Claude Opus 5 ships with thinking on by default, breaks old API calls

Anthropic dropped Claude Opus 5 today, and the biggest shock isn’t the benchmark scores — it’s that your existing API integration is probably broken. The model now forces thinking on by default, a hard reversal from Opus 4.8. If you’re setting thinking: { "type": "disabled" } with effort set to xhigh or max, you’ll eat a 400 error immediately. Anthropic’s fix is blunt: cap effort at high or strip the field entirely.

Under the hood, the model is a legitimate beast. It scored 43.3% on FrontierBench at max effort, more than doubling Opus 4.8’s 18.7%. On SWE-bench Verified it hit 96.0%. The agentic numbers are where things get spooky — 70.57% on OSWorld 2.0, up from 55.7%. Anthropic positions this as Fable 5-level intelligence at half the price, and on the GDPval-AA v2 leaderboard, the xhigh effort setting actually beats every other model while using 25% fewer output tokens than max.

But the safety story has a twist. Opus 5’s classifiers are far less jumpy than Fable 5’s — they flagged just 5% of API calls versus 42%. Anthropic also relaxed one restriction: vulnerability finding in source code is now unblocked because, as they see it, Opus 5 is nearly as good as Mythos 5 at finding bugs but substantially worse at exploiting them. Binary scanning, pentesting, and exploit generation remain behind the guardrails. On the ARC-AGI-3 benchmark, it posted a verified 30.16%, roughly quadruple the previous best, though max-effort results weren’t available at launch.

The practical advice buried in the docs? Delete your verification prompts. Instructions like “include a final verification step” now cause over-verification because the model already checks its own work. The minimum cacheable prompt drops to 512 tokens. Output caps at 128k tokens on the sync API, or 300k if you use the new beta header on Message Batches. Same price as before. Same model ID. Radically different behavior.