Claude Sonnet 4.6

Overview. Claude Sonnet 4.6 is a proprietary hybrid-reasoning model built by Anthropic, released in February 2026. It pairs a 1M-token context window with up to 128K tokens of output and lets you dial reasoning effort up or down per request. Developers reach for it when a task needs either a large context window, an agentic tool-calling loop, or an adjustable depth of reasoning within a single model.

Key strengths. The model's hybrid-reasoning design exposes selectable effort levels (low through max), so you can trade latency for depth of thinking on a per-call basis rather than switching models. Its 1M-token context window supports long documents, multi-file codebases, and extended agentic sessions without chunking. Native tool calling, optional parallel tool calls, and prompt/context caching make it well suited to coding and research workflows that lean on repeated context or multi-step tool use.

When to use it. Use Claude Sonnet 4.6 for coding, research, or agentic tasks that benefit from extended reasoning traces and a large working context, such as reviewing an entire codebase, synthesizing long research documents, or running multi-step tool-calling agents. Set a lower reasoning effort for interactive, latency-sensitive requests, and a higher effort for tasks that need deeper deliberation before answering. Its vision input also makes it a fit for tasks that combine text and images, such as reviewing diagrams alongside code or documents.

Limitations. Anthropic now positions this model as a legacy Sonnet release: Claude Sonnet 5 has replaced it as the default Sonnet model as of June 2026, though Sonnet 4.6 remains active with no retirement scheduled before February 2027. It has no native audio input or output, and Anthropic does not disclose its underlying architecture or parameter count. Its training data cutoff is January 2026, so it won't have knowledge of events after that point.

Sovereignty note. Claude Sonnet 4.6 is built by Anthropic, a US-based creator, which means Inference through this model can be subject to CLOUD Act access requests rather than qualifying as Sovereign under EU-residency terms.