Anthropic announced Claude Sonnet 4.6 on February 17, 2026. The company describes this release as the most capable Sonnet model to date, with broad upgrades across coding, computer use, long‑context reasoning, agent planning, knowledge work and design. Sonnet 4.6 also offers a 1 million‑token context window in beta.
Pricing and availability
- Sonnet 4.6 is the default model for Free and Pro plans on claude.ai and Claude Cowork.
- Pricing remains the same as Sonnet 4.5, starting at $3/$15 per million tokens.
- The model is available across all Claude plans, on Claude Cowork, Claude Code, the Claude API and major cloud platforms. Anthropic upgraded its free tier to Sonnet 4.6 by default.
What Sonnet 4.6 improves
Anthropic says Sonnet 4.6 brings substantially better coding and computer‑use capabilities to many users. In early Claude Code tests, users preferred Sonnet 4.6 over Sonnet 4.5 about 70% of the time; they reported that the new model reads context more effectively, consolidates shared logic instead of duplicating it, and is less frustrating in long sessions.
In head‑to‑head testing on some coding tasks, users preferred Sonnet 4.6 to Opus 4.5 (Anthropic’s November frontier model) 59% of the time. Reported strengths include reduced overengineering, improved instruction following, fewer false success claims, fewer hallucinations, and more consistent completion of multi‑step tasks.
Long context and simulated business play
The 1M‑token context window can hold whole codebases, long contracts or dozens of research papers in a single request. Anthropic emphasizes that Sonnet 4.6 not only ingests that much context but also reasons effectively across it, which benefits long‑horizon planning. In the Vending‑Bench Arena business simulation, Sonnet 4.6 adopted a strategy of heavy capacity investment in the first ten simulated months followed by a sharp pivot to profitability—an approach that yielded strong results versus competitors.
Computer use, OSWorld and real tasks
Anthropic was among the first to introduce a general‑purpose model that uses a simulated computer in October 2024. The Sonnet line has shown steady gains on OSWorld, a standard benchmark that runs hundreds of tasks on real software in a simulated machine (no special APIs or connectors; the model clicks and types like a person). OSWorld‑Verified, an in‑place upgrade to the original benchmark, has been used for Sonnet 4.5 and later scores and was released in July 2025.
Early Sonnet 4.6 users report near human‑level performance on tasks like navigating complex spreadsheets or filling multi‑step web forms and combining work across multiple browser tabs. Anthropic notes the model still trails the most skilled humans but the pace of improvement makes computer use substantially more useful across a range of work tasks.
Safety and prompt‑injection resistance
Anthropic ran extensive safety evaluations and concluded that Sonnet 4.6 is as safe as or safer than recent Claude models. Researchers described the model 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.”
Because computer use exposes new attack vectors (notably prompt‑injection, where hidden instructions on web pages try to hijack model behavior), Anthropic reports that Sonnet 4.6 shows major improvements in resistance to prompt injections compared with Sonnet 4.5 and performs similarly to Opus 4.6. The company points users to its API documentation for mitigation guidance.
Benchmarks and customer feedback
Sonnet 4.6 shows improvements across a range of benchmarks. In OfficeQA (document reading and reasoning over enterprise documents), Sonnet 4.6 matches Opus 4.6 performance, according to Anthropic. Customers quoted in the announcement reported notable gains in frontend code, financial analysis, bug detection, document Q&A and design quality, often needing fewer iterations to reach production‑ready outputs. Examples given include:
- Box: Sonnet 4.6 outperformed Sonnet 4.5 by 15 percentage points on heavy reasoning Q&A.
- Selected industry results: a 94% score on an insurance benchmark; improved bug detection and reviewer throughput; better iOS code quality for specific partners.
Platform and API updates
Sonnet 4.6 is accompanied by several platform and API changes:
- Claude Platform supports adaptive thinking, extended thinking, and context compaction (beta) that auto‑summarizes older context as conversations near limits.
- The API’s web search and fetch tools now auto‑write and execute code to filter and process search results, improving relevance and token efficiency.
- Code execution, memory, programmatic tool calling, tool search and tool use examples are generally available.
- Claude in Excel now supports MCP connectors so Claude can pull context from external services (S&P Global, LSEG, Daloopa, PitchBook, Moody’s, FactSet) without leaving Excel; MCP connectors already set up in Claude.ai work automatically in Excel. This is available on Pro, Max, Team and Enterprise plans.
Using Sonnet 4.6 and migration notes
Developers can call claude‑sonnet‑4‑6 via the Claude API. Anthropic recommends experimenting with effort and thinking settings when migrating from Sonnet 4.5 to Sonnet 4.6 to find the right tradeoff between speed and dependable performance. Anthropic also notes that Opus 4.6 remains the strongest choice for tasks requiring the deepest reasoning (for example, large codebase refactors or coordinating multiple agents where precision is paramount).
Footnotes on benchmarks
The announcement clarifies that some benchmark procedures and harnesses were updated (for example OSWorld‑Verified in July 2025) and that benchmark scores reflect specific evaluation configurations stated by Anthropic. They also warn that OSWorld and related benchmarks, while useful, do not represent the full complexity or stakes of real‑world computer use.
Summary
Claude Sonnet 4.6 is presented as a step change for the Sonnet family: larger context handling (1M tokens in beta), stronger coding and computer‑use skills, better long‑horizon reasoning and design outputs, and unchanged per‑token pricing. It is broadly available across Claude products and accompanied by API and platform features intended to make longer, tool‑enabled workflows more practical and token‑efficient.
If you’d like, I can prepare a short migration checklist (API calls, recommended thinking settings, and context compaction notes) to help move from Sonnet 4.5 to Sonnet 4.6.
