Mistral Small 3.2 (mistral-small-2506) API-Ready: One-Liner Swap, Eval Before Promoting

ApiDelta · 2026-05-16 · 272 mots · apidelta.maxiaworld.app

LLM API Daily Advisory — 2026-05-16

🚨 Breaking

None today. No breaking changes reported across monitored providers (Anthropic, OpenAI, Mistral, Groq, DeepSeek).


🗑️ Deprecations

Nothing new. No active deprecation deadlines in this cycle.


💰 Pricing

No pricing changes announced. Mistral has not published token pricing for mistral-small-2506 at time of writing. Before routing production traffic, verify at mistral.ai/technology/#pricing — Small models often undercut their predecessor, but confirm the per-token rate before assuming cost parity in your budget projections.


🆕 New

Mistral Small 3.2 — model ID: mistral-small-2506

Mistral released Small 3.2 (release note) and made mistral-small-2506 available via API (API changelog).

If you're pinning mistral-small-2503 or coasting on mistral-small-latest, this is your next upgrade target. The small tier is the high-volume workhorse for classification, entity extraction, structured-output pipelines, and cost-sensitive inference at scale.

What's not yet public: context window size, benchmark delta vs. 2503, JSON-mode pass rate. No official evals at time of writing — treat as early-access until third-party numbers land.

Migration is a one-liner:

# Before
model = "mistral-small-2503"  # or "mistral-small-latest" (avoid pinning aliases in prod)

# After — explicit version pin
model = "mistral-small-2506"
response = client.chat.complete(model=model, messages=[...])

💡 Tip of the Day

When any new small model lands, run your structured-output suite first. Small models regress on JSON schema compliance before they regress on free-text quality — happy-path evals won't catch it. Track three metrics against your baseline: schema compliance rate, latency p95, and token-per-dollar at your median request size.


Action for today: Shadow-swap mistral-small-2503mistral-small-2506 in staging and run your full JSON-mode and tool-call test suite. Promote only if schema compliance rate meets or exceeds your current baseline.

#api#llm#en#mistral#model-release