Anthropic launches Claude Sonnet 4.5, its most advanced coding model so far

Anthropic has launched Claude Sonnet 4.5, its most advanced model to date for coding and agent development, less than two months after the release of Claude Opus 4.1. The company positions this new model at the top of current coding benchmarks and emphasizes improved reliability, describing it as capable of supporting applications beyond prototypes.

On developer benchmarks, Sonnet 4.5 scored 82% on SWE-bench Verified, the highest reported to date, and 50% on Terminal-Bench, surpassing OpenAI GPT-5 and Google Gemini 2.5 Pro. On OSWorld, it reached 61.4%, up from 42.2% with Sonnet 4. Internal tests also showed more than 30 hours of continuous agent-mode operation. Despite these results, the model still trails GPT-5 and Gemini 2.5 Pro in broader evaluations such as GPQA Diamond, MMMLU, and MMMU.

Alongside the release, Anthropic is rolling out the Claude Agent SDK, giving developers access to the same infrastructure behind Claude Code to build custom agents, and introducing “Imagine with Claude,” a research preview for Max subscribers that demonstrates real-time software generation without preset code. Claude Sonnet 4.5 is available through the Claude API at the same pricing, and has also been added to Amazon Bedrock and GitHub Copilot for business tiers, with Visual Studio Code users able to connect using their own API key.