failure mode that makes a legal brief or academic paper's bibliography useless to parse programmatically. The model misses fewer blocks on complex pages. Double quotes stay as written, no longer silently converted to single quotes — a correctness bug that would corrupt any document where the distinction is semantically load-bearing (code samples, quoted legal language). A wider range of checkbox styles is recognized. Right-to-left tables parse more accurately — the one fix that reads as a real capability gap closing, since RTL table layout is a structurally different parsing problem, not a fine-tuning nudge.
Every one of those is the profile of a bug tracker item, not a benchmark gap. Compare this to the benchmark-table-first releases this month has been full of — Grok 4.6, DeepSeek-V4-Pro, Qwen3.8-Max all led with a comparison chart against named rivals. OCR 4.1 leads with a list of things that were wrong for actual paying customers. That's not a lesser kind of release; document OCR is a domain where "the citation list didn't collapse into one box" matters more to whether a product ships than a percentage point on an aggregate score does.
The quiet part: the alias just moved under you
The mistral-ocr-latest alias now points to OCR 4.1, which means any integration built against latest rather than a pinned version number gets these six fixes with zero code changes — and, less discussed, would have gotten any regressions the same way. That's a real operational choice worth naming: pointing a latest alias at a new model version is a silent behavior change for every downstream system that trusted the alias to be stable within a release family. It's convenient until the day one of those six fixes turns out to break a workflow that depended on the old, buggy behavior — the RTL table parsing change, for instance, could plausibly alter output for any document that was being parsed "wrong but consistently" before today.
The pricing context, for anyone catching up
Worth restating since it's easy to lose: Mistral's OCR API pricing doubled on June 23, when OCR 4 itself launched — $2 to $4 per 1,000 pages for OCR, with Document AI at $5. OCR 3 stays available at the old $2 rate for anyone who doesn't want to move. That's not new today, but it means OCR 4.1's six fixes arrive on top of a price increase that's already six weeks old, not alongside a fresh one — the improvement and the price hike are decoupled in time, which is worth knowing before assuming this release is what justified the June price change.
What to expect next
- Watch whether "no benchmark table" becomes Mistral's pattern for incremental OCR updates, versus a one-off for a release this narrow. A model category defined by document-parsing correctness may simply not benchmark well as a leaderboard number, which would make this the honest choice rather than an omission.
- Watch for regression reports from the alias move. Anyone whose pipeline quietly depended on OCR 4's specific quirks — including the ones just fixed — is the population most likely to notice this release went out.
References: Mistral's OCR 4.1 announcement (direct text supplied; no independently reachable URL as of this post — mistral.ai is unreachable from this environment and the specific 4.1 page did not surface in search) · Mistral — OCR 4 launch, for pricing/background · related coverage: Mistral Stops Selling Compute and Starts Selling Trust · DeepSeek-V4-Pro Closes the Gap and Opens a New One · Frontier Arcade: trends & predictions