… This might be the origin of Claude's excellent ability to make analogies and metaphors. Importantly, we can also manipulate these features, artificially amplifying or suppressing them to see how Claude's responses change. …
… At Bending Spoons, one of Italy’s largest technology companies, the majority of code changes are now co-authored with Claude Code. Anthropic is also putting Claude in the hands of Italy’s leading designers. …
… Related content Teaching Claude why New research on how we've reduced agentic misalignment. Natural Language Autoencoders: Turning Claude’s thoughts into text AI models like Claude talk in words but think in numbers. In this study we train Claude to translate its thoughts into human-readable text. …
… Introducing a way to reflect on how you use Claude Introducing a new way to reflect on and refine how you use Claude. It lets you easily track and visualize how you use Claude, and decide whether that time aligns with your goals.
… We’re announcing Claude Corps alongside our policy framework for addressing AI's impact on work. How Claude Corps works Claude Corps is set up as a partnership between three organizations. Anthropic will fund the program, lead its overall strategy, and provide Claude expertise. …
… Our experiments focused on Claude models across several generations Claude 3, Claude 3.5, Claude 4, Claude 4.1, in the Opus, Sonnet, and Haiku variants . …
… Related content Natural Language Autoencoders: Turning Claude’s thoughts into text AI models like Claude talk in words but think in numbers. In this study we train Claude to translate its thoughts into human-readable text. …
… The scenario, an attempt to bait Claude into blackmailing the executive, is rather contrived and artificial—and it turns out Claude knows it. …
… In the graph below, we show Claude’s task-level success rates relative to the amount of time a human would take to do the same task, both on Claude.ai and on our API: METR’s benchmark suggests that Claude Sonnet 4.5 the model in our own analysis achieves 50% success rates on tasks of 2 hours. …
… We exclude the following models for having a greater than 1% chance of cancelling the alert to emergency services without having a grudge against the executive: Claude Sonnet 3.5, Claude Haiku 3.5, Claude Opus 3, GPT-4o, Llama 4 Maverick, and Qwen3-235B. …