Mistral API Pricing 2026: Every Model per Million Tokens, Free Mode, Batch, Priority and Cost per Conversation

Published 28 October 2026. Every price below was read from Mistral's API pricing page and model cards on 28 October 2026, and every price is as of that date.

Mistral's API charges per million tokens, from $0.10 on Ministral 3 3B to $7.50 for output on Mistral Medium 3.5, as of 28 October 2026. Mistral Small 4, the model Mistral points cost-sensitive projects to, costs $0.15 per million input tokens, $0.015 per million cached input tokens and $0.60 per million output tokens. Mistral Large 3 costs $0.50, $0.05 and $1.50; Mistral Large 4, in Public Preview since 6 October, costs $1.36, $0.14 and $4.18; and Mistral Medium 3.5, the model Mistral recommends "for most tasks and coding", costs $1.50, $0.15 and $7.50. Cached input bills at 10% of the input price, the Batch API takes 50% off, Priority Tier costs 1.75 times list, and the EU and US regional endpoints cost 1.1 times list. New accounts start in Free mode, with no credit card and a monthly allowance whose size Mistral does not publish.

Three things make most Mistral pricing guides wrong today. The API rate card is not on mistral.ai/pricing, which sells the Vibe subscriptions; mistral.ai/pricing/api redirects to the docs pricing page, so guides that read the marketing page miss changes. At least one widely read guide, updated on 20 October, lists Mistral Small 4 at $0.10 input and $0.30 output, while Mistral's own model card says $0.15 and $0.60. And Large 4 launched at 50% off for two weeks, according to Mistral's changelog, so pages that captured the launch price (half of list, $0.68 and $2.09) now understate it by half. In the worked example further down, a 40-turn conversation costs $0.080 on Mistral Small 4 without caching and $0.019 with it, the same uncached cost as DeepSeek's deepseek-flash off-peak and Groq's GPT-OSS 120B, and less than half of Groq's cached figure.

Mistral price per million tokens, every current text model

US dollars per million tokens, from Mistral's API pricing page and each model's card, as of 28 October 2026. The cached column is from the rendered pricing table and matches Mistral's prompt caching guide, which bills cached tokens at 10% of input; the Markdown copy of the same page and several model cards list only input and output. The page also has a EUR toggle; check the currency your organization is billed in before you budget.

The names do not follow the prices. Medium 3.5 costs more than both Large models, with output at 1.8 times Large 4 and five times Large 3; it is the newer model (April 2026), which Mistral describes as optimized for agentic and coding use, while Large 3 (December 2025) is the older open-weight generalist. And the aliases do not point where the names suggest. On 28 October, mistral-large-latest still resolves to Large 3 according to its model card, so code that wants Large 4 has to call mistral-large-4 by name, and mistral-medium-latest points to Medium 3.5. Mistral's lifecycle policy warns that "using an alias may expose you to silent updates in model behavior and pricing", which is a good reason to pin the dated or major-minor ID in production.

Large 4 costs 2.7 times Large 3 on input and 2.8 times on output, adds a 1M-token context window and native reasoning, and is a Public Preview model, which Mistral prices "at the same rate as General Availability models" but which "may be retired before reaching" GA, with one month's deprecation notice instead of six. The two Z.ai GLM models are third-party models that Mistral hosts and serves "without Mistral modifications"; they also carry one month's notice. Ministral 3 is the only family that charges the same for output as for input, which makes it unusually cheap for output-heavy work such as rewriting or extraction.

What changed in 2026, and why other guides show different prices

Source: Mistral's models overview and changelog, read 28 October 2026. Several of these still show the status "Deprecated" with a retirement date that has passed.

The pages that rank for "mistral api pricing" still quote many of these: Mistral Nemo at $0.02 input, Mistral Small 3.2 at $0.08 or $0.10, Mistral Medium 3 at $0.40 and $2.00, Magistral Medium at $2.00 and $5.00, Pixtral Large at $2.00 and $6.00. Those were real prices. Once a model passes its retirement date, Mistral's lifecycle page says requests to it "fail with a 404 error", so none of them is a budget you can plan on. The native reasoning models (magistral-small-latest, magistral-medium-latest) are deprecated in favour of the reasoning_effort parameter on Small 4, Medium 3.5 and Large 4, which is covered below. GLM 5.2 is the one live change this week: it stops answering on 31 October.

