Opus 5.5 is here! 🎉 It's honestly such a nice model to work with. Feels like Fable, but ~30% faster and ~40% cheaper per task than Opus 5. Go try it out!↗

On harnesses, I vacillate between three beliefs: - the less harness, the better. Models are the magic - post training a model and harness is dramatically better and the model providers win - harnesses have real independent value from the model I have no idea which is right.↗

Introducing SWE-1.7, the most capable model we’ve trained yet. It scores within a few points of the strongest frontier models at a fraction of the cost, and is now available at 1000 tok/s. RL is not hitting its limit: after refining our recipe, we keep seeing gains as we scale↗

Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use. Its capabilities exceed those of any model we’ve ever made generally available.↗