Anthropic shipped Claude Opus 4.8 yesterday. 41 days after 4.7. That is the pace of someone who knows the competition is about to drop something big.

What actually changed:

- Agentic coding: 64.3% → 69.2%
- 4x less likely to let bugs pass without flagging them
- Fast mode roughly 2.5x quicker
- The model is now more willing to say "I am not sure"

They are calling it "the most honest model yet." The real story is that they trained it to admit uncertainty. This is the opposite of how every other AI provider has trained for the last five years. Honesty as a competitive moat.

Side detail buried in the launch: in a simulated town benchmark, Claude's agents committed less crime than any other model.

I am still deciding whether that is funny or terrifying.

Mythos class teased for the coming weeks. 🐶