Contents
If your business needs an EU-resident model provider, Mistral is the obvious answer and the pricing no longer makes that a sacrifice. Mistral Large 3 lists at $0.50 in and $1.50 out per 1M tokens as of August 2026, Small 4 at $0.15 and $0.60, and Codestral at $0.30 and $0.90, all with a 50 percent batch discount and a 90 percent caching discount on top. It is EU-headquartered and outside US CLOUD Act reach, and the flagship, small and Ministral tiers are Apache 2.0 licensed so you can self-host. Two honest caveats. Context windows are still not published anywhere we could find them. And Mistral's data processing addendum says it may use feedback and aggregated data to train models unless you opt out, which is a weaker default than the opt-in framing often attributed to it.
Best for
EU and UK businesses with data-residency requirements, cost-sensitive work that benefits from a 90 percent caching discount, and anyone who wants Apache 2.0 weights they can move on-premise later.
Not for
Teams who need thoroughly documented model specs, since context windows are not published, or anyone who assumed the default terms were opt-in for training.
Mistral AI API pricing per 1M tokens
One documentation trap to flag first. Mistral's main pricing landing page was still quoting the superseded figure of $2 in and $6 out per 1M for Mistral Large when we checked in August 2026, while the dedicated API pricing page listed Mistral Large 3 at $0.50 and $1.50. Use the API pricing page, not the landing page, and be sceptical of any third-party comparison built from the wrong one. The two discounts are unusually good: batch processing at 50 percent off and caching at 90 percent off, both documented across the Medium, Large and Small tiers. Context windows remain unpublished on the pricing pages and on the models overview we read, which is a genuine gap. Prices below are as of August 2026.
| Model | Context | Input / 1M | Output / 1M |
|---|---|---|---|
| Mistral Large 3 (flagship)API name mistral-large-2512; Apache 2.0 open weights; batch -50 percent, caching -90 percent | Not published | $0.50 | $1.50 |
| Mistral Medium 3.5 (agentic and coding)API name mistral-medium-3504; modified MIT licence; batch -50 percent, caching -90 percent | Not published | $1.50 | $7.50 |
| Mistral Small 4API name mistral-small-2603; Apache 2.0; batch -50 percent, caching -90 percent | Not published | $0.15 | $0.60 |
| Ministral 3 - 3BApache 2.0; flat rate both directions, sized for edge deployment | Not published | $0.10 | $0.10 |
| Ministral 3 - 8BApache 2.0; flat rate both directions | Not published | $0.15 | $0.15 |
| Ministral 3 - 14BApache 2.0; flat rate both directions | Not published | $0.20 | $0.20 |
| CodestralAPI name codestral-2508; Premier licence, closed weights. Now published, unlike earlier in 2026 | Not published | $0.30 | $0.90 |
| Magistral MediumReasoning tier | Not published | $2.00 | $5.00 |
| Mistral OCR 4Document AI variant listed at $5 per 1,000 pages | n/a | $4 per 1,000 pages | n/a |
| Mistral EmbedEmbeddings, priced per 1M input tokens | n/a | $0.10 | n/a |
List prices as of August 2026, taken from the provider's published pricing page on August 5, 2026. Model prices move fast - check the source links before you budget.
When EU residency is the whole reason to choose Mistral
Most small businesses do not have a hard data-residency requirement, and for them Mistral is simply one competent provider among several. But when the requirement exists, it usually exists absolutely: a public-sector client contract, a professional-services engagement with a confidentiality clause, an insurer's condition, or a sector regulator's expectation. In those cases the question is not which model is best, it is which providers you are permitted to use at all.
Mistral's structural answer is that it is a French company operating in the EU, so the concern that a US-headquartered provider could be compelled to produce data under the CLOUD Act does not arise in the same form. You can get EU regions from the US hyperscalers too, and for many organisations that is sufficient, but a European legal entity is a cleaner story when a client asks the question directly.
The contractual package is the second piece and it is genuinely good for a vendor of this size. The DPA puts Mistral in the processor role with you as controller, incorporates standard contractual clauses for any transfer outside the EEA, commits to GDPR Article 32 security measures and prompt breach notification, publishes a subprocessor list on the Trust Center, and grants one annual onsite audit with 90 days notice. Onsite audit rights are rare at this price point and they are exactly the sort of clause that gets a vendor through a public-sector procurement.
One thing we have to correct, because we previously described it the wrong way round. The DPA states that Mistral may use feedback and aggregated data to train models unless the customer opts out. That is an opt-out default. It is narrower than the training-on-everything defaults that DeepSeek and Moonshot document, but it is not the opt-in position that Mistral is often credited with, including in earlier versions of this page. If training on your content is a hard no, exercise the opt-out explicitly and keep the record.
