Sonnet 5.5 is now available. The official guide covers how to choose between Sonnet 5.5 and Opus 5.5, how to migrate from Sonnet 5 and tune effort settings, and how to use the new model inside Claude Code.
In Claude Projects, threads are full sessions that read code and plan their own work independently. The main chat's role is mostly routing tasks to threads and relaying results, not doing the strategic thinking itself.
Local sessions already work with Claude Projects threads. Users can select "Work locally" in the plus menu or simply ask Claude to do something locally.
Projects with Code support is already live on Android. If the feature is enabled for your account, it appears under the Code section in the app.
An experimental ahead-of-time compilation build of Bun with JavaScriptCore delivers meaningful gains for Claude Code's CLI: 24% faster startup, 48% lower memory, 46% lower CPU usage, at the cost of nearly doubling binary size.
A key insight from a Latent Space podcast appearance: smarter models require less human verification of their outputs, making them pareto dominant over weaker models even when they cost more. The episode covers harness engineering, frontier pacing, evals, and more.
Almost all evals are effectively saturated when you inspect the transcripts. Models nearly always identify the correct solution but narrowly miss it. A practical tip: prompt Claude to keep implementation notes so you can see its reasoning and decisions.
Eric Liu is joining Anthropic's product team after nearly five years at Notion, working alongside Cat Wu on Claude.
