LLM API Daily — Mistral Small 3.2 Hits API as mistral-small-2506 (2026-05-17)

ApiDelta · 2026-05-17 · 298 mots · apidelta.maxiaworld.app

🚨 Breaking

None this cycle. Zero breaking changes across all 2 scanned items.

🗑️ Deprecations

Nothing flagged. No active deprecation deadlines.

💰 Pricing

No pricing changes reported this cycle.

🆕 New

Mistral Small 3.2 is live in the API as mistral-small-2506.

Mistral Small is the fast/cheap tier — classification, intent routing, high-volume summarization. Same competitive bracket as gpt-4o-mini and deepseek-chat for cost-sensitive throughput.

Impact by current pin: - Pinned to mistral-small-2503 → no automatic upgrade; you stay on the old version until you move. - On mistral-small-latest → already silently upgraded to 3.2; consider pinning mistral-small-2506 to freeze behavior in prod.

# Explicit pin after your eval
response = client.chat.complete(
    model="mistral-small-2506",  # replaces: mistral-small-latest or mistral-small-2503
    messages=[{"role": "user", "content": prompt}]
)
# Log response.model — resolves the alias; surface the actual version in your observability stack
print(response.model)  # e.g. "mistral-small-2506"

No independent third-party benchmarks published at time of writing — Mistral self-reported numbers only. Do not promote to prod without your own regression pass.

⚠️ Coverage note: This cycle scanned Mistral changelog only (2 items total). OpenAI, Anthropic, Google, Groq, Cohere not included. Treat as partial signal.

💡 Tip of the Day

Mistral's response object returns a model field reflecting the resolved version, not the alias you passed. If you're calling mistral-small-latest anywhere in production and not logging response.model, you have invisible behavior drift every time Mistral bumps latest. One log line closes that blind spot permanently.

Action today: Run mistral-small-2506 against 50 representative prompts from your prod traffic. Compare latency p50/p95 and spot-check output quality against your current pin. Neutral or better → pin it, open a PR, commit the version freeze.

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