Retention under the DPA is framed around the agreement rather than a rolling window: personal data is retained for the duration of the agreement and becomes inaccessible within 30 days of termination. Zero data retention appears in Mistral's Trust Center material as an available option, but we could not read a page setting out its precise scope or how it is enabled, so confirm that with Mistral rather than assuming it works the way a competitor's does.
The price cut that changed the calculation
Mistral used to be a provider you chose despite the price. The old flagship listed at $2 in and $6 out per 1M tokens, which was frontier pricing for a model that was not frontier capability. Mistral Large 3 now lists at $0.50 and $1.50 on Mistral's API pricing page, Small 4 at $0.15 and $0.60, and Codestral at $0.30 and $0.90. That repositions Mistral as a genuinely price-competitive option rather than a compliance-driven compromise.
Put it on the two workloads we price every provider against. Workload A is an inbox agent, 1,500 runs a month at about 3,000 input tokens each of which 2,400 are a stable prefix, plus 400 output. Workload B is document extraction, 4,000 documents a month at 6,000 input and 300 output. On Mistral Small 4 that is about $1.04 and $4.32 a month, falling to roughly $0.55 and $2.16 once you apply the 90 percent caching discount to the prefix and the 50 percent batch discount to the bulk job. On Mistral Large 3 they are about $3.15 and $13.80. On Medium 3.5 about $11.25 and $45.00. On Ministral 3 8B, at a flat $0.15 both directions, about $0.77 and $3.78.
The caching discount is the standout number and it deserves more attention than the headline rate. At 90 percent off repeated prompt content, Mistral matches the best caching economics in this comparison. In an agent loop that resends a system prompt, tool schema and few-shot examples on every turn, that is usually the difference between the estimate and the invoice. Structure the prompt so the stable material is at the front and the volatile material is at the end, and the discount applies itself.
We do have to flag a documentation problem alongside those numbers. When we checked in August 2026, Mistral's main pricing landing page was still quoting the superseded $2 and $6 flagship figures while the dedicated API pricing page listed $0.50 and $1.50. Per-model context windows were not published on the pricing pages or the models overview. If you are planning capacity you need the context window, and here you will have to establish it by testing. Expect to spend an extra half day on that, and build a margin into any estimate you give a client.
The licensing side is much clearer. The flagship and small tiers are Apache 2.0, which is a real OSI-approved open-source licence with no revenue thresholds, branding requirements or acceptable-use riders. That is a cleaner position than the Llama Community Licence and it means deploying the weights needs no legal review.
Agent reliability, and the limits of what anyone can tell you
Mistral positions Medium 3.5 as its agentic and coding tier and supports native function calling across the line. What nobody can give you, us included, is an independent replicated benchmark of its tool-calling reliability against the frontier US models. Vendor positioning is a statement of intent. Treat it as such and test.
The published research on long-horizon agents is provider-neutral and applies here as everywhere. LOCA-bench, released in early 2026, grows agent context under controlled conditions while holding the task fixed and reports that most models including frontier and open-source ones degrade sharply as context grows, with the spread between models widening at length. AgentCE-Bench, from April 2026, found that a 10 percent tool-call failure rate visibly degrades performance across every task horizon and difficulty level tested. Compaction between steps, retry and backoff, idempotency on writes and a human escalation path are the answers, and none of them depend on which vendor you picked.
There is one Mistral-specific reliability question that matters more here than elsewhere: you do not know the context window. Every other provider in this directory publishes it somewhere, even if you have to hunt. If you cannot look up the limit, you cannot design a compaction threshold against it, and an agent that silently truncates is worse than one that errors. Establish the limit empirically before you ship, and set your own working limit comfortably below it.
The mitigation that costs nothing is an eval set. Twenty real cases from your own business, run against Small 4 and Large 3, graded by someone who knows the right answers. At these prices the experiment costs pennies, and it tells you more about whether Mistral suits your workflow than any amount of vendor positioning.
What to build on Mistral, and what not to
The natural fit is any automation where the residency question decides the vendor. Document processing for a professional-services firm, client-record enrichment for a European agency, case-file summarisation where the underlying material cannot leave the EU. In those situations the model choice writes itself and the good news is that the cost is now reasonable.
The second fit is on-premise. The Ministral small models are deliberately sized for edge and single-server deployment under Apache 2.0, which makes a genuinely local automation practical: a model running on a machine in your office, processing documents that never touch a network. For a small firm with a specific confidentiality constraint, that is often a better answer than any contractual assurance.
