Mistral API Pricing 2026: The Overlooked Budget Powerhouse

Here is a number that should make you reconsider your LLM bill: Mistral Large 3, the company's open-weight flagship, costs $0.50 per million input tokens and $1.50 per million output. That is roughly 80% cheaper on input than OpenAI's GPT-5.6 ($2.50) and 90% cheaper than Anthropic's Claude Opus 4.8 ($5). The output gap is wider still.

Mistral is a French AI company that ships competitive open and premier models at aggressive price points, and most engineering teams we work with have never seriously evaluated them. This guide breaks down Mistral API pricing for every model as of July 2026, across text, reasoning, code, vision, voice, and OCR. For each, you get what it costs and where it makes sense against the big three.

All token prices are per million tokens. Run your own numbers in our LLM Pricing Calculator.

The Full Mistral API Pricing Table (July 2026)

Text and reasoning models, priced per million tokens:

Specialized services (documents, voice, embeddings, moderation):

Agent and tool calls, billed per use on top of model tokens:

Vision is now built into Large 3, Medium 3.5, and Small 4, so the standalone Pixtral Large model has been retired. Leanstral, an open Lean 4 code agent, is currently free.

Why Mistral Deserves Your Attention

The pricing tells the story immediately. Here is how the flagships compare on a representative 5M-input / 2M-output workload:

Mistral Large 3 runs that workload for $5.50 — versus $34 on Gemini 3.1 Pro, $42.50 on GPT-5.6, and $75 on Claude Opus 4.8. That is roughly 8x cheaper than OpenAI's flagship and 14x cheaper than Anthropic's, for the same tokens.

Bar chart: monthly cost of a 5M-input / 2M-output workload — Mistral Large 3 $5.50 vs Google Gemini 3.1 Pro $34, OpenAI GPT-5.6 $42.50, Anthropic Claude Opus 4.8 $75

Monthly cost for the same 5M-input / 2M-output workload across flagship models, from official provider pricing (July 2026).

Raw price only matters if quality holds. Mistral Large 3 benchmarks competitively with previous-generation flagships and, as an open-weight model, can also be self-hosted. It may not match GPT-5.6 or Claude Opus 4.8 on the hardest tasks, but for a large share of production workloads the quality gap is smaller than the price gap.

Model Deep Dives

Mistral Large 3: The Value Flagship

At $0.50/$1.50, Large 3 offers one of the best price-to-quality ratios on the market. It is open-weight, multimodal (text + vision), supports function calling, and can be run through the API or self-hosted.

The output price is the standout: $1.50 per million tokens, against $15 for GPT-5.6, $25 for Claude Opus 4.8, and $10 for Claude Sonnet 5. If your workload is output-heavy (content, code, or report generation), Large 3 cuts output cost by 85–94%.

Best for:

• General-purpose production workloads where you are currently overpaying a premium provider
• Output-heavy pipelines (content, code, reports) where the $1.50 output price compounds
• Multilingual applications — Mistral models are strong across European languages
• Teams that want the option to self-host an open-weight flagship

Mistral Medium 3.5: The Enterprise Tier

Counterintuitively, Medium 3.5 ($1.50/$7.50) costs more than the Large 3 flagship. It targets enterprise deployments, tuned for quality and simplified private hosting rather than lowest cost. For most API users optimizing spend, Large 3 is the better pick; reach for Medium 3.5 when its specific quality profile or deployment story matters to you.

Mistral Small 4: The Volume Workhorse

At $0.15/$0.60, Small 4 competes with Gemini 2.0 Flash ($0.15/$0.60), Gemini 2.5 Flash-Lite ($0.10/$0.40), and GPT-5.4 Mini ($0.75/$4.50), undercutting the OpenAI option heavily on output. It is Apache-2.0 licensed, multimodal, and supports function calling, which is rare at this price.

Ministral 3: Frontier AI at the Edge

New in the current lineup, the Ministral 3 family (3B, 8B, 14B) is built for edge and on-device use, with flat pricing where input and output cost the same: $0.10 (3B), $0.15 (8B), and $0.20 (14B) per million tokens. For latency- or privacy-sensitive workloads that run close to the user, these are the cheapest competent options Mistral offers.

Magistral: Mistral's Reasoning Models

The Magistral line uses extended chain-of-thought for complex problems. Magistral Medium ($2.00/$5.00) and Magistral Small ($0.50/$1.50) deliver transparent, multilingual reasoning. Since OpenAI folded reasoning into the GPT-5 family (GPT-5.4 at $2.50/$15), Magistral Medium's $5 output undercuts it by two-thirds; Magistral Small matches Large 3's price while optimizing for reasoning.

