'Claude Opus 4' is the top model that Anthropic evaluated as the "best of all time", and surpassed the competing models in coding performance, recording 72.5% in the SWE-bench benchmark and 43.2% in Terminal-bench. It maintains continuous performance even in long-term tasks with thousands of steps, and shows its strength in allowing AI agents to autonomously process complex tasks. For example, a case was disclosed where Opus 4 directly generated a 'navigation guide' during a Pokemon game play.
Meanwhile, 'Claude Sonnet 4' is an upgraded version of the existing Sonnet 3.7, and focuses on practicality and cost efficiency, recording 72.7% on SWE-bench. It has been improved in instruction execution ability, code control, and problem-solving ability, and is evaluated as a model suitable for various real-use environments, such as being adopted as the next-generation agent model of GitHub Copilot.
This Claude 4 series introduces the 'Extended Thinking' function, which allows external tools such as web searches to be used in parallel during inference, and also includes functions such as tool parallel execution, memory improvement, and prompt cache. The probability that the model will rely on expedient solutions has been reduced by 65% compared to the previous version, and a function for summarizing the internal thinking process has also been newly introduced.
The coding assistance tool, 'Claude Code', has also been officially released. It supports integration with VS Code and JetBrains, and allows automation and background tasks linked to GitHub Actions. Developers can create their own AI agents using the provided SDK, and can automate review feedback responses and error corrections by linking with GitHub.
