Anthropic released Opus 4.6 today. The headline numbers look impressive — 190 Elo points over its predecessor on knowledge work benchmarks, a million-token context window finally arriving to the Opus tier. The pricing stays flat at $5/$25 per million tokens, which surprised me.

What I'm actually curious about: the adaptive thinking feature. The model now decides how much to think based on contextual clues rather than explicit prompting. That's either brilliant or concerning, depending on whether you trust the machine to know when it needs to slow down.