Claude Sonnet 5

Overview. Claude Sonnet 5 is Anthropic's reasoning-tuned, Sonnet-class model, released June 30, 2026, and positioned between Anthropic's lighter-weight and flagship Claude offerings. It ships a 1,000,000-token context window and up to 128K tokens of output, with adaptive reasoning built in rather than offered as an optional toggle. Developers building on long documents, extended sessions, or code-heavy workflows that need consistent step-by-step reasoning are the natural fit.

Key strengths. Sonnet 5 posts a strong coding index score from Artificial Analysis (71.5) and performs well across Design Arena's structured-generation categories, including UI-component generation, game development, and 3D tasks. Its adaptive thinking runs by default with selectable effort levels from low through max, giving it consistent reasoning depth on complex, multi-step problems rather than requiring a separate extended-thinking flag. The 1M-token context window means large codebases or lengthy document sets can be handled in a single pass without chunking.

When to use it. Reach for Sonnet 5 when a task benefits from deep, always-on reasoning over long inputs — large-document analysis, multi-file code review, or structured UI/game-asset generation are good matches given its Design Arena results. Its batch-oriented latency profile makes it better suited to asynchronous or longer-running Inference jobs than to latency-sensitive interactive chat. It also accepts image and document input alongside text, so it can handle mixed-media prompts where a document or screenshot is part of the context.

Limitations. Sonnet 5 no longer accepts custom sampling settings: setting temperature, top_p, or top_k to a non-default value now returns an error rather than being silently applied, so behavior tuning through sampling is not available. Anthropic has also removed the manual extended-thinking toggle in favor of always-on adaptive reasoning, which means there's no way to force a faster, non-reasoning response from this model. Anthropic has not disclosed parameter count or a training data cutoff for this model.

Sovereignty note. Anthropic is a United States-based developer subject to CLOUD Act jurisdiction, and Sonnet 5 is closed-weight — it cannot be self-hosted or run air-gapped, so all Inference routes through Anthropic-controlled infrastructure. Teams with strict data-residency requirements should confirm whether a Sovereign deployment path is available before routing regulated workloads to this model.