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-2503 → mistral-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.