Anthropic Releases Claude Opus 5.5 With Fable 5.1 Performance

Anthropic has released Claude Opus 5.5, the inaugural model in its new Claude 5.5 family, which the company states achieves performance levels comparable to Claude Fable 5.1 across most tasks. This new model also boasts a 40% reduction in running costs compared to its predecessor, Opus 5, when utilized for typical workloads at default settings. Anthropic's internal benchmarks indicate that Opus 5.5 demonstrates leadership in areas such as agentic coding, computer use, and knowledge work. The model is accessible as a managed API service through the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. Anthropic has not released the model's weights, precluding self-hosting options. Similar to previous Opus models, Opus 5.5 offers zero data retention capabilities for enhanced privacy. On Anthropic's proprietary benchmarks, Opus 5.5 exhibits strong performance, though it does not achieve a clean sweep across all metrics. Specifically, Opus 5.5 scored 66.4% on Terminal-Bench 4.0, surpassing Fable 5.1's 55.8% and Opus 5's 52.3%. It also achieved 54.4% on FrontierCode v1.1, exceeding Fable 5.1's 50.3% and Opus 5's 48.0%. In the CursorBench 4.0 benchmark, Opus 5.5 scored 57.8%, significantly outperforming Fable 5.1 (51.8%) and Opus 5 (46.6%). On OSWorld 2.0, it reached 81.8%, compared to Fable 5.1's 80.7% and Opus 5's 74.0%. However, GPT-6 Astra, another advanced model, still leads Opus 5.5 on the Terminal-Bench-Science benchmark with 64.6% versus 58.7%, and on AutomationBench with 41.4% versus 40.0%. Anthropic notes that Terminal-Bench 4.0 was evaluated at a higher effort setting for Opus 5.5. Furthermore, Zapier's evaluation of AutomationBench excluded fallback models, meaning safeguard interventions were counted as failures, potentially impacting the comparison. Anthropic also cautions that benchmark margins are becoming less indicative of real-world performance differences. The cost-effectiveness of Opus 5.5 is a key highlight. At default medium effort, Opus 5.5 achieves 54.6% on FrontierCode, outperforming GPT-6 Astra's top score of 53.3% at approximately one-fifth of the cost per task. Similarly, on CursorBench, a medium effort score of 52.5% for Opus 5.5 is 11 points higher than GPT-5.6 Sol's best performance, at roughly one-third of the cost. The pricing structure for Opus 5.5 reflects its reduced computational requirements. The cost per 1 million tokens for input is $4, compared to $5 for Opus 5. Output tokens are priced at $20 for Opus 5.5, down from $25 for Opus 5. Cache reads are priced at $0.20 per 1 million tokens for Opus 5.5, a reduction from $0.50 for Opus 5, and cache writes are priced at $0.20 for Opus 5.5, down from $0.50 for Opus 5. This pricing strategy underscores the model's improved efficiency and cost savings for developers and businesses utilizing the Claude API.