Blog · 2026-10-23 · By Dhrumil Barot, Founder

Mistral API Pricing: Every Model, Cache and Batch Rate

Mistral API pricing verified 2026-10-23: Medium 3.5, Large 4, Large 3, Small 4, Ministral 3, Codestral, OCR and Voxtral rates, cache and Batch discounts, the 1.1x EU/US regional upcharge and cost per task.

TL;DR: Mistral API pricing on 2026-10-23 runs from $0.10 per million tokens in and out (Ministral 3 3B) to $1.50 input / $7.50 output (Mistral Medium 3.5). Mistral Large 3 costs $0.50 / $1.50, and the newer Mistral Large 4 costs $1.36 / $4.18. Cached input costs 10% of the input rate, Batch halves everything, and EU or US regional endpoints add 10%.

Mistral API pricing is a pay-as-you-go rate card from Mistral AI: each model has a USD price per 1 million input tokens, cached input tokens and output tokens, billed separately from Le Chat and Vibe subscriptions. Prices verified 2026-10-23 against Mistral's API pricing page. On a task of 10,000 input tokens and 1,000 output tokens, one call costs $0.0011 on Ministral 3 3B, $0.0021 on Mistral Small 4, $0.0065 on Mistral Large 3 and $0.0225 on Mistral Medium 3.5. The counterintuitive result: on Mistral's current card the "Medium" model is the most expensive one, and the older Large 3 is the cheapest big model.

How much does the Mistral API cost per million tokens?

The Mistral API price is set per model, per 1 million tokens, on the global endpoint at Standard processing. These are Mistral's generalist models:

Output costs 5x input on Medium 3.5, about 3x on Large 4 and Large 3, 4x on Small 4, and exactly 1x on the three Ministral 3 models. That flat Ministral pricing makes them unusually cheap for generation-heavy jobs such as drafting, rewriting and translation. Mistral Large 4 costs 2.7x Large 3 on input and 2.8x on output, so check that your eval actually needs it. Look up model details in our /models catalog.

What do Codestral, embeddings and other Mistral APIs cost?

Mistral's specialist APIs are priced by their own units, per the Mistral pricing page:

The Mistral API also serves two third-party models from Z.ai with 1M-token context windows: GLM 5.3 at $1.40 / $0.14 / $4.40 and GLM 5.3 Flash at $0.15 / $0.03 / $0.50. Transcribing one hour of audio costs $0.21 on Voxtral Mini Transcribe 3 and $0.48 on the Realtime model. OCR 4.1 at $0.004 per page makes a 300-page contract $1.20.

How does Mistral prompt caching pricing work?

Cached input is Mistral's discount for repeated prompt prefixes: tokens read from cache cost 10% of the model's input price, which Mistral's FAQ describes as reducing input cost "by up to 90%". On Mistral Medium 3.5 a cached token costs $0.15 per million against $1.50; on Mistral Small 4 it costs $0.015 against $0.15.

Worked example: an agent with a 30,000-token system prompt and tool schemas on Mistral Medium 3.5 pays $0.045 in input per uncached call and $0.0045 when the prefix is cached. Over 10,000 calls that is $450 versus $45 on the prefix alone. Keep stable content (instructions, tool definitions, reference documents) at the start of the prompt and per-request content at the end, so the prefix matches. Our prompt caching in production guide covers why hit rates fall in real traffic.

How much does Mistral Batch save?

Mistral Batch is an asynchronous API that runs requests at a 50% discount, per the Mistral batch docs. You upload a JSONL file of up to 1 million requests (or send fewer than 10,000 inline), and Mistral processes the queue over a 24-hour window. Each line holds a custom_id and a body identical to a normal chat completions request:

{"custom_id": "0", "body": {"max_tokens": 100, "messages": [{"role": "user", "content": "Classify this ticket: refund not received"}]}}

At 50% off, Mistral Large 3 drops to $0.25 / $0.75 and Mistral Small 4 to $0.075 / $0.30. A nightly job classifying 1 million support tickets at 2,000 input and 50 output tokens each costs $330 on Small 4 Standard and $165 on Batch. Use Batch for evals, backfills, embeddings refreshes and bulk extraction; keep it off anything a user is waiting on. We compared the same trade on other providers in batch APIs: when latency is free.

What do Mistral regional endpoints and Priority Tier cost?

Regional inference is Mistral's option to keep processing inside the EU or the US, through dedicated endpoints api.eu.mistral.ai and api.us.mistral.ai. Per the regional inference docs, it is billed at 1.1x list price on input, output, cached reads and cache writes. The global endpoint api.mistral.ai has no upcharge but makes no commitment about where inference runs.

The regional limits matter for cost: you cannot combine EU residency with the 50% Batch discount, so an EU-bound bulk job pays 2.2x what the same job costs as a global batch. Priority Tier is a separate option for real-time workloads: it needs account setup with Mistral sales, uses service_tier: "auto" per request, falls back to Standard when limits are exceeded, and is the only tier with an uptime SLA (99.5%, per the Priority Tier docs). Mistral does not publish a Priority rate; it is negotiated.

