Claude Fable 5 Performance and Pricing Overview
Better Stackgo watch the original →
the gist
Anthropic's new Mythos-class model, Fable 5, leads on coding and reasoning benchmarks but introduces restrictive pricing, high usage costs, and mandatory 30-day data retention policies.
Benchmark Performance and Coding Capabilities
Claude Fable 5 demonstrates significant performance gains over existing models, particularly in coding and long-running tasks. On the SWE Bench Pro benchmark, the model shows a 10% improvement in agentic coding performance compared to previous iterations. It also outperforms competitors on the Frontier Code benchmark, which evaluates whether code generated by a model would be accepted by a human maintainer. The model exhibits improved memory management, enabling it to maintain focus across millions of tokens for complex, long-running tasks, such as performing a codebase-wide migration of a 50-million-line Ruby project in one day.
Practical Application and Design
In vision-based tasks, the model demonstrates the ability to solve complex puzzles, such as beating Pokemon Fire Red using a vision-only harness without additional tools. When tasked with generating a website from a single screenshot, Fable 5 successfully reconstructed the UI, including layout and section structure, without external web search. In a comparative test building a finance dashboard app, Fable 5 completed the task in 8 minutes, compared to 12 minutes for Opus 4.8 and 15 minutes for GPT 5.5, while producing a more polished UI design than the latter.
Safeguards and Operational Constraints
Fable 5 incorporates strict safety measures, particularly regarding cybersecurity tasks, where it currently refuses to execute requests that trigger safety filters. When a request is flagged, the system attempts to route the task to Opus 4.8 to determine if it can be processed safely. Users should expect a high rate of false positives, with the model failing to pass cyber-evaluation benchmarks due to these refusals. Additionally, the model consumes usage limits at twice the rate of Opus 4.8.
Pricing and Data Policy
The model is priced at $10 per million input tokens and $50 per million output tokens. Anthropic has implemented a temporary access model where Fable 5 is available in Pro, Team, and Enterprise plans until June 23, after which it will transition to a usage-credit-based model. Furthermore, using the model requires a mandatory 30-day data retention policy for all traffic on first and third-party tools, which Anthropic states is for security monitoring rather than model training.