Mistral AI Pricing 2026 — Live Rates, Training & Opt-Out

Last updated:

How much does Mistral AI cost? The table below renders the current per-token rates from our daily canonical dataset, including cached-input pricing. The documented training opt-out adds no token fee; Vibe, Studio/API, feedback, and Zero Data Retention still have different controls.

How much does each Mistral model cost per million tokens?

Showing 7 current models. .
Relative priceLowestLowerHigherHighestAutomatic log scale across current offers; input, cache and output are graded separately. Color shows cost, not quality.
Showing 7 grouped models from 7 offers · USD per 1M tokens
Last synced:

Sponsored links may earn us a commission at no extra cost to you. Affiliate status never changes model ordering.

All product names, logos, and brands are property of their respective owners and are used for identification purposes only.

Last synced:

September 8 funding watch

Mistral raises €3B; current API rates are unchanged

Mistral raised €3 billion at a post-money valuation above €21 billion to expand frontier research, training compute, infrastructure, commercial growth, and international reach. The announcement names no new model, API endpoint, or token rate, so the daily-synced table above remains the pricing source of truth.

Read the Mistral €3B funding, sovereign AI, and pricing analysis for the buyer impact and Labs scope blocker.

September 2 policy watch

Mistral trains on Vibe data by default unless you opt out

Mistral's current help center says Vibe input and output data are used for model training by default unless the user opts out. Vibe Enterprise is opted out by default; Mistral Studio Free data may be used, while pay-as-you-go customers can opt out. Feedback and Zero Data Retention follow separate rules.

No model or token rate changed with the documentation update. Read the Mistral data-training, opt-out, ZDR, and pricing analysis before sending confidential prompts.

August 4 launch

Shieldstral 1.0: open weights, no hosted rate

Shieldstral is a 3.8B-parameter, policy-adaptive moderation model for text and images. Mistral says it fits on one 16GB GPU in BF16 and recommends staying within its 32K-token training range. The weights carry an Apache 2.0 license, but production inference still has infrastructure and operations costs.

Read the Shieldstral pricing and deployment analysis for benchmark caveats, the Labs availability blocker, and a buyer checklist.

Key facts about Mistral AI pricing

  • Codestral: $0.3 input / $0.9 output per 1M tokens in the current canonical dataset.
  • Ministral 3 3B: $0.1 input / $0.1 output per 1M tokens in the current canonical dataset.
  • Ministral 3 8B: $0.15 input / $0.15 output per 1M tokens in the current canonical dataset.
  • Ministral 3 14B: $0.2 input / $0.2 output per 1M tokens in the current canonical dataset.
  • Mistral Large 3: $0.5 input / $1.5 output per 1M tokens in the current canonical dataset.
  • Mistral Medium 3.5: $1.5 input / $7.5 output per 1M tokens in the current canonical dataset.
  • Mistral Small 4: $0.15 input / $0.6 output per 1M tokens in the current canonical dataset.
  • Shieldstral 1.0 is open-weight and self-deployable; it is not listed as a priced hosted Shieldstral API SKU.
  • Open-weight licensing removes a separate weight-license charge, not GPU, storage, scaling, monitoring, or engineering costs.

Why choose Mistral over OpenAI or Anthropic?

Mistral's positioning comes down to European deployment options, open-weight availability, and a compact active API catalog. The daily-synced table is the reliable place to compare unit rates; use the calculator to measure your actual input/output mix rather than carrying an old launch price into a budget.

Shieldstral strengthens the openness case for moderation workloads. Teams can inspect and self-host the checkpoint, but they should compare end-to-end moderation cost, latency, false-positive review, and maintenance against a managed classifier before moving production traffic.

When Mistral is not the right choice

A managed provider can be simpler when your priority is a predictable invoice, first-party service-level commitments, or minimal model operations. Compare the live OpenAI and Anthropic rates, then test cost per accepted task or correctly moderated item under the same workload.

Price History

Only models with a recorded price change are charted here.

Ministral 3 3B

Ministral 3 8B

Mistral Large 3

Mistral Small 4

Price history tracking started April 2026. Flat model charts stay hidden until a price change is detected.
View pricing changelog →

Frequently asked questions

Does Mistral AI train on user data?

Mistral says Vibe input and output data are used for model training by default unless the user opts out. Vibe Enterprise is opted out by default. Mistral Studio Free data may be used for training, while pay-as-you-go customers can opt out.

Does paying for Mistral automatically opt data out of training?

Do not infer the training setting from payment alone. Mistral documents separate controls for Vibe, Vibe Enterprise, Studio Free, and pay-as-you-go. Zero Data Retention is also a separate, approval-based control for supported stateless API calls.

How much does Mistral AI charge per token?

Current active Mistral rates in our 2026-10-16 dataset are: Codestral $0.3 input / $0.9 output; Ministral 3 3B $0.1 input / $0.1 output; Ministral 3 8B $0.15 input / $0.15 output; Ministral 3 14B $0.2 input / $0.2 output; Mistral Large 3 $0.5 input / $1.5 output; Mistral Medium 3.5 $1.5 input / $7.5 output; Mistral Small 4 $0.15 input / $0.6 output. Rates are USD per 1 million tokens and are synced from Mistral's official API pricing page.

How much does Shieldstral cost?

Mistral released Shieldstral 1.0 as Apache-2.0 open weights without publishing a hosted Shieldstral API model ID or per-token rate. Downloading the weights has no separate license fee, but self-hosting still incurs GPU, storage, scaling, observability, and engineering costs.

Is Shieldstral available through the Mistral API?

Not as a priced hosted SKU at launch. The official release and model card point to downloadable weights and local deployment, while Mistral's API rate card does not list Shieldstral. Do not treat the separate Mistral Moderation classifier service as a Shieldstral endpoint.

Which active Mistral model is cheapest?

Ministral 3 3B has the lowest active input rate in today's dataset at $0.1 per 1 million input tokens. Use the live table because Mistral changes its active catalog over time.

What hardware does Shieldstral need?

Mistral says the 3.8B-parameter checkpoint fits on one 16GB GPU in BF16. The model card recommends a maximum model length of 32,768 tokens for deployment because that is the training range, even though the architecture theoretically supports more.

How should buyers compare Shieldstral with a moderation API?

Measure cost per correctly moderated item, including GPU-hours, autoscaling headroom, image preprocessing, false-positive review, false-negative risk, monitoring, and model maintenance. A free-to-download checkpoint is not the same as zero-cost production moderation.

Methodology

Hosted API pricing sourced from Mistral's official API rate documentation on . Data-training claims are checked against Mistral's official training-use policy, opt-out guidance, and ZDR guidance. Shieldstral availability is checked separately; no Shieldstral API price is inferred from the open-weight release.

Compare Mistral to other providers

View All Models OpenAI Pricing Anthropic Pricing Groq Pricing Token Calculator

Further reading: Full AI API pricing comparison · Cheapest AI APIs in 2026.