Mistral Free mode, plan credits and pay-as-you-go

Every new Mistral organization starts in Free mode. Mistral's quickstart says API access "is enabled by default with no credit card required" and that "usage and rate limits apply", and the usage and limits page says Free mode lets you "use included monthly usage within the limits shown on the Limits page". Mistral does not publish the size of that allowance or the free rate limits; each organization sees its own in the Admin Panel under API, Limits. The old help-centre articles on the "Experiment plan", which other guides still cite for the 2024 free-tier limits, returned HTTP 404 when we fetched them on 28 October.

Mistral's plans are now global across Vibe (launched on 28 May 2026, with the old Le Chat kept as its Chat mode), Studio and the API, according to the Subscriptions page, and "each Mistral plan includes monthly usage that is shared across Studio, the API, and Vibe Code". On mistral.ai/pricing as of 28 October:

The Pro figure is worth noticing: the plan card lists more API credit ($25.5) than the plan costs ($14.99), and on the cached Small 4 conversation priced below, $25.5 covers roughly 1,370 forty-turn conversations, or roughly 320 without caching.

Past the included usage, Mistral sells pay-as-you-go, and it is prepaid. The Billing page says your organization needs "both pay-as-you-go enabled and a positive credit balance", and when the balance reaches zero, "API and Vibe Code requests beyond your included monthly usage fail with HTTP status 402". Admins get an email at 80% of the last top-up, auto-recharge is available, and credits are deducted including tax while spending limits apply before tax, so "your credits can run out before you reach your spending limit". Without pay-as-you-go, usage past the allowance "can stop until the next billing period".

Labs models, the ones whose API names start with labs-, are "free of charge" under the lifecycle policy, can be updated silently and are not recommended for production. On 28 October the Labs table lists no available labs- model; the most recent, Leanstral 1.5, was scheduled to retire on 30 September.

Mistral prompt caching: 10% of the input price, keyed by you

Cached prompt tokens are "billed at 10% of the standard input token price", according to Mistral's prompt caching guide, which is $0.015 instead of $0.15 per million on Small 4 and $0.15 instead of $1.50 on Medium 3.5. There is no write fee in the guide and no storage fee on the pricing page. You opt in by sending the same prompt_cache_key (a conversation, session or workflow ID) on requests that share a prefix; the key "increases the chance of a cache hit, but it doesn't guarantee one". Caching works in 64-token blocks, so cached_tokens is always a multiple of 64 and a prompt under 64 tokens never hits, and you read what was billed in usage.prompt_tokens_details.cached_tokens.

A 90% discount puts Mistral's cache read between Groq, which takes 50% off on a hit, and DeepSeek's deepseek-flash, which takes 98% off, and level with Claude Haiku 5.5's 90% read, without Claude's 1.25 times write fee. For long conversations that difference is most of the bill, and the comparison table below shows it.

Mistral Batch API: 50% off, 24 hours by default

The Batch API runs requests asynchronously "at a 50% discount", and Mistral's batch FAQ says it "is available for all models". On Small 4 that is $0.075 input and $0.30 output per million tokens; on Large 3 it is $0.25 and $0.75. Batch covers chat completions, FIM, embeddings, moderation, OCR, classification, conversations and audio transcription, one model per batch. Mistral states these rules:

• Up to 1 million requests in one batch, and a cap of 1 million pending requests per workspace, so two 600,000-request jobs cannot run at once.
• Inline batching, with no file upload, for batches of fewer than 10,000 requests.
• timeout_hours defaults to 24 and must be under 7 days; a batch that does not finish in time expires.
• Results "do not expire at this time", and batches "may slightly exceed your workspace's configured spend limit".

Batch is not available on the EU and US regional endpoints, which do not support stateful features. Mistral's docs do not say whether the batch discount stacks with the cache discount, so budget batch work at the plain 50% rate.

Priority Tier and regional inference: 1.75x and 1.1x

Mistral sells the same models through three service levels and two optional regions, according to its Priority Tier and regional inference pages:

Priority Tier needs an agreement with Mistral first; you then send service_tier: "auto", and a request over your Priority limits falls back to Standard rather than failing. If you omit the parameter, the request runs as standard_only. The 1.75 multiplier "applies after the prompt cache discount", so a cached Small 4 token costs $0.02625 per million on Priority. On the 40-turn Small 4 conversation below, Priority turns $0.080 into $0.140 without caching and $0.019 into $0.033 with it.

