Claude Opus 5.5 Launch and Capability Demos

Anthropic released Claude Opus 5.5, which matches its Mythos-grade Fable 5.1 on most work tasks while costing 40% less. The model is 30% faster in output, uses fewer tokens, and has a cache read price drop from $0.50 to $0.20 per million tokens. It is available on Claude platforms, AWS, Google Cloud, and Azure.
Key insights
Performance and Cost Improvements
Claude Opus 5.5 achieves comparable performance to Fable 5.1 on most benchmarks while costing 40% less than Opus 5. Input/output prices are $4 and $20 per million tokens, 20% lower than Opus 5. The biggest saving is cache reads, dropping from $0.50 to $0.20 per million tokens, which benefits heavy agent and coding workloads. Output speed increased by over 30%. An internal test showed Opus 5.5 rewrote HAProxy from C to Rust in 9.5 hours at 51% lower cost than Fable 5.1's 12 hours.
Capability Demos: LEGO Design and Code Audit
HuggingFace team member Victor Mustar used Opus 5.5 to design a Microduck robot from scratch using 1,113 real LEGO pieces in 6 colors. The model generated a 141-page building manual with 237 steps and 13 chapters, plus LDraw files, a CSV part list, and a BrickLink XML order. All 3,204 connections were collision-free, and the center of mass was validated. Separately, an early tester used Opus 5.5 to audit and fix a 200,000-line codebase in under 3 hours, compared to 20 hours with Opus 5.
Factual Accuracy and Safety Measures
In a test where models wrote quarterly earnings reports using only web search, Opus 5.5 produced 16 of 18 reports with no fabricated numbers or quotes. Fable 5.1 and Opus 5 failed all. The model also uses improved writing style, placing conclusions first. However, due to capabilities near Mythos 5.1 in cybersecurity and biology, safety restrictions from Fable 5.1 apply: most cybersecurity tasks are routed to older Opus 4.8, and biological research requires a Life Sciences Verification Program.
Developer and API Considerations
Opus 5.5's thinking mode cannot be disabled. API accounts registered after August 31, 2026, cannot extract the model's reasoning process through historical context manipulation, a measure to prevent distillation attacks. Anthropic advises existing integrations to check compatibility documentation. The model is available on all Claude platforms (Pro, Max, Team), AWS, Google Cloud, and Azure. Sonnet 5.5 and Haiku 5.5 will follow in subsequent weeks.