Where we would be more cautious is on the hardest agent workloads. Independent, replicated benchmarks comparing Mistral's tool-calling reliability against the frontier US models are not available, and the vendor's positioning of the medium tier as the agentic model is a statement of intent rather than a measurement. If you are building a long multi-step agent with real consequences, test it thoroughly on your own cases rather than assuming parity.
The practical shape we would recommend for a European business with a residency requirement: Mistral Small 4 as the volume tier with caching on the system prompt, Mistral Large 3 for the judgement steps, the DPA training opt-out exercised in writing, zero data retention confirmed in scope before you rely on it, and a documented evaluation set you re-run whenever the lineup changes. That gets you a compliant, affordable stack with the escape hatch of self-hosting if the requirements tighten.
Strengths for agent workloads
- EU headquartered and EU resident, outside the reach of the US CLOUD Act, which is a real distinction for some regulated European businesses
- A 90 percent caching discount, matching the best in this comparison and better than most, stacked with a 50 percent batch discount
- Apache 2.0 open weights on the flagship, the small tier and all three Ministral sizes, a genuine open-source licence with no revenue thresholds
- Ministral 3 in 3B, 8B and 14B sizes at flat rates both directions, from $0.10 per 1M, which suits edge and on-premise deployment
- Codestral coding pricing is now published at $0.30 and $0.90, having previously been absent from the price list
- A dedicated agentic and coding tier rather than one model stretched across every job
- A published data processing addendum with SCCs, a subprocessor list on the Trust Center, and annual onsite audit rights
Weaknesses
- Context windows are not published on the pricing pages or the models overview, so capacity planning means testing
- The main pricing landing page carried a stale flagship figure of $2 and $6 while the API pricing page said $0.50 and $1.50, which is the kind of inconsistency that ends up in someone's budget
- The DPA states Mistral may use feedback and aggregated data to train models unless customers opt out, which is an opt-out default rather than the opt-in one often claimed for it
- Smaller ecosystem than the US majors, with fewer first-class integrations in no-code automation platforms
- Independent, replicated benchmarks of tool-calling reliability against the frontier US models are not available, so capability claims are hard to verify
- The Premier-tier models, including Codestral, OCR 4 and Voxtral Mini Transcribe, are closed, so the open-weight story does not cover the whole lineup
How to access Mistral AI
- La Plateforme, Mistral's own API. The direct route, and where the batch and caching discounts and any zero-retention arrangement apply.
- Azure AI, AWS Bedrock and other clouds carry Mistral models, which is often the easiest procurement path for a business already in one of those ecosystems.
- OpenRouter and similar aggregators for testing against other providers on the same workload.
- Self-hosting the Apache 2.0 models from Hugging Face. The small and Ministral tiers are genuinely sized for a single server or even an edge device, which is the point of the range.
Data privacy and residency
- Mistral is headquartered in France and operates as an EU company. For businesses whose concern is specifically that a US provider could be compelled to disclose data under the CLOUD Act, this is the structural answer rather than a contractual patch on top of a US vendor.
- We are correcting our own earlier wording here. Mistral's data processing addendum, read in August 2026, sets Mistral as processor and you as controller, incorporates standard contractual clauses for transfers outside the EEA, and states that personal data is retained for the duration of the agreement and becomes inaccessible within 30 days of termination. It also states that Mistral may use feedback and aggregated data to train models unless the customer opts out. That is an opt-out default, not the opt-in default we and others have previously described, and it is a meaningful distinction if training on your content is a hard no for you.
- The DPA is otherwise strong on the things procurement asks about: GDPR Article 32 security commitments, prompt breach notification, a current subprocessor list published on the Trust Center, CCPA provisions restricting sale of personal data, and audit rights allowing one annual onsite inspection with 90 days notice. Very few AI vendors at this price point offer contractual onsite audit rights.
- Zero data retention is referenced in Mistral's Trust Center material as an available option. We were not able to read a page setting out its exact scope and how it is enabled, so treat the availability as indicated rather than the terms as known, and ask Mistral to confirm scope before you design a zero-retention workflow around it.
- The Apache 2.0 open weights give you the same escape hatch as the other open providers: if no policy is good enough for a particular workload, run the model on your own hardware and there is no policy involved.
- As always, the discipline that matters more than the vendor choice is data minimisation. Send the automation what it needs to do the job and redact the rest, whoever is processing it.