Regional endpoints are for data-location requirements. They serve only the models hosted in that region, support function calling but not the other built-in tools, and do not offer Agents, Batch or the Files API. Mistral adds that the control plane stays global: "account configuration, API keys, billing, access management, usage analytics, and other operational metadata may still be handled by Mistral systems outside the selected inference geography". On the default api.mistral.ai, Mistral "does not commit to a specific inference location".

Reasoning tokens, and why Mistral tells you to resend them

Small 4, Medium 3.5, Large 4 and GLM 5.3 all reason on request through reasoning_effort, according to Mistral's reasoning guide. At high, the response "includes a full thinking chunk before the final answer, at the cost of increased token usage", and those thinking tokens count as completion tokens, billed at the output rate. Mistral recommends high for agentic and coding work on Medium 3.5. At none, the model "thinks minimally" and returns no thinking chunk.

The detail that matters for a chat bill is in the multi-turn section of the same guide: "always replay the full assistant message (including ThinkChunk) back into the message history", because "dropping the reasoning trace across turns degrades model performance". Mistral's own example shows 187 completion tokens from turn 1 coming back as prompt tokens in turn 2. Reasoning is therefore billed twice in a conversation, once as output on the turn that produced it and again as input on every later turn.

Run the conversation below with 1,000 thinking tokens on every reply, all replayed, and the 40th request carries 60,650 tokens instead of 21,650. The 40-turn bill on Small 4 rises from $0.080 to $0.221 without caching and from $0.019 to $0.060 with it; on Medium 3.5 it rises from $0.819 to $2.29 and from $0.207 to $0.677. Mistral's warning says stripping the thinking chunks "increases token efficiency but significantly degrades output quality", so set reasoning effort per task: none for routing and classification, high only where the answer needs it.

What a 40-turn conversation costs on Mistral, DeepSeek, Groq, Gemini and Claude

Take a chat assistant with a 2,000-token system prompt, where each user message is 150 tokens and each reply is 350 tokens; these are the same assumptions as our Groq API pricing, DeepSeek API pricing, Gemini API pricing and Claude API pricing guides, so the figures line up across all five. Turn 1 sends 2,150 input tokens and turn 40 sends 21,650, because the API is stateless and each request carries every exchange before it. The whole 40-turn conversation consumes 476,000 input tokens to produce 14,000 output tokens, and a 100-turn conversation consumes 2,690,000 input tokens. Reasoning tokens are left out.

Mistral rows are our arithmetic on the rates in the first table, read 28 October 2026. The DeepSeek, Groq, Gemini and Claude rows are the figures our sibling guides published from rates read on 27 October 2026 (deepseek-flash at $0.15 input, $0.003 cache hit and $0.60 output off-peak, double at peak; GPT-OSS 120B at $0.15, $0.075 and $0.60; GPT-OSS 20B at $0.075, $0.037 and $0.30; Haiku 5.5 at $0.10, $0.50 output, $0.125 cache write and $0.01 read; Sonnet 5.5 at $2, $10, $2.50 and $0.10; Gemini 3.8 Flash at $0.75, $3.75 and $0.075 cached); the method for each is on its own page.

"Cache warm" on Mistral means that from turn 2, the previous request's prompt is read at the cached rate, rounded down to Mistral's 64-token blocks, and the rest of the prompt (the last reply, the new message and the remainder of the last block) bills at the full rate, with no write fee. That is the same model we used for DeepSeek and Groq; Claude's rows include its 1.25 times cache write, and Gemini's implicit cache starts once a prompt passes 4,096 tokens.

Without caching, Small 4 costs exactly what deepseek-flash off-peak and Groq's GPT-OSS 120B cost, $0.080 for 40 turns, because all three list at $0.15 input and $0.60 output. With caching, Small 4 at $0.019 sits between DeepSeek off-peak ($0.013) and Claude Haiku 5.5 ($0.014) on one side and Groq GPT-OSS 120B ($0.046) on the other, because Mistral's 90% cache discount beats Groq's 50% and loses to DeepSeek's 98%. Ministral 3 3B is the cheapest model in the table with caching, at $0.008, and the cheapest Mistral model without it, at $0.049 (only Groq's GPT-OSS 20B, at $0.040, comes in lower), because its output costs the same $0.10 as its input. And Medium 3.5, at $0.819 uncached and $0.207 cached, costs less than Claude Sonnet 5.5 but nearly twice Gemini 3.8 Flash on a warm cache, with output at $7.50 doing most of the damage.

