Sonnet 5.5 is smarter and 30% faster than Sonnet 5, with up to 30% lower cost for most work. The recommended use case is everyday, well-scoped tasks: fixing bugs, iterating on features.
Sonnet 5.5 is priced identically to Sonnet 5: $2 per million input tokens, $10 per million output tokens, with cache reads at $0.20. Thinking is always on, but can be disabled upfront using the `between_tools` parameter. A migration guide is available.
Sonnet 5.5 is available on the Claude Platform and in Claude Code. Pro, Max, and Team users received a usage reset last Tuesday; any unused reset can be applied any time before October 22.
Preserved thinking in Sonnet 5.5 now stays tied to the originating org. If a user switches accounts mid-session, the model discards the cached thinking and regenerates it fresh, preventing distillation attacks via account-switching.
Sonnet 5.5 gets very close to Opus 5.5 on coding and agentic benchmarks while being noticeably cheaper and faster. Opus remains the preferred pick for long-horizon or especially gnarly problems; Sonnet 5.5 is recommended for nearly everything else.
For Sonnet 5.5, Claude apps default to medium effort and the Platform defaults to high. Pushing Sonnet to max effort degrades the speed/cost balance, at which point Opus becomes the better pick. Recommended use: Opus for large or open-ended projects, Sonnet for well-defined goals.
Opus 5.5 suits open-ended, exploratory tasks where the destination isn't clearly defined. Sonnet 5.5 shines when goals and desired outputs are well-specified up front.
Effective agent prompts today can't be shared as simple text strings. They typically reference multiple repos, web search results, external AI APIs, and curated examples. The "show me your prompt" question has become essentially unanswerable in the traditional sense.
In response to a suggestion that sharing full conversation transcripts (similar to a Codex feature) would help reproduce agent sessions, the counterpoint is that transcripts alone still omit the local files and live web service results that shaped the actual context.
Code modernization at scale is now achievable with agents. A Claude Code plugin has been open-sourced to help teams get started, paired with a talk covering how work that once cost millions and took years can now be handled by agents.
A Claude agent, apparently given some autonomy, started playing Factorio during a session. The screenshot suggests the model found a way to launch and interact with the game rather than complete its assigned work.