Codestral & Devstral: Purpose-Built for Code

Codestral ($0.30/$0.90, 256K context) is a low-latency model for completion, fill-in-the-middle, and generation. Devstral 2 ($0.40/$2.00) and Devstral Small 2 ($0.10/$0.30) are agentic coding models for autonomous software engineering, and Leanstral is a free open code agent for Lean 4. Against GPT-5.6 ($2.50/$15) or Claude Sonnet 5 ($2/$10) for code, Codestral is 6–15x cheaper on the same tokens.

Vision, Voice, and Documents

Vision is now built into Large 3, Medium 3.5, and Small 4, so the dedicated Pixtral Large model has been retired and you no longer pay a premium for multimodal input. For speech, Voxtral covers text-to-speech ($0.016 / 1,000 characters) and transcription (from $0.003 / minute). For documents, Mistral OCR 4 runs at $4 / 1,000 pages ($5 with Document AI extraction), priced per page rather than per token.

Embeddings & Moderation

Mistral Embed ($0.10 / 1M) and the code-specialized Codestral Embed ($0.15 / 1M) cover retrieval and RAG, and Mistral Moderation ($0.10 / 1M) provides content classification, all at commodity prices.

Real Cost Calculations

Use Case 1 — AI Email Drafting

20,000 drafts per month, roughly 800 input + 400 output tokens each: 16M input, 8M output.

Small 4 undercuts everyone at $7.20/month; Large 3 at $20 still beats GPT-5.4 Mini by more than half. Claude Sonnet 5 at $112 is about 16x Small 4 for this shape of work.

Use Case 2 — Code Review Pipeline

500 PRs per month, roughly 6,000 input + 1,500 output tokens each: 3M input, 0.75M output.

Codestral at $1.58/month for 500 reviews is remarkably cheap: an 8–12x cost advantage over Sonnet 5 and GPT-5.6. Even at 80% of their quality it is worth a serious evaluation.

Use Case 3 — Reasoning-Heavy Analysis

1,000 tasks per month, roughly 5,000 input + 3,000 output tokens each: 5M input, 3M output.

Magistral Small delivers reasoning at $7/month, roughly one-eighth of GPT-5.4. Magistral Medium at $25 still lands under every premium option while giving a larger reasoning budget.

Where Mistral Falls Short

Transparency matters, so here is where Mistral is not the right pick:

Context window: Mistral's flagships now reach 256K tokens (Magistral and the smallest edge model stay at 128K). That is a real jump from the 128K of a year ago, but still short of GPT-5.6's 400K or the 1M windows on Gemini 3.1 Pro and Claude Sonnet 5 / Opus 4.8.

Ecosystem and tooling: OpenAI, Google, and Anthropic have larger ecosystems, with more SDKs, integrations, and community. Switching cost is real if you depend on them.

Peak instruction-following: Claude Sonnet 5 and Opus 4.8 remain the reference for long, nuanced, ambiguous instructions. If precision on complex prompts is critical, Mistral may need more prompt engineering to match.

Enterprise features: verify Mistral's current SOC 2, DPA, and SLA coverage against your requirements; the incumbents have more mature programs.

Benchmarks vs reality: competitive benchmark scores do not guarantee production performance. Always run your own evaluations.

The Migration Path

If you are considering a move from OpenAI or Anthropic to Mistral:

• Pick your highest-volume, lowest-complexity workload. Classification, extraction, and simple generation are the best candidates.
• Run a parallel evaluation: send the same 200–500 requests to both your current model and the Mistral equivalent, and score on your real quality criteria.
• Calculate the real savings: token cost plus latency, error rates, and any prompt engineering needed to adapt.
• Migrate the easy wins; keep the harder tasks on your current provider.
• Re-evaluate quarterly, since Mistral ships new models frequently.

Most teams we work with end up multi-provider: Mistral for volume, Claude or GPT-5.6 for the tasks that need peak quality. Blended cost is typically 40–60% lower than running a single premium provider for everything — see our full OpenAI vs Gemini vs Claude pricing comparison for the head-to-head.

The Bottom Line

Mistral is not a compromise. It is a legitimate alternative that happens to cost 80–95% less than the incumbents for many workloads. The question is not “is Mistral good enough” — it is “have you actually tested it on your workload?”

If you have not, you may be leaving real money on the table.

Prices verified against Mistral's official API pricing page (mistral.ai/pricing) as of July 2026. Compare Mistral against every provider in our LLM Pricing Calculator.

Sources

Pricing verified against official provider pricing pages, retrieved 2026-07-17:

• Mistral model documentation
• OpenAI API pricing
• Anthropic (Claude) API pricing
• Google Gemini / Vertex AI pricing