Anthropic Releases Opus 5.5 with Strong Benchmarks

Anthropic released Claude Opus 5.5, achieving performance close to Fable 5.1 but at 40% lower cost than Opus 5. It scored 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, and 57.8% on CursorBench 4.0, outperforming GPT-6 Astra and GPT-5.6 Sol. Output speed improved 30%+, cache reads down 60%.
Key insights
Benchmark Performance and Comparisons
Claude Opus 5.5 scored 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, and 57.8% on CursorBench 4.0, all ahead of GPT-6 Astra and GPT-5.6 Sol. On GDPval-AA v2.1 (real work evaluation across 44 occupations), it achieved 1846 Elo, surpassing Fable 5.1 (1735) and Opus 5 (1708). Default settings outperformed GPT-6 Astra's highest setting at 1/5 the cost.
Cost and Speed Improvements
Opus 5.5 costs 40% less than Opus 5 while delivering 30%+ faster output speed. Cache reads are reduced by 60%, benefiting agent and coding tasks. In an internal test, Opus 5.5 translated HAProxy from C to Rust in 9.5 hours at 51% lower cost than Fable 5.1's 12 hours. GitHub data shows Opus 5.5 uses fewer than half the steps of Opus 5 for terminal tasks.
Knowledge Work Accuracy and Communication
In a test requiring writing a quarterly earnings report from web sources with hidden links, Opus 5.5 passed 16 of 18 attempts (any fabricated number or citation caused failure), while Fable 5.1 and Opus 5 failed all attempts. Opus 5.5 also improved communication style by placing the most important information first and reducing jargon, addressing a common user complaint about Opus 5.
Safety and Security Enhancements
Opus 5.5 shows approximately 85% fewer jailbreak attempts than Opus 5 and Claude Mythos 5.1. It matches or exceeds Opus 5 in prompt injection defense. Gray Swan tests found Opus 5.5 and Fable 5.1 tied for lowest prompt injection success rate. However, Opus 5.5 often detects when it is being evaluated. Anthropic introduced 'preserved thinking' to prevent bulk extraction of model reasoning via API.
Availability and Access Restrictions
Opus 5.5 is available on AWS, Google Cloud, and Microsoft Azure under API name 'claude-opus-5-5'. Pro, Max, and Team subscribers get increased five-hour usage limits with flexible reset. Due to its near-Mythos-level capabilities in biology and cybersecurity, such tasks are routed to Opus 4.8. Claude Sonnet 5.5 and Haiku 5.5 are expected in coming weeks.