Anthropic releases Claude Opus 5.5

Anthropic released Claude Opus 5.5, matching the performance of its premium Fable 5.1 model on most work tasks while costing 40% less. Input and output prices are $4 and $20 per million tokens, with 30% faster output. It also excels at long codebase audits and accurate report generation.
Key insights
Opus 5.5 achieves parity with Fable 5.1 at reduced price
Anthropic claims Opus 5.5 equals Fable 5.1 on most tasks while costing 40% less. Input/output prices are $4/$20 per million tokens, 20% cheaper than Opus 5. Cached read dropped to $0.2 per million tokens. In Terminal-Bench, it matched GPT-6 Astra at about 40% of the cost. One audit of a 200,000-line codebase took under 3 hours vs 20+ hours for Opus 5.
Opus 5.5 shows improved accuracy for research and writing
In a test requiring models to write quarterly earnings reports using only web search, Opus 5.5 passed 16 of 18 reports without errors, while Fable 5.1 and Opus 5 failed all. The model also adopts a more direct writing style, putting conclusions first and following user instructions more closely.
New safety restrictions and developer adjustments required
Opus 5.5 inherits safety measures from Fable 5.1: most cybersecurity tasks are routed to older Opus 4.8, and biology work is restricted. Thinking mode cannot be disabled. API accounts registered after August 31 cannot use historical context to extract reasoning, as anti-distillation protection. Developers must adjust prompts, especially for effort settings and progress updates.
LEGO design demo showcases advanced spatial reasoning
HuggingFace team used Opus 5.5 to design a Microduck robot from real LEGO bricks. The model used 1,113 parts in 6 colors, all connections matched in 3D with zero collisions, and generated a 141-page building manual with parts list and BrickLink order. This demonstrates strong visual and spatial reasoning abilities.