Mistral API Pricing in 2026: Models, Free Tier, and Cost Comparison

Mistral Large 3 costs $0.50 input / $1.50 output per million tokens. According to CloudZero's analysis, that makes it 80% cheaper on input and 90% cheaper on output than GPT-5.4, and 83% cheaper on input and 90% cheaper on output than Claude Sonnet 4.6.

Those numbers position Mistral as the cheapest frontier API in 2026. But the pricing structure is more complicated than a single rate card. Mistral runs two completely separate billing systems under one URL: per-token API charges through la Plateforme, and monthly subscriptions through Le Chat. Teams that also self-host open-weight models add GPU compute as a third cost line with no connection to either billing system. The result is AI spend that is difficult to attribute across projects or providers.

This guide covers every model's price, explains the billing splits, and puts the costs in context against the major alternatives.

Every Model and Its Price

Mistral API pricing spans from $0.02 per million input tokens (NeMo) up to $2.00 per million input tokens (Magistral Medium). The range reflects a catalog designed to cover everything from budget batch processing to frontier reasoning.

Flagship and Generalist Models

Mistral Large is designed for complex reasoning and multilingual tasks. Mistral Small is optimized for speed and cost-efficiency in high-volume processing. That split matters for model routing: teams paying Large prices for tasks Small handles fine are leaving money on the table.

Reasoning (Magistral)

A section below covers why Magistral's real cost is higher than the rate card implies.

Coding and Agentic

Codestral is trained on 80+ programming languages including Python, JavaScript, Java, and C++.

Edge (Ministral)

All three Ministral models use symmetric input/output pricing:

Multimodal (Pixtral)

Audio (Voxtral)

Specialist APIs

How Mistral Large Pricing Changed: A 75% Drop Since Large 2

Teams that evaluated Mistral 12 to 18 months ago saw a different price card. Mistral Large 2 cost $2.00 input / $6.00 output per million tokens. Mistral Large 3, released in December 2025, dropped that to $0.50/$1.50, a 75% price reduction.

At $2.00/$6.00, Large 2 was in the same range as competing frontier models. At $0.50/$1.50, Large 3 undercuts them by a wide margin. If your cost projections are based on the old pricing, they are four times too high.

Is the Mistral API Free?

Three paths to zero cost, each with its own constraints.

API evaluation tier. The Mistral API free tier provides rate-limited evaluation access. Enough to test models and validate integration code. Not enough to serve production traffic.

Le Chat free plan. Roughly 25 messages per day, access to state-of-the-art models, web search, image generation, and 40+ connectors. This is the consumer chat interface, not API access.

Open weights. Mistral publishes open weights for Small 4, Medium 3.5, Magistral Small, Devstral Small, and the full Ministral lineup (14B/8B/3B) under Apache 2.0 or Modified MIT. Self-hosted models have zero per-token API cost. Your Mistral inference cost becomes GPU compute.

The distinction matters. "Free" in Mistral's world means free to test (evaluation tier), free to chat (Le Chat), or free of per-token charges but not free of infrastructure (open weights).

Le Chat Plans: A Separate Bill from the API

This is where mistral api pricing gets confusing. Le Chat and the API are two completely separate billing systems under one URL.

Le Chat Pro at $14.99/month is cheaper than ChatGPT Plus ($20) and Claude Pro ($20). But here is the billing trap: a Pro subscription does not cover API calls.

Codestral inside Le Chat is included with a Pro subscription. Codestral through the API is a separate per-token charge. A developer using Codestral in their IDE through the API and also through Le Chat for quick questions is paying twice, through two different mechanisms.

For a broader view of how Anthropic's chat and API pricing compare, see our Claude API pricing breakdown.

Mistral vs GPT and Claude: Cost Comparison

CloudZero's pricing analysis puts the gap in concrete terms:

By those figures, Mistral Large 3 is 80% cheaper on input and 90% cheaper on output than GPT-5.4. Against Claude Sonnet 4.6, it is 83% cheaper on input and 90% cheaper on output.

Price is not the only differentiator. Mistral offers EU data residency and open weights, two advantages no US LLM provider matches. For teams operating under GDPR or those that need to self-host for compliance reasons, those features may matter more than the token rate.

For model-by-model pricing of the alternatives, see our guides on OpenAI API pricing, Gemini API pricing, and Grok API pricing. We also have a direct Anthropic vs OpenAI pricing comparison.

What Reasoning Models Actually Cost in Practice

Magistral Medium's rate card says $2.00/$5.00 per million tokens. In practice, reasoning models generate 2 to 5 times more output tokens than generalist models because chain-of-thought reasoning is verbose by design.

A request that produces 500 output tokens from Mistral Large 3 might produce 1,000 to 2,500 tokens from Magistral Medium. At $5.00 per million output tokens, that 2 to 5x multiplier turns a $0.0025 call into a $0.005 to $0.0125 call. The rate looks lower than competing reasoning models, but the token volume means the real cost per request is not what the price table suggests.

For agentic pipelines that chain multiple reasoning calls, budget for the token amplification, not the headline rate.

Cost Optimization: Caching, Tiers, and Model Routing

Prompt caching. Mistral offers a 90% discount on repeated token prefixes. System prompts, few-shot examples, and shared context that stays constant across requests benefit directly. For applications with long, stable prompts, this is the single largest cost lever.

Model selection by use case. Not every request needs the flagship. A practical split:

• Mistral Small 4 ($0.15/$0.60): agentic tasks, coding, and multimodal understanding at production scale
• Mistral Large 3 ($0.50/$1.50): complex reasoning, multilingual, high-accuracy requirements
• Codestral ($0.30/$0.90): code generation with FIM support
• Ministral 3B/8B ($0.10 to $0.15): edge deployment, on-device inference
• NeMo ($0.02/$0.03): budget multilingual tasks where cost matters more than capability

Model routing with a gateway. The optimization above only works if your application can route requests to different models based on task complexity. An AI gateway sitting between your application and Mistral's API (and other providers) can handle that routing, applying rules like "send simple classification to NeMo, send reasoning tasks to Large 3, fall back to another provider if Mistral is down." Without a routing layer, teams default to one model for everything and overpay.

The Split-Billing Problem

Mistral runs two completely separate billing systems under one URL: per-token API charges through la Plateforme, and monthly subscriptions through Le Chat. Teams that also run self-hosted open-weight models add GPU compute as another cost line outside either system.

Each billing system has its own usage metrics and its own cost curve. Neither talks to the other. A team running Mistral Large 3 on the API while paying Le Chat Pro subscriptions for product managers has two separate line items with no unified view of total Mistral spend.

Scale that across providers. Add OpenAI API costs, Claude API costs, and maybe Gemini. Now you have Mistral's split billing plus two or three other provider dashboards. Attributing AI spend to a specific team, product, or feature becomes a manual spreadsheet exercise.

This is where AI FinOps tooling becomes necessary. An AI gateway that proxies all provider traffic through a single layer can tag requests by team, project, or use case, then aggregate costs across providers and billing models into one view. The token savings from model routing are a bonus. The real value is knowing what you are spending, and where.