What does a Mistral API task actually cost?

Cost per task is the number that goes in a budget. Assumptions: 10,000 input tokens, 1,000 output tokens, no caching, global endpoint, Standard tier. Formula: input/1M x input price + output/1M x output price. Change the token counts for your workload.

Medium 3.5 costs 3.5x Large 3 per call on this shape, and its 5x output multiplier widens the gap on long answers. Mistral itself recommends Medium "for most tasks and coding" and Small "for cost-sensitive projects"; that is a quality claim, so test both on your own prompts before paying 10x for Medium over Small 4. Compare these against other providers with the Vynaris cost calculator.

How does Mistral compare with other providers on price?

Mistral's cheapest models sit at the bottom of the market. Ministral 3 3B at $0.10 / $0.10 undercuts gemini-2.5-flash-lite ($0.10 / $0.40) on output, and Mistral Small 4 at $0.15 / $0.60 is in the same band as gpt-oss-120b on AWS Bedrock ($0.15 / $0.60). At the top end, Mistral Medium 3.5 at $1.50 / $7.50 costs less than Claude Sonnet 5.5 ($2 / $10) on Anthropic's API.

Mistral Large 3 is also on AWS Bedrock at the same $0.50 / $1.50 in us-east-1, so AWS customers can buy it on their existing invoice without a price penalty; see our AWS Bedrock pricing and Gemini API pricing breakdowns for the full cross-provider picture. Because Large 3, Small 4 and the Ministral 3 models are open weight under Apache 2.0, you can also self-host them; whether that beats $0.10 to $0.50 per million tokens depends on utilization, which we modeled in self-hosting vs API break-even.

Is Mistral's Le Chat or Vibe subscription the same as API access?

Mistral's consumer plans are subscriptions to the Le Chat and Vibe apps, not the API rate card. Pro costs $14.99 per month, Team $24.99 per user per month, and verified students pay $5.99. The pricing page notes that paid plans can keep going past their limits with pay-as-you-go credits "at API rate", and paid plans list a monthly allowance of API credits. Production API traffic is billed per token through Mistral Studio, where the dashboard breaks spend down by endpoint (completions, OCR, fine-tuning).

Which Mistral models are being retired?

Mistral retires old model versions on published dates, and calls to a retired ID fail. Per the Mistral models overview, mistral-large-2411 (Mistral Large 2.1) was retired on 2026-05-31 and open-mistral-nemo-2407 on 2026-07-31. Older IDs such as open-mistral-7b and open-mixtral-8x7b were retired in 2025. If your code pins a dated ID, move it to a current model and re-run your cost math: Large 3 at $0.50 / $1.50 and Large 4 at $1.36 / $4.18 differ by almost 3x.

When is the expensive Mistral model worth it?

Mistral Medium 3.5 at $22.50 per 1,000 calls is worth it over Large 3 at $6.50 only if Large 3 needs more than about 3.5x the tokens or attempts to finish the same task. That can happen on agentic coding and multi-step tool use, which is what Mistral positions Medium 3.5 for. It rarely happens on classification, extraction, translation or summarization, where Ministral 3 8B at $1.65 or Small 4 at $2.10 per 1,000 calls is the right default. Measure success rate and tokens per task before choosing by list price.

Where Vynaris fits

Vynaris is an OpenAI-compatible API that routes each request to the cheapest model that can handle it, across Mistral, OpenAI, Anthropic, Google and open-weight hosts. With a 20x price spread inside Mistral's own card, routing easy requests to Ministral and hard ones to Large or Medium is where savings come from. See Vynaris pricing. If you need EU-only processing, use Mistral's EU endpoint directly; a router adds nothing to a residency requirement.

FAQ

How much does the Mistral API cost?

From $0.10 per 1M tokens (Ministral 3 3B, input and output) to $1.50 input / $7.50 output (Mistral Medium 3.5). Mistral Large 3 costs $0.50 / $1.50.

What is the cheapest Mistral model?

Ministral 3 3B at $0.10 per 1M input and output tokens, or $0.05 on Batch.

Is the Mistral API free?

The Free plan gives limited access to Le Chat, Vibe and Mistral Studio. Production API usage is billed per token; Mistral Moderation 2 is free.

How much is Mistral Large?

Mistral Large 3 costs $0.50 input / $1.50 output per 1M tokens; Mistral Large 4 costs $1.36 / $4.18. Cached input is $0.05 and $0.14.

Does Mistral offer a batch discount?

Yes, 50% off Standard prices for asynchronous jobs of up to 1 million requests, processed within 24 hours. Batch is not available on regional endpoints.

Is Mistral cheaper than OpenAI?

Not across the board. OpenAI's gpt-6-luna at $0.10 / $0.50 is cheaper than Mistral Small 4 at $0.15 / $0.60, while Ministral 3 3B at $0.10 / $0.10 is cheaper than both on output. See our OpenAI API pricing page for every GPT tier.