Claude Fable 5.1: Performance, Costs, and Agentic Benchmarks

Performance and Benchmarks

Claude Fable 5.1 represents a significant leap in agentic and coding capabilities, achieving a 92.5% score on the KingBench suite, the first model to surpass the 90% threshold. The model demonstrates superior performance in complex 3D generation, notably scoring a 9 out of 10 on the difficult wristwatch task, where its predecessor, Fable 5, scored only a 4. While gains in general knowledge work and terminal tasks are incremental, ranging from 2 to 5 points over Opus 5 and Fable 5, the model's ability to handle end-to-end agentic workflows—such as dataset generation and local model fine-tuning—is robust.

Cost Structure and Efficiency

API pricing remains at $10 per million input tokens and $50 per million output tokens. The primary economic shift is a 75% reduction in cache-read pricing, now at $0.25 per million tokens. Despite this, session costs are frequently dominated by cache-write operations, which remain priced at $12.50 per million for five-minute caches and $20 per million for one-hour caches. Users running long-horizon agentic loops can expect 20% to 45% cost savings, but short, one-shot requests see no financial benefit from these changes.

Operational Limitations and Safety

Several operational friction points persist. The model exhibits a tendency to perform whole-file rewrites rather than targeted edits, increasing output token consumption. Additionally, the safety alignment layer can trigger silent fallbacks to Opus 5 for security-related queries, often without user notification. The API now enforces append-only conversation rules and restricts access to thinking blocks to prevent model distillation, which complicates debugging for users who rely on inspecting raw chain-of-thought traces. Furthermore, the writing style is notably denser and less polished, which some users find exhausting for documentation tasks.