Mistral OCR 4
Mistral's document OCR model — bounding boxes, block classification and per-element confidence scores across 170 languages, priced per page.
Updated 2026-08-10
No free tier is documented for Mistral OCR 4.
Listed pricing: API $4 / 1,000 pages ($2 batch); Document AI $5 / 1,000 pages — check site for current rates.
Overview
Mistral OCR 4 converts PDFs, scans and images into structured output rather than a flat wall of text. Every detected element comes back with a bounding box, a block classification (heading, paragraph, table, figure and so on) and a per-element confidence score, across 170 languages. That structure is the point: downstream code can reason about layout, pull named fields out of forms and tables, and route low-confidence regions to a human instead of silently propagating a misread digit into an invoice total.
It is an API-first model, not an app. There is no consumer scan front end — you call it, or you use Mistral's higher-level Document AI product when you want the extraction-to-fields workflow handled rather than assembled. The natural buyers are teams with document volume: invoice and receipt processing, contract and form parsing, insurance and legal intake, and archive digitization feeding RAG systems where layout context changes the answer. The batch tier is what makes back-catalogue projects budgetable, since a million-page scan job is exactly the workload that does not care about latency.
The competitive picture is what has moved since this entry first went up, not the model. Dedicated OCR now sits between two kinds of pressure: the entrenched enterprise stack (Google Document AI, AWS Textract, Azure Document Intelligence) above it, and general-purpose multimodal models below, which will read a document and answer a question about it in a single call with no separate OCR hop to operate. My read: Mistral OCR 4 wins where the structure genuinely matters — high-volume, schema-driven extraction with an audit trail and a confidence threshold — and loses where a team just wants an answer out of a PDF and would rather not run two services to get it.
Is Mistral OCR 4 free?
No free tier is documented for Mistral OCR 4.
Listed pricing: API $4 / 1,000 pages ($2 batch); Document AI $5 / 1,000 pages — check site for current rates.
No figure is estimated here for what is not published. Check Mistral OCR 4's own site for the current terms.
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
Mistral OCR 4 pricing
| Plan | Price | What's included |
|---|---|---|
| OCR API | $4 per 1,000 pages | Raw structured OCR — bounding boxes, block classification, confidence scores, 170 languages. Pay-as-you-go via the Mistral API. |
| OCR API (Batch) | ~$2 per 1,000 pages | Roughly 50% off for large, non-latency-sensitive jobs such as archive digitization. |
| Document AI | $5 per 1,000 pages | Higher-level product layering extraction-to-fields workflows on top of the OCR model. |
Raw structured OCR — bounding boxes, block classification, confidence scores, 170 languages. Pay-as-you-go via the Mistral API.
Roughly 50% off for large, non-latency-sensitive jobs such as archive digitization.
Higher-level product layering extraction-to-fields workflows on top of the OCR model.
Pricing verified . That is the date the plans were last re-checked against the vendor's own pages, not today's date. Confirm on the official site before you pay.
Is Mistral OCR 4 worth it?
Worth it for Engineering teams building high-volume document pipelines — invoice, contract or claims extraction — who need positioned text and a confidence threshold they can audit, not just a transcript of the page.
There is no free tier to test it on, so the plans above are the entry cost. The recorded trade-offs are listed below, and any one of them can settle the question on its own.
The 8/10 AI Score is an editorial read of published capability, price and shipping pace. Nobody here has hands-on hours with Mistral OCR 4. How we verify.
Worth it if
The strengths recorded against this entry.
- Structured output — bounding boxes, block types and per-element confidence, not just extracted text
- Batch tier roughly halves the per-page rate, which is what makes bulk digitization budgetable
- 170-language coverage including non-Latin scripts
- Confidence scores support a real auto-accept vs. human-review threshold, useful for audited workflows
- Per-page pricing is transparent and undercuts much of the entrenched enterprise document stack
Not worth it if
Any one of these blocks your use case.
- No free tier — you pay from the first page, where several rivals offer a monthly free page quota
- General-purpose multimodal models increasingly read documents and answer questions in one call, making a separate OCR hop harder to justify for simple extraction
- API-only: no consumer or business-user front end, so someone has to build the pipeline around it
- Accuracy on messy handwriting and poor-quality scans is not covered by independent third-party benchmarking that we can point to
What sets Mistral OCR 4 apart
- Bounding boxes, block classification and per-element confidence scores in one response, rather than text plus a separate layout pass
- Batch tier at roughly half the standard per-page rate for bulk jobs
- 170-language coverage including non-Latin scripts
- Positional output is no longer rare in 2026 — the differentiator is getting it at this per-page price point rather than enterprise-stack rates
Key features
Positional extraction
Returns bounding boxes for every detected element, so text stays tied to where it sits on the page. This is what keeps multi-column layouts, tables and forms from collapsing into unusable reading-order soup.
Block classification
Labels each region by type — heading, paragraph, table, figure — giving downstream code semantic structure to parse against instead of an undifferentiated blob. Table regions in particular can be handled by dedicated logic.
Confidence scoring
Per-element confidence lets a pipeline set a threshold: auto-accept above it, queue for human review below it. For regulated workflows this is the difference between an auditable process and a black box.
Batch processing tier
Non-latency-sensitive jobs run at roughly half the standard per-page rate, which is what makes large archive digitization projects financially realistic rather than a line item someone kills.
How it compares
| Tool | Best for | Pricing | Score |
|---|---|---|---|
| Mistral OCR 4 | Engineering teams building high-volume document pipelines — invoice, contract or claims extraction — who need positioned text and a confidence threshold they can audit, not just a transcript of the page. | API $4 / 1,000 pages ($2 batch); Document AI $5 / 1,000 pages — check site for current rates | 8/10 |
| n8n | Developer-adjacent ops and engineering teams who want AI agents and internal integrations running on infrastructure they control, with server costs instead of a per-task meter. | Free self-hosted + Cloud from ~$20/mo | 8.7/10 |
| Make vs Make → | Teams building multi-step, branching automations who want to see the whole flow on a canvas rather than describe it as a list of steps. | Free tier with 1,000 credits/mo; paid from $9/mo for 10,000 credits | 8.7/10 |
| Fathom vs Fathom → | Founders, consultants and small sales teams living in Zoom or Google Meet who want unlimited call recording and CRM-synced summaries without a visible bot joining client calls. | Freemium — generous free tier; paid Premium and Team plans priced per seat (check website for current rates) | 8.5/10 |
Compare head-to-head
Related reading
Claude Riemann Hypothesis Result: 41.6% to 67.2%
Anthropic says an unreleased Claude raised the proven fraction of Riemann zeta zeros on the critical line from 41.6% to 67.2%. Here is what that means.
GPT-5.6-Cyber vs GPT-5.6 Sol: The Daybreak Split
OpenAI's GPT-5.6-Cyber is gated on authorized cybersecurity work. The Daybreak Blue and Red tiers, the 95% claim, and who each model is meant for.
Muse Glimmer and the 30B Open-Weights Ceiling
Meta released Muse Glimmer's 30B weights under Apache 2.0, landing in the same consumer-GPU size class as every other recent open model release.
Ready to try Mistral OCR 4?
Head to the official site to start with Mistral OCR 4 — pricing and plans are listed above.
Visit Mistral OCR 4