Complete cost breakdown for Mistral Large, Small, Nemo, and Codestral — with comparisons to Claude, GPT-4o, and Gemini.
All prices are per million tokens (MTok). Mistral bills input and output tokens separately.
| Model | Input (per MTok) | Output (per MTok) | Context | Best For |
|---|---|---|---|---|
| Mistral Large | $2.00 | $6.00 | 128K | Complex reasoning, multilingual |
| Mistral Small | $0.20 | $0.60 | 32K | Simple tasks, classification |
| Mistral Nemo | $0.15 | $0.15 | 128K | Budget-tier inference, RAG |
| Codestral | $0.20 | $0.60 | 32K | Code generation, autocomplete |
| Mistral Embed | $0.10 | — | 8K | Embeddings, semantic search |
Comparing flagship models and their budget tiers across all major providers.
| Model | Input (MTok) | Output (MTok) | Cache Discount | Context |
|---|---|---|---|---|
| Mistral Nemo | $0.15 | $0.15 | None | 128K |
| GPT-4o-mini | $0.15 | $0.60 | 50% | 128K |
| Claude Haiku 4.5 | $0.25 | $1.25 | 90% | 200K |
| Mistral Small | $0.20 | $0.60 | None | 32K |
| Mistral Large | $2.00 | $6.00 | None | 128K |
| GPT-4o | $2.50 | $10.00 | 50% | 128K |
| Claude Sonnet 4.6 | $3.00 | $15.00 | 90% | 200K |
| Gemini 1.5 Pro | $3.50 | $10.50 | None | 1M |
| Claude Opus 4.7 | $15.00 | $75.00 | 90% | 200K |
Mistral is a French company with EU data residency options. Ideal for EU-regulated workloads where data locality matters.
Mistral Large has particularly strong French, German, Spanish, Italian performance — better than GPT-4o-mini for EU language tasks.
Mistral Nemo at $0.15/MTok (both input and output) is one of the cheapest capable models available for non-critical tasks.
At $0.20/MTok, Codestral is 12.5× cheaper than Claude Sonnet for autocomplete and boilerplate-heavy coding pipelines.
| Model | Input (50M tokens) | Output (50M tokens) | Monthly Total |
|---|---|---|---|
| Mistral Nemo | $7.50 | $7.50 | $15 |
| Mistral Small | $10 | $30 | $40 |
| Claude Haiku (cached 90%) | $1.25 cached | $62.50 | ~$64 |
| Mistral Large | $100 | $300 | $400 |
| GPT-4o | $125 | $500 | $625 |
| Claude Sonnet (cached 90%) | $15 cached | $750 | ~$765 |
| Factor | Mistral Large | Claude Sonnet 4.6 |
|---|---|---|
| Input price | $2.00/MTok (cheaper) | $3.00/MTok |
| Output price | $6.00/MTok (cheaper) | $15.00/MTok |
| Prompt caching | Not available | 90% discount |
| Context window | 128K | 200K |
| Reasoning quality | Good | Excellent |
| EU data residency | Available | US-based |
| Tool use / function calling | Good | Best-in-class |
| Open weights option | Yes (Mistral 7B/8x7B) | No (API only) |
Yes. Mistral's la Plateforme offers a free tier with rate-limited API access for testing. The free tier covers all models but is not suitable for production. Commercial use requires a paid plan with per-token billing.
Yes — Mistral 7B, Mistral 8x7B (Mixtral), and Mistral 8x22B are open-weight models you can download and self-host via Ollama, vLLM, or llama.cpp. Self-hosting eliminates per-token costs (you pay only for GPU time). Mistral Large is API-only.
Free tier: 1 request/second. Paid tiers vary by plan — most paid tiers start at 60 RPM and scale up with spend. Contact Mistral for enterprise volume limits.