Anthropic rolls out Claude Sonnet 4.6 at same prices

Anthropic has launched Claude Sonnet 4.6, updating its Sonnet line with improvements to coding, computer use, long-context reasoning, agent planning, knowledge work, and design. The model includes a 1 million token context window in beta and is now the default for Free and Pro users in Claude.ai and Claude Cowork.

Pricing remains unchanged from Sonnet 4.5, starting at USD $3 per million input tokens and USD $15 per million output tokens. As a result, more Claude users can access the latest Sonnet model without upgrading.

The release reflects a broader industry shift in which AI vendors bring more advanced models to mainstream tiers while competing on cost and responsiveness. Anthropic is positioning Sonnet 4.6 as close to its Opus line on several tasks while keeping Sonnet pricing the same.

Coding focus

Anthropic says Sonnet 4.6 improves consistency and instruction following for software development. It also reports stronger performance in longer sessions, with fewer incorrect assertions and better completion of multi-step work.

In Claude Code testing, users preferred Sonnet 4.6 over Sonnet 4.5 about 70% of the time, according to Anthropic. Testers said the new model read surrounding context more effectively before making changes and reduced duplication by consolidating shared logic.

In the same setting, users preferred Sonnet 4.6 over Claude Opus 4.5 59% of the time, Anthropic reports. Testers rated Sonnet 4.6 as less prone to overengineering and "laziness," with better instruction following and fewer hallucinations.

Computer use

A major theme of the update is progress on "computer use," in which an AI model operates software by interacting with an on-screen environment. This approach is aimed at organisations that rely on older tools and bespoke systems without modern integration interfaces, where automation often requires custom connectors.

Anthropic introduced what it called a general-purpose computer-using model in 2024, but described the first version as experimental and sometimes error-prone.

It pointed to OSWorld, a computer-use benchmark that covers tasks across applications including Chrome, LibreOffice, and VS Code in a simulated environment. OSWorld does not use special APIs or purpose-built connectors; instead, the model clicks and types in the interface.

Anthropic says its Sonnet models have made steady gains on OSWorld over 16 months. Early users of Sonnet 4.6 also reported human-level performance on tasks such as navigating complex spreadsheets and completing multi-step web forms across multiple browser tabs.

Anthropic notes that the model still trails the most skilled humans at using computers. It also highlighted security risks tied to computer use, including prompt-injection attacks in which malicious instructions are hidden on websites.

Safety checks

Anthropic says safety evaluations found Sonnet 4.6 to be as safe as, or safer than, other recent Claude models. Its safety team reported no major concerns about high-stakes misalignment.

"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," said Anthropic researchers.

On prompt injection, Anthropic says Sonnet 4.6 is a major improvement over Sonnet 4.5, with performance similar to Opus 4.6.

Long context

The 1 million token context window, currently in beta, is large enough to include entire codebases, long contracts, or dozens of research papers in a single request. Anthropic also says Sonnet 4.6 can reason effectively across the expanded context, linking it to longer-horizon planning.

Anthropic cited results from a Vending-Bench Arena evaluation that simulates running a business over time and pits models against each other. It said Sonnet 4.6 invested heavily in capacity for 10 simulated months before shifting focus to profitability later, helping it finish ahead of competitors.

Product updates

On the Claude Developer Platform, Sonnet 4.6 supports adaptive thinking and extended thinking. It also supports context compaction in beta, which summarises older context as a conversation approaches its limits.

Anthropic also updated Claude's web search and fetch tools in its API. The tools now write and execute code to filter and process search results, aiming to keep only relevant content in the active context and reduce token usage. It added that code execution, memory, programmatic tool calling, tool search, and tool-use examples are now generally available.

For Claude in Excel, Anthropic says the add-in now supports MCP connectors. It listed services including S&P Global, LSEG, Daloopa, PitchBook, Moody's, and FactSet, and said existing MCP connector setups in Claude.ai will work in Excel automatically for eligible plans.

Sonnet 4.6 is available on all Claude plans, Claude Cowork, Claude Code, the API, and major cloud platforms, Anthropic says. It added that the free tier now uses Sonnet 4.6 by default and includes file creation, connectors, skills, and compaction.