Claude Sonnet 4 expands its context window to 1 million tokens

Anthropic has expanded Claude Sonnet 4's context limit to 1 million tokens, five times more than before. In practice, the model can receive a codebase of more than 75,000 lines or dozens of research papers in a single request.

Tokens are the units a model uses to process text. They do not correspond exactly to words, but they help measure how much information it can keep in view during a conversation or task. The larger the context, the more material it can analyze without forcing you to split it into parts.

What you can do with such a large context window

The feature is mainly designed for tasks that previously required you to break information into pieces:

• Analyze complete software projects, including source files, tests and documentation. Claude can connect dependencies across files and suggest changes while taking the overall architecture into account.
• Review large collections of documents, such as contracts, scientific papers or technical specifications, and compare relationships across hundreds of files.
• Build more persistent agents, capable of retaining API documentation, tool definitions and history across hundreds of calls and multi-step processes.

Here is a concrete example: instead of pasting only the file where an error appears, a team can upload an entire project and have Claude look for the source of the problem across the code, tests and documentation. That does not guarantee that all of its conclusions will be correct, but it gives the model much more information to work with.

The feature is available in public beta on the Claude developer platform for Tier 4 customers, with customized usage limits. It is also coming to Amazon Bedrock and Google Cloud Vertex AI, with broader availability expected in the following weeks.

The price goes up for large requests

The 1 million token limit does not come at the same price as standard requests. Anthropic charges the following:

• Requests of up to 200,000 tokens: $3 per million input tokens and $15 per million generated tokens.
• Requests of more than 200,000 tokens: $6 per million input tokens and $22.50 per million generated tokens.

Temporary instruction storage through prompt caching can reduce latency and cost when you reuse the same context. Batch processing also offers an additional 50% saving.

Companies such as Bolt.new and iGent AI are already testing this capability in coding tools. Its value is not just in reading more text, but in enabling longer work sessions on real projects without losing track of the thread so quickly.

For you, the change matters if you work with code, contracts, research or automated workflows that accumulate a lot of information. For simpler tasks, the previous limit was probably enough. The next step will be seeing how much of that context Claude can use reliably, rather than just how much it can theoretically accept.