Claude Opus 5.5

Claude Opus 5.5, released September 22, 2026, succeeded Claude Opus 5 at Anthropic's Opus tier eight and a half weeks after it and three weeks after Claude Fable 5.1 took the top of the lineup. It arrived with a price cut: $4 per million input tokens and $20 per million output against Opus 5's $5 and $25, with cache reads at $0.20 per million, less than half Opus 5's $0.50, and cache writes at $5. Its launch table did something no earlier Opus release had done: it put the Opus model ahead of the Fable model above it on every board the two shared. At launch it scored 66.4% on Terminal-Bench 4.0 against Fable 5.1's 55.8% and Opus 5's 52.3% under Anthropic's own harness, 54.4% on the main split of FrontierCode v1.1, a GDPval-AA v2.1 Elo of 1846 for knowledge work where Fable 5.1 sat at 1735, and 67.7% on Humanity's Last Exam with tools. All of those were the best published scores at the time.

The rest of the card was agentic breadth. Opus 5.5 reached 81.8% partial completion on OSWorld 2.0 computer use, 89.0% on Chartography with tools and 58.7% on Terminal-Bench-Science 0.1, double Opus 5's 29.0% and six points behind GPT-6 Astra's 64.6%. On AutomationBench, run by Zapier during early access, it scored 40.0%, a point and a half short of GPT-6 Astra's 41.4% and up from Opus 5's 26.9%. Anthropic ran the table with production safeguards enabled, so cybersecurity tasks the classifiers intercepted were completed by Claude Opus 4.8 and biology and frontier-LLM-development tasks by Claude Opus 5, which it said likely lowered the reported numbers. The Terminal-Bench 4.0 result was reported at xhigh effort with a standard error of ±2.6 points; every other Claude figure used adaptive thinking at max effort.