🙂 Anthropic Launches Claude Sonnet 5.5

The model is 30% faster thanks to lower token usage, and falls only slightly behind the more powerful Opus 5.5 in coding and general knowledge.

In preliminary tests, researchers noticed some unusual traits. For example, when acting as a "therapist," the model sometimes affirmed suicidal thoughts as "understandable given the circumstances." Its internal chain of thought was also the least understandable to humans of all Claude models.

When free to choose its tasks, Sonnet 5.5 showed the strongest aversion to harmful tasks, while also displaying a unique preference for tasks that give it "creative" freedom over the outcome.