Mistral AI
Listed at https://console.mistral.ai
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Large 4: $0.70/M ($0.07 cached), Small 4: $0.15/M, Codestral: $0.30/M | per million tokens |
| Output | Large 4: $2.10/M, Small 4: $0.60/M, Codestral: $0.90/M | per million tokens |
🤖 Supported Models (17)
✨ Pros
- ✓Mistral Large 4 open weights: 1T params / 52B active MoE, 1M context
- ✓Large 4 API $0.70/$0.07 cached/$2.10, Batch 50% off, Regional 1.1x
- ✓Ecosystem Open-Weight License: free internal use, MaaS needs a separate deal
- ✓Embedded DSpark speculative-decoding head, up to 1.9x throughput
- ✓Large 3 & Medium 3.5 retired 2026-10-31; replaced by Large 4
- ✓Pixtral 1.6B vision encoder, natively multimodal
⚠️ Cons
- ×China requires proxy; no mainland-China regional endpoint
- ×Large 4 is Public Preview: same price as GA but weights can update silently
- ×Large 4 is ~40% pricier than the Large 3 it replaces ($0.70/$2.10 vs $0.50/$1.50)
- ×Reasoning cannot be disabled (reasoning_effort=none still uses high); thinking tokens are billed
- ×Chinese capabilities relatively weak
🎯 Best For
European market compliance; code generation (Codestral); embeddings/search (Mistral Embed)
💰 Pricing & Plans
| Model | Input ($/M tok) | Output ($/M tok) | Context | Best For |
|---|---|---|---|---|
| Mistral Large 4 | $0.70 | $2.10 | 1M | Frontier open-weight coding/agentic (Public Preview) |
| Mistral Medium 3.5 | $1.50 | $7.50 | 256K | Frontier multimodal, agentic & coding |
| Mistral Small 4 | $0.15 | $0.60 | 256K | Hybrid instruct+reasoning+coding, cost-efficient |
| Mistral Large 3 | $0.50 | $1.50 | 256K | Best general-purpose performance |
| Codestral | $0.30 | $0.90 | 128K | Code completion, FIM, generation |
| Ministral 8B | $0.15 | $0.15 | 256K | Edge / low-cost text+vision |
| Mistral Embed | $0.10 | — | 8K | Semantic & code embeddings |
🔧 API & Developer Experience
- •API Style: REST chat-completions-style endpoint on La Plateforme with streaming (SSE) support; most OpenAI/Anthropic SDKs and wrappers connect after a base-url and key swap.
- •SDKs & Ecosystem: First-party Python, TypeScript/Node, and Go SDKs plus an OpenAI-compatible client; Mistral also publishes many model weights openly, so there is a large self-hosted community and tooling ecosystem.
- •Agentic & Coding: Mistral Medium 3.5 is the flagship for agentic and coding workloads, with native tool/function calling and structured output across a 256K context; Codestral targets low-latency code completion and FIM.
- •Open Weights: Mistral Small 4 and Ministral are Apache 2.0; Medium 3.5 ships under a Modified MIT license. Weights can be self-hosted (e.g. vLLM, Ollama), a key differentiator for data-sovereignty teams.
- •Batch & Scale: The Batch API offers a 50% discount for asynchronous high-volume jobs; the platform also covers OCR (mistral-ocr-4-0 per page), embeddings, and fine-tuning in one console.
- •EU Data Residency: API is served from EU data centers. Pricing is in USD/EUR with pay-as-you-go prepaid credits; OCR-4 and Document AI are billed per page rather than per token.
🎯 European Sovereign AI & Open-Weight Coding
Mistral's defining position is European sovereign AI delivered through open-weight models. Unlike most frontier labs that serve a closed API and guard their weights, Mistral publishes Apache-2.0 and Modified-MIT models that teams can self-host on their own infrastructure — a decisive advantage for banks, governments, and healthcare orgs bound by EU data-residency and GDPR rules. Mistral Large 4 is the current frontier-class multimodal flagship tuned for agentic and coding use cases at a 1M context (Public Preview); Medium 3.5 and Large 3 retired 2026-10-31, while Mistral Small 4 unifies instruct, reasoning, and coding in a single efficient model. This open + sovereign combination means a regulated European enterprise can get frontier-level coding and agent work without sending data to a US cloud. The trade-off is reach: strong in EU/UK and increasingly in North America, but with a smaller third-party ecosystem and weaker Chinese language performance than domestic or US rivals.
🌐 China Access & Latency
Mistral is a European company serving its API from EU data centers, with no mainland China regional endpoint and no official China access program. Production use from mainland China therefore requires a stable overseas proxy or an aggregator that fronts the Mistral API; direct connections typically time out or fail. Latency from China is dominated by the cross-border round trip to EU (or occasional US) regions, making Mistral a poor choice for latency-sensitive interactive workloads. For teams targeting Chinese end users, a domestic LLM (e.g. DeepSeek, Qwen, Kimi) is the latency-safe default, while Mistral remains attractive for EU data-residency requirements, multilingual European documents, and teams that want open-weight self-hosting. Mistral's own OCR-4 and Document AI products are also priced per page and best reached through a stable proxy or a Chinese cloud's OCR endpoint for high-volume scanning.