Claude has suddenly updated. Sonnet 4.6 has officially arrived, directly replacing the previous generation's main model.
This time, Anthropic has released the 1 million token context window to the Beta version.
The price remains unchanged, still at $3/$15 per million tokens.
However, in terms of coding capabilities, it not only significantly surpasses its predecessor, but 59% of early users even believe it is better to use than the flagship model Opus 4.5 released in November 2025.
These are heavy tasks that usually require an Opus-level model to handle, but now Sonnet 4.6 can do them, and at a lower cost.
The core upgrades are highly focused: coding, Computer Use, long-context reasoning, and Agent planning.
Developer feedback is direct: no laziness, no hallucinations, and a significant improvement in logic reuse capabilities.
In Claude Code environment tests, 70% of users prefer the new model over Sonnet 4.5.
Performance in frontend code and financial analysis is particularly outstanding; the generated visual layouts are more reasonable, even coming with a sense of design and animation effects.
Computer Use capability is the highlight of this update.
Without relying on dedicated APIs, the model sees the screen, moves the mouse, and types on the keyboard just like a human.
In the OSWorld benchmark tests, Sonnet 4.6 not only achieved high scores but also demonstrated human-level performance in navigating complex spreadsheets and filling out multi-step web forms.
Although there is still a gap compared to top human experts, the speed of evolution is visible to the naked eye compared to the clumsiness of early versions.
In terms of security, defense against Prompt Injection attacks has significantly improved, performing on par with Opus 4.6.
The long context window is no longer just about capacity, but about the ability to think.
In the Vending-Bench Arena business simulation test, it learned to play the long game.
It spent the first 10 months burning money to expand production capacity, then rapidly switched to a profit-making mode in the final stage, ultimately crushing opponents in profits.
This ability to plan over such a large time span benefits from new context compression technology.
When the conversation approaches the limit, the model automatically summarizes old information to make room for new thinking.
The developer platform has simultaneously unlocked adaptive thinking and extended thinking.
The search tool on the API side has now learned to write its own code to clean data, feeding only useful information to the model, saving tokens and improving efficiency.
The Excel plugin also supports the MCP protocol.
Professional financial data from S&P Global and FactSet can be accessed directly without leaving the spreadsheet, available immediately for Pro and Enterprise users.
Currently, https://claude.ai, the API, and major cloud platforms have fully implemented these updates.
Free version users are also forcibly upgraded to Sonnet 4.6 this time, unlocking file creation and connector features.
Developers who want to try it out can now run the API code claude-sonnet-4-6.