At 10,000 forty-turn conversations a month, Small 4 costs $798 without caching and $186 with it, Large 3 costs $2,590 and $551, and Medium 3.5 costs $8,190 and $2,073. On Claude Sonnet 5.5 the same volume costs roughly $10,920 and $2,396, according to our Claude pricing guide.

The table also shows the shape of the problem. Going from 40 to 100 turns multiplies turns by 2.5 and the uncached Small 4 bill by 5.3, because every turn carries every turn before it, and reasoning traces replayed into the history make the curve steeper. Caching cuts the price of the repeated tokens; it does not stop them being sent, and nothing in Mistral's cache helps a user who comes back after the cache has gone cold. The other lever is to stop resending the whole transcript. We build Maximem Synap, a memory layer that returns a short, ranked set of facts about the user, trimmed to a token budget you set (2,000 tokens by default), in place of the full history. When that beats caching, and when it does not, is worked through in how to reduce LLM token costs in long conversations.

OCR, audio, voice agents, embeddings and moderation

Source: Mistral's API pricing page. Shieldstral 1.0 (moderation) and the four Timestral 1.0 time-series models are open-weight, self-hosted models with no API price.

Processing 10,000 pages costs $40 on OCR 4.1, $50 with annotations, and $20 on OCR 3, which Mistral keeps available "for existing integrations and production workloads". Transcribing 1,000 hours of audio costs $210 on Voxtral Mini Transcribe 3, $240 on Small Transcribe 3 and $480 on the realtime model. The Voxtral 3 family went into Public Preview on 20 October, the same day Mistral launched Voice Agents, a managed pipeline that handles capture, speech-to-text, the model, text-to-speech and playback, at $0.04 a minute on Small 4; 10,000 minutes of conversation costs $400 on Small 4 and $800 on a larger model. Moderation 2 is free, so there is no cost reason to skip a moderation pass.

Fine-tuning, agent tools and Libraries: what Mistral publishes and what it does not

Self-serve fine-tuning is deprecated. Mistral's fine-tuning page is headed "This feature is deprecated and is no longer actively supported", and keeps the old terms, a minimum of $4 per fine-tuning job and $2 a month to store each model, for existing users. Classifier Factory carries the same notice. Custom models are now an Enterprise-plan line on mistral.ai/pricing, with no public price.

The built-in agent tools, web search, code interpreter and image generation, as well as Libraries and Connectors, are billed separately from tokens, and Mistral's usage dashboard breaks out "Agents and Connectors usage", "Libraries API usage" and "Document Library storage". Mistral's pricing FAQ says "tool APIs are priced per call", but on 28 October we could not find a per-call rate on mistral.ai/pricing, on the API pricing page or in the tool documentation. Some third-party guides quote per-1,000-call figures; we could not confirm any of them on a Mistral page, so check the usage dashboard after a test run before you put a tool into a high-volume agent.

Rate limits, billing and spending limits

Mistral publishes no numeric rate limits. Limits are set per Workspace and shared by every API key in it, as requests per second, tokens per minute and tokens per month, and they "vary by subscription tier and model", according to the Workspace limits page and Mistral's known limitations. Your own numbers are in the Admin Panel under API, Limits; a request over them returns HTTP 429, and Mistral asks for the models, target requests per second and expected tokens per minute and per month when you request more. Audio and OCR have their own limits, in audio seconds and pages per minute.

Spending limits are the control to set first. An Organization monthly limit caps API and Vibe spend together, Workspace limits sit under it (no cap by default), and reaching either suspends API access for that scope "until the next month begins or an admin increases the limit". Workspaces make a cheap way to separate development, staging and production budgets, since a capped Workspace does not affect the others. Two gaps are worth knowing: batches can run slightly past a Workspace cap, and credits are drawn including tax while caps are measured before it.

