Messages API responses now report how many billed output tokens were extended thinking.

Billed thinking is now a separate usage detail, so builders can inspect reasoning spend directly.

For streaming builds, the breakdown arrives on the final message_delta event and works without a beta header.

• Jun 12, 2026Anthropic publicly documented stop_details on refusal responses, with category and explanation fields for routing.
• Jun 12, 2026Claude Opus 4.8 defaults the effort parameter to high across all surfaces.
• Jun 12, 2026Claude Opus 4.8 lowers prompt caching’s minimum cacheable prompt length to 1,024 tokens versus Opus 4.7.

This is the world we build in. It moves this fast. We move with it.