Anthropic: Claude Opus 5.5

Claude Opus 5.5 is Anthropic's flagship Opus-class large language model, released on 22 September 2026 as the recommended starting point for demanding API workloads. It extends the Claude 5.5 generation with adaptive extended thinking, a one-million-token context window, up to 128K tokens of output per request, and multimodal input (text and images). Anthropic positions it as a broad upgrade over Claude Opus 5, with especially strong gains in agentic software engineering, terminal and computer-use tasks, long-horizon professional work, and visual reasoning, while costing roughly 40% less than Opus 5 on typical token-billed workloads at $4 per million input tokens and $20 per million output tokens.

On Anthropic's published capability suite, Opus 5.5 leads or matches frontier peers on SWE-bench Pro (89.9%), SWE-bench Multilingual (93.9%), DeepSWE v1.1 (74.2%), Humanity's Last Exam with tools (67.7%), Toolathlon Verified (77.8% pass@1), GDPval-AA v2.1 (1846 Elo), and AutomationBench (40.0%). Independent leaderboards report competitive scores on GPQA Diamond, MMLU-Pro, LiveBench, SimpleBench, Agents' Last Exam (38.2% pass rate), and creative-writing Elo ratings. The model ships with domain safeguards in biology, cybersecurity, and frontier-AI development that can fall back to older Claude snapshots when classifiers trigger.

Claude Opus 5.5 is available only through hosted APIs and cloud platforms (Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry) under the model ID claude-opus-5-5. Parameter count and training architecture details are not disclosed; inference runs on Anthropic's infrastructure rather than on customer hardware. Default reasoning effort is medium, with higher effort settings available for maximum performance on difficult agentic and reasoning benchmarks.