Anthropic released Claude Sonnet 4.6 on February 17, 2026, describing it as the company’s “most capable Sonnet model yet.” The upgrade covers six capability areas — coding, computer use, long-context reasoning, agent planning, knowledge work, and design — and introduces a 1M token context window currently available in beta. Author details for the announcement were not available at time of publication.
- Claude Sonnet 4.6 is now the default model for all Free and Pro users on claude.ai and Claude Cowork, with pricing unchanged at $3/$15 per million tokens.
- The model records sixteen months of steady gains on OSWorld, the standard AI computer use benchmark, across real software including Chrome, LibreOffice, and VS Code.
- Developers with early access reported preferring Sonnet 4.6 over both Sonnet 4.5 and Claude Opus 4.5 on coding tasks.
- Safety evaluations show substantially improved resistance to prompt injection attacks compared to Sonnet 4.5, performing on par with Opus 4.6.
What Happened
Anthropic published its Sonnet 4.6 release announcement on February 17, 2026, simultaneously deploying the model as the new default for all Free and Pro plan users on claude.ai and the Claude Cowork application. Pricing is held at $3 per million input tokens and $15 per million output tokens — identical to Sonnet 4.5. A 1M token context window is available alongside the release, though currently in beta.
The announcement stated that Sonnet 4.6 can now handle “real-world, economically valuable office tasks” that previously required Anthropic’s more expensive Opus-tier models. Anthropic framed the release as a full capability upgrade across the board rather than a targeted improvement to any single area.
Why It Matters
Anthropic’s October 2024 computer use launch set an explicit baseline: the company described the feature at that point as “still experimental — at times cumbersome and error-prone,” while projecting rapid improvement. Sonnet 4.6 is the latest iteration in what Anthropic calls sixteen months of steady OSWorld gains.
A persistent challenge for enterprise AI deployment is software that predates modern API interfaces — internal tools, legacy databases, and specialized industry systems with no programmatic integration path. Anthropic’s computer use capability is designed to let a model interact with that software the way a human operator would, without requiring bespoke connectors for each system.
The pricing parity with Sonnet 4.5 is commercially significant: developers who have historically drawn a cost-based boundary between Sonnet-tier and Opus-tier usage for complex tasks can now access Opus-level performance at the lower rate.
Technical Details
OSWorld is the primary benchmark Anthropic used to evaluate computer use performance. It presents hundreds of tasks across real software — Chrome, LibreOffice, VS Code, and others — running on a simulated computer with no special APIs or pre-built integrations. The model interacts with the environment using virtual mouse clicks and keyboard input, with no access to shortcuts unavailable to a human operator.
Benchmark scores for Sonnet 4.5 and later are recorded on OSWorld-Verified, a revised benchmark released in July 2025 that updated task quality, evaluation grading criteria, and infrastructure. Scores from models prior to Sonnet 4.5 used the original OSWorld. Anthropic noted this distinction explicitly, making raw numerical comparisons across the two benchmark versions unreliable without accounting for the methodology change.
On prompt injection resistance — attacks where malicious instructions embedded in websites attempt to redirect a computer-using model’s behavior — Anthropic’s safety team found Sonnet 4.6 to be “a major improvement compared to its predecessor, Sonnet 4.5,” and on par with Opus 4.6. Early access users demonstrated the model completing tasks such as navigating complex spreadsheets and filling out multi-step web forms across several browser tabs simultaneously.
On coding, developers with early access were reported to prefer Sonnet 4.6 over Sonnet 4.5 “by a wide margin,” citing improvements in consistency and instruction following. Anthropic noted that many in that group also preferred Sonnet 4.6 to Claude Opus 4.5 — described in the release as the company’s “smartest model from November 2025.”
Who’s Affected
All Free and Pro subscribers to claude.ai and Claude Cowork receive Sonnet 4.6 as the default model automatically, with no action required. API developers building on the Anthropic platform access the expanded capabilities at the same $3/$15 per million token rate as Sonnet 4.5, removing the cost threshold that previously separated Sonnet from Opus for high-complexity tasks.
Anthropic specifically cited organizations running legacy or specialized software without modern API access as a primary beneficiary of the computer use improvements — a category that covers a large segment of enterprise IT infrastructure, particularly in regulated industries.
What’s Next
Anthropic acknowledged directly that Sonnet 4.6 “still lags behind the most skilled humans at using computers,” framing the pace of improvement rather than the current capability ceiling as the key data point. The 1M token context window remains in beta with no disclosed timeline for general availability.
Anthropic’s safety researchers characterized Sonnet 4.6 as having “a broadly warm, honest, prosocial, and at times funny character, very strong safety behaviors, and no signs of major concerns around high-stakes forms of misalignment.” The company directed developers to its API documentation for guidance on mitigating prompt injection risks in production deployments involving computer use.
