About 19; min
Mistral AI’s API powers applications that need strong multilingual support, function calling, and European data residency. The model lineup spans from tiny edge models to flagship multimodal systems. Here’s the complete guide to Mistral API pricing in 2026.
Model Pricing Overview
| Model | Input (per M tokens) | Output (per M tokens) | Context | Best For |
|---|---|---|---|---|
| Mistral Large 2 | $2.00 | $6.00 | 128K | Top-tier reasoning |
| Mistral Small 3 | $0.20 | $0.60 | 32K | Most production tasks |
| Codestral | $0.20 | $0.60 | 32K | Code generation |
| Pixtral Large | $2.00 | $6.00 | 128K | Vision + text |
| Mistral Embed | $0.10 | N/A | 8K | Embeddings |
| Mistral Nemo | $0.15 | $0.15 | 128K | Lightweight tasks |
| Ministral 3B/8B | $0.04 / $0.10 | $0.04 / $0.10 | 128K | Edge deployment |
Mistral Large 2 — The Flagship
Mistral Large 2 (123B parameters) is Mistral’s flagship for tasks that demand top-tier reasoning. At $2.00/M input and $6.00/M output, it’s priced below Claude Sonnet ($3/$15) and at the same input price as GPT-4o ($2.50/$10) — though output is slightly cheaper than both. The 128K context window handles long documents and codebases. Native function calling supports parallel tool execution for complex AI agents. Multilingual performance is exceptional across 12 languages, particularly French, German, Spanish, and Italian. The model excels at code generation, mathematical reasoning, and instruction following. For European businesses or applications requiring strong multilingual capability, Mistral Large 2 provides frontier-class performance with EU-based infrastructure options.
Mistral Small 3 — The Workhorse
Mistral Small 3 (24B parameters) handles most production tasks at a fraction of Large’s cost. At $0.20/M input and $0.60/M output, it competes directly with GPT-4o mini ($0.15/$0.60) and Claude Haiku ($0.80/$4.00) — landing between them on price. The model achieves quality close to Mistral Large on standard benchmarks while running 3-10x faster. For high-volume applications (chatbots, classification, content generation, summarization), Small 3 hits the sweet spot of capability and cost. The 32K context window is smaller than competitors but sufficient for most production use cases. Mistral Small 3 is also available as an open-weight model under Apache 2.0 for self-hosting.
Codestral — Code Generation Specialist
Codestral is Mistral’s coding-specialized model. At $0.20/M input and $0.60/M output, it’s priced like Mistral Small 3 but tuned specifically for code completion, generation, and explanation. The model supports 80+ programming languages with native fill-in-the-middle for editor integrations. Codestral 2501 (latest version) achieves competitive results against larger general-purpose models on coding benchmarks while running faster and cheaper. For IDE plugins, code review tools, and automated programming assistants, Codestral provides production-grade quality at the budget tier price point. Available through the Codestral API endpoint or Mistral’s standard API.
Pixtral Large — Vision + Text
Pixtral Large is Mistral’s multimodal model that processes images alongside text. Pricing matches Mistral Large 2 at $2.00/M input and $6.00/M output. The 128K context window accommodates multiple images plus extensive text. Pixtral handles document understanding, chart analysis, screenshot interpretation, and general visual reasoning. The model is built on Mistral Large 2 with vision capabilities added — text quality matches the flagship while gaining multimodal input. For applications that need to analyze images, extract information from documents, or combine visual and textual reasoning, Pixtral Large provides European AI’s answer to GPT-4o vision and Claude Sonnet vision capabilities.
Ministral — Edge Models
Ministral 3B and 8B target on-device and edge deployment. The 3B model at $0.04/M input/output is among the cheapest API options available — designed for tasks like classification, routing, and simple chat where small model capability is sufficient. Ministral 8B at $0.10/M offers stronger reasoning while still running on consumer hardware (8GB VRAM). Both models support 128K context. The open-weight versions enable on-device deployment on phones, edge devices, and constrained environments. For applications needing AI inference at the network edge or on user devices, Ministral provides Mistral’s compact options.
