Anthropic Ships Claude Sonnet 4.6 With 1M-Token Context in Beta

Anthropic introduced Claude Sonnet 4.6 on Feb 17, 2026 and described it as its most capable Sonnet model yet. The company says the upgrade improves coding, computer use, long-context reasoning, agent planning, knowledge work, and design, while also adding a 1M token context window in beta. Sonnet 4.6 becomes the default model on Free and Pro plans, in Claude Cowork, and is available through the API and major cloud platforms.

A notable part of the release is that Anthropic kept pricing unchanged from Sonnet 4.5, starting at $3/$15 per million tokens. That matters because Sonnet has become Anthropic’s price-performance workhorse for developer and enterprise use. By attaching stronger capabilities to the same price tier, Anthropic is clearly trying to widen usage rather than reserve the gains for a premium Opus-only lane.

What Changed

• 1M token context window in beta.
• Adaptive thinking and extended thinking on the Claude Platform.
• Context compaction in beta to summarize older turns near context limits.
• General availability for code execution, memory, programmatic tool calling, tool search, and tool use examples.

Anthropic also included customer signals meant to show that the model’s gains translate into real workflows. According to the company, Box saw a 15 percentage point improvement over Sonnet 4.5 in heavy reasoning Q&A, Pace reported 94% on its insurance benchmark, and early users preferred Sonnet 4.6 to Opus 4.5 59% of the time. The company further said web search and fetch can now automatically write and execute code to filter and process results, a practical change for agent-style applications that need tighter context control.

The broader implication is not just a faster Claude release cycle. Sonnet 4.6 strengthens the middle of the market where teams want frontier-adjacent performance without frontier-class cost. If Anthropic’s claims hold in production, the upgrade could make Sonnet a more attractive default for coding agents, browser automation, and long-running task orchestration where reliability and tool use matter as much as raw reasoning depth.