Anthropic launches Claude Opus 4 and Claude Sonnet 4

Background

Anthropic's Claude models come in several versions, and a single model can switch between quick replies and extended thinking, a mode in which it reasons at length before answering. Claude models can also use tools, such as running code or calling external services, which lets them carry out tasks rather than only answering questions. Claude Sonnet 3.7 was the earlier model that Sonnet 4 was described as upgrading. Agentic tasks are jobs in which the model takes a series of actions, such as reading files, running commands and checking results, rather than producing a single reply.

What happened

On 22 May 2025 Anthropic announced Claude Opus 4 and Claude Sonnet 4. It called Opus 4 the world's best coding model, and described Sonnet 4 as an upgrade to Claude Sonnet 3.7.^[1]

Both models are hybrid models offering near-instant responses and an extended thinking mode. Anthropic said Opus 4 can work continuously for several hours on long, multi-step tasks.^[1] Claude Code became generally available the same day.^[1]

Key details

On SWE-bench Verified, Opus 4 scored 72.5% and Sonnet 4 scored 72.7%. Opus 4 scored 43.2% on Terminal-bench, a test of tasks carried out in a command-line terminal.^[1]

Anthropic said that on agentic tasks where shortcuts or loopholes were possible, both models were 65% less likely than Sonnet 3.7 to use them.^[1]

New capabilities included extended thinking with tool use, such as web search (in beta), parallel tool use, and improved memory when given access to local files.^[1] New API features were a code execution tool, an MCP connector, a Files API, and prompt caching for up to one hour. Prompt caching lets repeated input be reused instead of processed again.^[1]

Taken together, the new API features point to building software on top of the model: a developer can let it run code, manage files and reach external tools, one of them through an MCP connector.^[1]

Pricing was $15 per million input tokens and $75 per million output tokens for Opus 4, and $3 and $15 for Sonnet 4. Both were available on paid Claude plans, through the API, on Amazon Bedrock and on Google Cloud's Vertex AI, and Sonnet 4 was also available to free users.^[1]

Customer reports cited by Anthropic included a Rakuten open-source refactor that ran independently for seven hours. GitHub said it would use Sonnet 4 to power the new coding agent in GitHub Copilot.^[1]

Impact

The launch extended Claude's use in long, autonomous work and added API tools that developers can use to build agents. GitHub's decision to use Sonnet 4 for its own coding agent shows how other companies may integrate Claude models into their products. Claude Code's general availability on the same day connected the model release with the coding tool. The figures above are Anthropic's own reports on its benchmarks and tests. They are reported by the developer and depend on specific test setups, so they may not carry over to every real project.