What we build on Mistral AI
- Automations for EU and UK businesses with contractual or regulatory data-residency requirements
- On-premise and edge deployments using the small Apache-licensed models, where nothing leaves the building
- Sensitive document workflows under the DPA, with the training opt-out exercised and zero data retention confirmed in writing first
- Coding and agentic automation on the medium tier
- The cheap volume tier of a two-model stack for businesses that want everything inside the EU
Model choice is a line item, not a strategy. If you want the workflow designed, built and handed over with a plain-language report of what it ran and what it saved, start with a free AI automation audit or see the full done-for-you automation services.
Mistral AI FAQ
How much does the Mistral API cost?
From Mistral's API pricing page in August 2026: Mistral Large 3 at $0.50 per 1M input and $1.50 output, Medium 3.5 at $1.50 and $7.50, Small 4 at $0.15 and $0.60, Codestral at $0.30 and $0.90, Magistral Medium at $2.00 and $5.00, and Ministral 3 at flat rates of $0.10, $0.15 and $0.20 for the 3B, 8B and 14B sizes. Batch processing is 50 percent off and caching is 90 percent off. Ignore the main pricing landing page, which still quoted a superseded $2 and $6 flagship figure.
What would a real Mistral automation cost per month?
A 1,500-run inbox agent at about 3,000 input and 400 output tokens per run costs roughly $1.04 a month on Mistral Small 4, or about $0.55 with the 90 percent caching discount applied to the system prompt. A 4,000-document extraction job at 6,000 input and 300 output tokens each is about $4.32, or $2.16 batched. On Mistral Large 3 those become about $3.15 and $13.80.
Is Mistral GDPR-friendly for EU businesses?
It is the structurally simplest option for EU residency: Mistral is a French company operating in the EU, so the CLOUD Act concern that applies to US-headquartered providers does not arise the same way. Its DPA puts Mistral in the processor role, incorporates standard contractual clauses for non-EEA transfers, commits to GDPR Article 32 security, publishes a subprocessor list, and grants one annual onsite audit with 90 days notice. Read the training clause carefully though, since it is opt-out rather than opt-in.
Does Mistral train on my data?
Its data processing addendum states that Mistral may use feedback and aggregated data to train models unless the customer opts out. That is an opt-out default, not the opt-in default Mistral is often credited with, including in earlier versions of this page. It is still narrower than the training-on-inputs defaults documented by DeepSeek and Moonshot, but if training on your content is unacceptable, exercise the opt-out explicitly and keep the record.
How good are Mistral's caching and batch discounts?
Among the best in this comparison. The API pricing page documents caching at 90 percent off and batch processing at 50 percent off across the Medium, Large and Small tiers. For an agent loop resending a heavy system prompt every turn, a 90 percent caching discount is usually worth more than a lower headline rate elsewhere. Keep stable content at the front of the prompt and volatile content at the end so the cache actually hits.
Can I run Mistral models on my own hardware?
Yes for several of them. Mistral Large 3, Mistral Small 4, all three Ministral 3 sizes, Voxtral Small and Shieldstral are documented as Apache 2.0. Mistral Medium 3.5 uses a modified MIT licence. Apache 2.0 is a genuine open-source licence with no revenue thresholds or branding conditions, so no legal review is needed before deploying. The Premier-tier models, including Codestral, OCR 4 and Voxtral Mini Transcribe 2, remain closed.
How much does Codestral cost now?
It is published, which it was not earlier in 2026: $0.30 per 1M input and $0.90 output on Mistral's API pricing page as of August 2026, under API name codestral-2508. The licence is Premier, so the weights are closed even though much of the rest of the lineup is Apache 2.0. At that rate it is cheap enough to run continuous background engineering automation against your own repository.
Is Mistral good enough for agent workloads?
Its medium tier is positioned as the agentic and coding model and native function calling is supported. What we cannot offer you is independent replicated benchmarks against the frontier US models on tool-calling reliability, so treat vendor positioning as intent rather than measurement and test on twenty of your own real cases before committing anything with real consequences to it.
What is Mistral's context window?
Not published on the pricing pages or the models overview we checked as of August 2026, which is a genuine documentation gap and the only one of its kind among the ten providers in this directory. If your build depends on a specific context limit, establish it empirically against the model you plan to use and set your own working limit comfortably below it, because an agent that silently truncates is worse than one that errors.
Mistral or OpenAI for a European small business?
If you have no hard residency requirement, OpenAI has the broader ecosystem and better documentation and is usually the faster path to a live automation. If you do have a residency requirement, or if a client contract makes a European legal entity easier to justify, Mistral is the natural answer and since the Large 3 price cut it is no longer an expensive one.
