Claude Sonnet 4.5

Overview. Claude Sonnet 4.5 is Anthropic's instruct-tuned mid-tier model in the Claude Sonnet 4 family, released 29 September 2025 and accessed through the Claude API. It sits between Anthropic's lighter Haiku models and its flagship Opus tier, and it's built specifically for agentic coding, computer-use, and long-context reasoning workflows. Developers reach for it when they need a model that can drive multi-step tool loops rather than just answer single-turn questions.

Key strengths. Sonnet 4.5's headline signal is agentic capability: it scored 61.4% on the OSWorld computer-use benchmark at launch, reflecting genuine competence at interpreting screenshots and operating software autonomously. It supports tool calling with multiple tool_use blocks per turn, native prompt caching, and a 200,000-token context window, which together suit sustained coding and document-heavy agent loops. Domain-wise, it's strongest on code and agentic tasks, with solid general writing and communication performance.

When to use it. Use Sonnet 4.5 for agentic coding assistants, computer-use automation, and tasks that require holding large amounts of context — long documents, multi-file codebases, or extended tool-use transcripts — within a single 200,000-token window. Its interactive latency profile and parallel tool-calling support make it a good fit for conversational agent loops rather than pure batch processing. Extended-thinking traces are available as an opt-in mode if a task benefits from visible step-by-step reasoning, but the model defaults to direct instruct-style responses.

Limitations. Context is capped at 200,000 tokens — the earlier 1M-token beta was retired — and output generation is capped at 64,000 tokens per response. The model accepts text and image input only; it does not process or generate audio, and Anthropic has not disclosed its parameter count, architecture, or training-data cutoff. Anthropic now also offers the newer Sonnet 4.6 within the same family, so teams starting new agentic-reasoning work may want to evaluate that variant alongside this one.

Sovereignty note. Claude Sonnet 4.5 is built by Anthropic, a US-based creator, and carries CLOUD Act exposure — relevant for teams with strict data-residency requirements. It's a closed-weights, proprietary model served over Anthropic's own API rather than deployed on Sovereign infrastructure.