Claude Sonnet 4.6 arrives with a 1M-token context window

Anthropic has introduced Claude Sonnet 4.6, an update that brings several capabilities previously reserved for the company’s more expensive models to a faster, more affordable version. The model improves at coding, document analysis, computer use and multi-step tasks.

Sonnet 4.6 is now the default model for Free and Pro users in Claude and Claude Cowork. It is also available in Claude Code, the API, major cloud services and all Claude plans.

Pricing remains unchanged from Sonnet 4.5: $3 per million input tokens and $15 per million output tokens. Tokens are the units of text the model processes, so the final cost depends on how much context it receives and how much content it generates.

More capacity for working with large volumes of information

The most visible change is a 1 million token context window in beta. In practice, this lets you bring an entire codebase, lengthy contracts or dozens of articles and research papers into a single conversation.

More context alone is not enough. The model also needs to find the important connections across all that material. Anthropic says Sonnet 4.6 improves at long-term planning and tasks that require following instructions over many steps.

In a simulated business management evaluation, the model invested heavily during the first ten months to increase its capacity, then changed strategy to focus on profitability. The timing of that shift allowed it to finish ahead of its competitors.

The computer as a work tool

Sonnet 4.6 also improves at what is known as computer use: it can see a screen, click, type and navigate software much like a person would, without needing a dedicated connection for each application.

This is useful for older or specialized software that does not offer an API, meaning a direct way for other programs to connect to it. Examples include:

• Completing multi-step web forms.
• Navigating complex spreadsheets.
• Working across documents spread over multiple tabs.
• Using business tools that were not designed to integrate with AI systems.

In the OSWorld test, which evaluates tasks in Chrome, LibreOffice, VS Code and other applications, Anthropic says its Sonnet models have improved steadily over the past 16 months. The company also warns that the model still falls short of the most expert people and that benchmark results do not represent everything that happens in a real-world environment.

There is an important risk: websites can hide malicious instructions designed to manipulate the model. This attack is known as prompt injection. Anthropic says Sonnet 4.6 resists these attempts better than Sonnet 4.5 and delivers performance similar to Opus 4.6 in its evaluations.

Coding and documents: the most practical improvement

According to Anthropic’s internal tests, Claude Code users preferred Sonnet 4.6 over Sonnet 4.5 approximately 70% of the time. They also chose it over Claude Opus 4.5 in 59% of cases.

Users said the new model does a better job of reviewing context before modifying a project, duplicates less code and follows instructions more consistently. They also reported fewer false claims that a task was complete, fewer invented errors and better tracking of lengthy processes.

For business documents, Sonnet 4.6 matches Opus 4.6’s performance on OfficeQA, a test that measures whether AI can read charts, PDFs and tables, locate specific data and draw correct conclusions. Anthropic also reports a 15 percentage point improvement over Sonnet 4.5 in an evaluation of complex questions about business documents conducted by Box.

The company also says the model reached 94% on its insurance test, the best result it has achieved in that computer-use evaluation. These results were reported by Anthropic and its collaborators, not guaranteed accuracy for every company or use case.

What changes for you

If you use Claude for free, you get a more capable model without paying more. The free tier now includes file creation, connectors, skills and compaction, a feature that summarizes older parts of a conversation to free up space and keep work active for longer.

For developers, the most relevant change is the balance between capability and cost. Tasks such as fixing errors in large codebases, analyzing contracts or coordinating processes across multiple tools can run on Sonnet 4.6 without always relying on an Opus model, although Anthropic continues to position Opus 4.6 as its option for problems that require the deepest reasoning.

The API adds adaptive and extended thinking, as well as compaction in beta. Its web search tools can also run code to filter results and retain only relevant information, which may reduce token usage.

Sonnet 4.6 does not eliminate the need to supervise an AI that operates software or makes decisions about documents. What changes is that these tasks are becoming reliable and affordable enough to test across more teams. The next thing to watch is whether that improvement holds up beyond controlled evaluations, in real workflows with real consequences.