Mistral’s API lineup is more focused than OpenAI’s or Anthropic’s. As of April 2026, you have one flagship, one coding specialist, and three small models — plus some legacy variants you should skip. That’s it.
Short version: Mistral Large 2512 for general coding, Devstral for code-heavy agentic work, Ministral 8B for budget tasks. Everything older than the 2512 generation is priced at old-tier rates and isn’t worth using.
The quick answer
| Model | Input/Output Cost | Context | Best For |
|---|---|---|---|
| Mistral Large 2512 | $0.50 / $1.50 | 262K | Default for general work |
| Devstral 2512 | $0.40 / $2.00 | 262K | Coding-focused agents |
| Ministral 14B 2512 | $0.20 / $0.20 | 262K | Budget Mistral |
| Ministral 8B 2512 | $0.15 / $0.15 | 262K | Tiny fast tasks |
| Ministral 3B 2512 | $0.10 / $0.10 | 131K | Edge deployment |
| Mistral Large 2407 / 2411 | $2 / $6 | 128K | Legacy — use 2512 |
| Mistral Small 3.1/3.2 | $0.08–$0.35 | — | Legacy — use Ministral |
Start with Mistral Large 2512 for general coding. Swap to Devstral if your OpenClaw workflow leans agentic (long tool-calling chains, write-test-iterate loops).
Mistral Large 2512 — the default pick
$0.50/M input, $1.50/M output, 262K context, function calling and vision. 2512 is Mistral’s flagship as of the December 2025 release, and it’s priced at roughly a quarter of the old Large 2407 tier.
262K context is the headline feature — competitive with GLM and DeepSeek, more than you get from GPT-5 Mini at this price point. Tool calling is reliable; Mistral put real engineering into the function-calling stack for the 2512 generation and it shows. Arguments are well-formed, fields are respected.
Where Mistral Large lands relative to alternatives at $0.50/M:
- vs DeepSeek V3.2 ($0.28/$0.40) — Mistral is pricier on input and output, but more reliable. DeepSeek’s API throws 503s during peak hours; Mistral mostly doesn’t.
- vs GLM-4.7 ($0.39/$1.75) — similar tier. GLM is slightly cheaper and has roughly equivalent coding quality. Mistral wins on European compliance story and multilingual support (French, German, Italian, Spanish are notably stronger).
- vs Gemini 2.5 Flash ($0.30/$2.50) — Mistral is cheaper on output, Gemini is cheaper on input. Pick based on workload shape.
If you’re building for a European market or have multilingual requirements, Mistral Large 2512 is a genuinely strong default.
Devstral 2512 — the coding specialist
$0.40/M input, $2.00/M output, 262K context, function calling. Devstral is Mistral’s coding-tuned variant, positioned for OpenClaw’s core use case.
Devstral is cheaper than Large on input but more expensive on output. The bet is that coding agents generate a lot of output tokens (writing files, producing diffs, explaining reasoning) so the output-heavy pricing matters. If your OpenClaw usage is read-heavy (analyzing code, asking questions), stick with Large. If you’re writing code, Devstral is typically cheaper end-to-end.
On coding benchmarks, Devstral outperforms Large on SWE-bench-style tasks but underperforms on general reasoning. That’s the specialization trade-off — it’s a coding model, not a generalist.
Devstral’s tool-calling reliability is the best in the Mistral lineup. For multi-step OpenClaw agent runs where the model chains tools autonomously, Devstral holds up better than Large does.
Ministral 8B 2512 — the budget pick
$0.15/M input, $0.15/M output (same price both directions), 262K context. Ministral 8B is the sweet spot of the Ministral small-model line.
At this price, Ministral 8B is competitive with DeepSeek V3 and GLM-4.7 Flash on simple tasks:
- File classification and triage
- Commit message generation
- Mechanical refactors
- Documentation generation
Don’t ask Ministral 8B to reason. Don’t give it complex tool chains. For what it is — a tiny, fast, cheap model — it punches above its weight.
Ministral 14B 2512 and 3B 2512
Ministral 14B ($0.20/$0.20, 262K context) is larger and more capable than 8B for a small price bump. If you’re running 8B and hitting its quality ceiling, 14B is the obvious upgrade before jumping to Large.
Ministral 3B ($0.10/$0.10, 131K context) is positioned for edge deployment — it fits in small VRAM budgets and runs fast on consumer hardware via Ollama. For cloud OpenClaw use, 8B is almost always better per dollar. 3B’s real home is on-device.
Legacy variants — skip
Mistral Large 2407 and 2411 ($2/$6). These are the pre-2512 Large models. Mistral priced them at the old tier and hasn’t dropped the API endpoints. There’s no scenario where these are the right pick over 2512 at $0.50/$1.50.
Mistral Small 3.1 and 3.2 ($0.08–$0.35). These predate the Ministral line. Context windows are unclear (OpenRouter reports null on some). Use Ministral 8B or 14B instead.
Mixtral 8x22B ($2/$6). The old Mixture-of-Experts flagship. Was the state of the art in 2024. There’s no active reason to run it in 2026 — Mistral Large 2512 is better on every axis at a quarter the price.
Setup in OpenClaw
Mistral isn’t a built-in provider. Two routes.
Running through haimaker.ai
All Mistral models are available through haimaker.ai:
{
"models": {
"providers": {
"haimaker": {
"baseUrl": "https://api.haimaker.ai/v1",
"apiKey": "your-haimaker-api-key",
"api": "openai-completions"
}
}
}
}
Add the models to your allowlist:
{
"agents": {
"defaults": {
"models": {
"mistralai/mistral-large-2512": {},
"mistralai/devstral-2512": {},
"mistralai/ministral-8b-2512": {}
}
}
}
}
Apply with openclaw gateway config.apply.
Direct Mistral setup
Sign up at console.mistral.ai:
{
"models": {
"providers": {
"mistral": {
"baseUrl": "https://api.mistral.ai/v1",
"apiKey": "your-mistral-api-key",
"api": "openai-completions"
}
}
}
}
What I’d do
Default to Mistral Large 2512 unless you have a specific reason to pick something else. If your OpenClaw workflow is heavy on agent loops and code generation, swap to Devstral 2512. For bulk tasks where quality matters less than cost, Ministral 8B is a legitimate choice.
Mistral isn’t the cheapest option (DeepSeek and GLM are cheaper) and isn’t the highest-quality (Claude and GPT-5 are stronger on hard reasoning). Where it wins is the combination: decent quality, European data residency story, strong multilingual support, reliable API. For teams that need to stay in the EU for compliance reasons, that combination is valuable.
Pair with a flagship (Sonnet 4.6, GPT-5.4) for the hard problems Mistral can’t solve.
For API key setup, see the OpenClaw API key guide. For a full model comparison, see the complete models guide.