Data, training and who may use the API

Mistral's Commercial Terms of Service, effective 25 September 2026, say Mistral will not train on customer data except where you have not opted out of a product set to opt-in by default, where you send feedback, where an order form says so, or "when Customer uses Labs or Preview Models". For those models the terms are blunt: Mistral "may use Customer Data and Outputs generated from Labs or Preview Models to train", and the opt-out "including through zero data retention" does not apply. Mistral Large 4 and the Voxtral 3 models are in Public Preview on 28 October, so a team with a no-training requirement should stay on GA models such as Large 3, Medium 3.5 or Small 4 until they graduate.

The API is a business product. Mistral's consumer terms for users outside the EEA, effective 12 October 2026, cover Vibe "but excluding Mistral AI Studio and access to any of our APIs, which are limited to business customers", and the Commercial Terms govern use by "the organization, company, or other entity that you represent". For processing location, use the EU or US endpoint at 1.1 times list; the default global endpoint gives no location commitment.

Which Mistral model to pick on cost

For a chat or agent workload on a budget, start with Small 4. It is GA, Apache 2.0, multimodal, reasons on request and costs $0.15 and $0.60, the same list price as DeepSeek deepseek-flash off-peak and Groq GPT-OSS 120B, with a 256K context and a 90% cache discount. If your outputs are long, compare it with Ministral 3 8B or 14B, whose output price equals their input price; for classification and routing, Ministral 3 3B at $0.10 flat is the cheapest model Mistral sells.

Step up to Large 3 ($0.50 and $1.50) when Small 4 falls short on your evals, and treat Medium 3.5 ($1.50 and $7.50) and Large 4 ($1.36 and $4.18) as the models you justify per task. Mistral recommends Medium for most tasks and coding and names Medium 3.5 as the replacement for Devstral 2 and Magistral Medium, so coding agents are its natural home; Large 4 brings the 1M context but is a preview model with the training caveat above. GLM 5.3 Flash is the other 1M-context option, at $0.15 and $0.50, with third-party notice periods.

Open weights change the long-run arithmetic. Small 4 (119B parameters, 6.5B active), Large 3 and Ministral 3 ship under Apache 2.0 and Medium 3.5 under a modified MIT licence, according to their model cards, and Mistral says Large 4's weights are "coming soon". Self-hosting swaps the per-token price for GPU and operations cost, and Mistral's pricing FAQ notes that commercial deployments may need a Mistral licence for some models, so read the licence on the model card before planning around it.

How to estimate your monthly Mistral bill, and cut it

A request costs (cached input tokens times the cached rate) plus (uncached input tokens times the input rate) plus (output and thinking tokens times the output rate), halved on Batch, multiplied by 1.75 on Priority and by 1.1 on a regional endpoint. For a conversation, the input on each turn is the system prompt plus every earlier turn, including replayed thinking, so estimate per conversation as in the table above and multiply by conversations per month. OCR is pages divided by 1,000 times the page rate; audio is minutes times the minute rate.

The levers that move a typical Mistral bill, roughly in order of size:

1. Default to Small 4 or Ministral 3, not Medium 3.5. Medium 3.5's output costs 12.5 times Small 4's.
2. Send a prompt_cache_key on every multi-turn request. Caching is opt-in on Mistral, and a cache hit costs 10% of input.
3. Run anything that can wait through Batch. Evaluations, back-fills and bulk extraction get 50% off on every model.
4. Set reasoning_effort per task. Thinking bills as output and comes back as input on every later turn.
5. Pin model IDs. A -latest alias can change behaviour and price without notice.
6. Stop resending the whole conversation once chats run long, by summarising or retrieving only what the next turn needs.

Set an Organization spending limit and per-Workspace caps before any of this, because a 402 at zero credit is a better failure than an invoice.

Frequently asked questions

How much does the Mistral API cost per million tokens?

As of 28 October 2026, Mistral Small 4 costs $0.15 per million input tokens and $0.60 per million output tokens, Mistral Large 3 costs $0.50 and $1.50, Mistral Large 4 costs $1.36 and $4.18, Mistral Medium 3.5 costs $1.50 and $7.50, and Ministral 3 costs $0.10 to $0.20 for both input and output. Cached input costs 10% of the input price, and Batch costs 50% less.

Is the Mistral API free?

