Claude Opus 4.6 Unveiled

Claude Opus 4.6 has been unveiled as the latest model with greatly improved coding performance. It plans more sophisticated approaches, sustains agentic work for longer periods, operates more reliably on large codebases, and has strengthened code review and debugging capabilities. It is the first Opus-class model to offer a 1M token context window in beta.

Its performance also leads across multiple benchmarks. It achieved the top score on Terminal-Bench 2.0 and also led on Humanity's Last Exam. On GDPval-AA, it outperformed OpenAI's GPT-5.2 by about 144 Elo, and exceeded its predecessor Claude Opus 4.5 by 190 Elo. On BrowseComp as well, it produced better results than any other model.

For long-form work, it tracks more documents over longer stretches without losing key information. Anthropic stated that on the 8-needle 1M variant of MRCR v2, Opus 4.6 scored 76% while Sonnet 4.5 scored 18.5%. This demonstrates improvements in both information retrieval over long contexts and the reasoning that follows it.

Safety has also been strengthened. In automated behavioral audits, rates of misaligned behavior—such as deception, sycophancy, encouraging user delusion, and cooperating with misuse—were low, and it showed the lowest over-refusal rate among recent Claude models. The new version applies a broader range of evaluations, including user wellbeing, refusal of dangerous requests, and detection of covertly harmful actions.

Products and the API have also been updated together. Claude Code has added agent teams for running multiple agents in parallel, and the API now includes compaction, adaptive thinking, and effort controls. Claude in Excel has been substantially improved, and Claude in PowerPoint has been released as a research preview. Opus 4.6 is immediately available on claude.ai, the API, and major cloud platforms, with pricing maintained at $5/$25 per million tokens.