Free Tier and Limits
Mistral provides a free tier through La Plateforme with rate-limited access to all models. New accounts include credit for testing. The free tier supports development and prototyping but requires a paid plan for production volume. The pay-as-you-go model has no minimum commitment — pay only for tokens used. For enterprise customers, Mistral offers committed-use pricing with volume discounts and dedicated infrastructure options including European data residency.
Cost Scenarios
Customer support chatbot (50K conversations/month)
Average: 500 input + 300 output tokens.
- Mistral Small 3: (25M × $0.20/M) + (15M × $0.60/M) = $5 + $9 = $14/month
- Mistral Large 2: (25M × $2.00/M) + (15M × $6.00/M) = $50 + $90 = $140/month
Code generation tool (10K requests/month)
Average: 1,000 input + 1,500 output tokens.
- Codestral: (10M × $0.20/M) + (15M × $0.60/M) = $2 + $9 = $11/month
- Mistral Large 2: (10M × $2.00/M) + (15M × $6.00/M) = $20 + $90 = $110/month
Document analysis with vision (1K PDFs/month, mixed text and images)
Average: 30K input + 2K output per document.
- Pixtral Large: (30M × $2.00/M) + (2M × $6.00/M) = $60 + $12 = $72/month
Mistral vs Competitors
| Tier | Mistral | OpenAI | Anthropic | |
|---|---|---|---|---|
| Flagship input | $2.00 (Large 2) | $2.50 (GPT-4o) | $3.00 (Sonnet) | $1.25 (Gemini Pro) |
| Mid-tier input | $0.20 (Small 3) | $0.15 (4o mini) | $0.80 (Haiku) | $0.10 (Flash) |
| Edge model | $0.04 (3B) | N/A | N/A | N/A |
| Multilingual strength | Excellent (12 languages) | Very good (50+) | Very good | Very good |
| EU data residency | Yes (native) | Via Azure | Limited | Yes |
Mistral’s pricing falls between Google’s aggressive low rates and OpenAI’s mid-tier. The differentiator isn’t price but European data residency, strong multilingual capability, and open-weight availability for self-hosting. For European businesses with data sovereignty requirements, Mistral provides the best combination of capability and compliance.
Mistral Small 3 is the best starting point for most Mistral API applications. At $0.20/M input tokens, it handles the vast majority of production tasks at a fraction of Large 2’s cost. Mistral Large 2 earns runner-up for tasks that demand top-tier reasoning and multilingual quality — the 12-language excellence beats most competitors for European applications. Codestral excels at code-specific work. The combination of competitive pricing, strong multilingual capability, EU data options, and open-weight availability makes Mistral the European AI choice that no US-based competitor matches.
FAQ
Can I self-host Mistral models?
Yes, for the open-weight versions. Mistral Small 3, Codestral, Mistral Nemo, and Ministral 3B/8B are available under Apache 2.0. Mistral Large 2 requires research/commercial licensing. For privacy-sensitive deployments, the open-weight options provide flexibility the API doesn’t.
Does Mistral offer European data residency?
Yes. Mistral AI is based in France and operates infrastructure in EU regions. For GDPR-strict applications and European businesses, Mistral provides native EU data handling without the complications of using US-based AI providers.
How does function calling work on Mistral?
All Mistral models support native function calling with parallel tool execution — call multiple tools in a single request and process results together. The implementation follows OpenAI-compatible patterns, simplifying migration from OpenAI’s function calling.
Should I use Codestral or Mistral Small 3 for coding?
Codestral is fine-tuned specifically for code and outperforms Small 3 on coding benchmarks at the same price. Use Codestral for IDE integrations, code review, and programming-focused applications. Use Small 3 for general tasks that include some coding.