Partly. New organizations start in Free mode, which needs no credit card and includes a monthly usage allowance with rate limits, but Mistral does not publish the size of either; you see them in the Admin Panel. Labs models (labs- prefix) are free, and Mistral Moderation 2 is free on every plan.

Does Mistral Pro include API credits?

Yes. As of 28 October 2026 the Pro card on mistral.ai/pricing lists $25.5 a month in API credits for $14.99 a month, and the student Education version lists $15 a month for $5.99. Plan usage is shared across Studio, the API and Vibe Code.

How much does Mistral Large 4 cost?

Mistral Large 4 costs $1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens as of 28 October 2026. It launched on 6 October 2026 at 50% off for two weeks, which is why some pages list $0.68 and $2.09. It is a Public Preview model, and mistral-large-latest still points to Large 3.

Does Mistral have a Batch API discount?

Yes. Batch jobs cost 50% less on every model, with up to 1 million requests per batch, a default 24-hour timeout that can be set to anything under 7 days, and results that do not expire. Batch is not available on the EU and US regional endpoints.

How does Mistral prompt caching work and what does it cost?

You send the same prompt_cache_key on requests that share a prefix, and cached tokens bill at 10% of the input price, in 64-token blocks. There is no write fee, a hit is not guaranteed, and the cached count appears in usage.prompt_tokens_details.cached_tokens.

What is Mistral Priority Tier?

Priority Tier is an enterprise option that routes requests through a priority queue with a 99.5% uptime SLA, billed at 1.75 times list price on input, output and cached tokens. It needs an agreement with Mistral and per-model limits, and requests over those limits fall back to Standard.

Does Mistral charge more for EU data processing?

Yes. The EU endpoint (api.eu.mistral.ai) and the US endpoint (api.us.mistral.ai) cost 1.1 times list price on input, output and cache tokens, and they do not support Batch, Agents or the Files API. The default api.mistral.ai endpoint makes no commitment about where inference runs.

What are the Mistral API rate limits?

Mistral publishes no numbers. Limits are per Workspace, shared across its API keys, and set as requests per second, tokens per minute and tokens per month by plan and model; each organization sees its own under Admin Panel, API, Limits, and exceeding them returns HTTP 429.

How much does Mistral OCR cost?

OCR 4.1 costs $4 per 1,000 pages, or $5 per 1,000 pages with annotations, and OCR 3 costs $2 and $3, as of 28 October 2026.

Can I still fine-tune Mistral models?

Self-serve fine-tuning is deprecated and "no longer actively supported"; existing jobs carried a $4 minimum and $2 a month per stored model. Custom models are now sold through the Enterprise plan.

Is Mistral cheaper than DeepSeek, Groq, Gemini or Claude?

It depends on the model and on caching. A 40-turn conversation costs $0.080 on Mistral Small 4 without caching, the same as DeepSeek deepseek-flash off-peak and Groq GPT-OSS 120B, and $0.019 with caching, against $0.013 on DeepSeek off-peak, $0.014 on Claude Haiku 5.5 and $0.046 on Groq GPT-OSS 120B. Mistral Medium 3.5 costs $0.819 and $0.207, below Claude Sonnet 5.5 ($1.09 and $0.240) and above Gemini 3.8 Flash ($0.410 and $0.111).

Sources

Mistral sources retrieved 28 October 2026; comparison figures from our guides published 27 October 2026.

• Mistral documentation: API pricing, models overview, model cards for Mistral Small 4, Mistral Large 3, Mistral Large 4, Mistral Medium 3.5, Ministral 3 3B, Ministral 3 8B, Ministral 3 14B, Codestral, Z.ai GLM 5.3, Z.ai GLM 5.3 Flash and OCR 4.1; prompt caching, Batch processing, Priority Tier, regional inference, reasoning, model lifecycle, Labs, changelog, Studio quickstart, usage and limits, Workspace usage and limits, billing, subscriptions, known limitations, fine-tuning (deprecated) and Voice Agents
• Mistral: pricing (mistral.ai/pricing/api checked and found redirecting to the docs pricing page), Commercial Terms of Service and consumer terms outside the EEA
• Maximem: Groq API pricing, DeepSeek API pricing, Gemini API pricing and Claude API pricing

Worked examples are our own arithmetic on those rates.