Fable 5 Hits 72.9% on CursorBench With 32K Reasoning Tokens

Anthropic’s technical case study on Claude Fable 5 details the model’s capacity to solve long-horizon engineering problems autonomously. The flagship Mythos-class model set a new state-of-the-art record on CursorBench, scoring 72.9% at the highest reasoning setting. With a 1 million token context window and a 128,000 token output capacity, Fable 5 targets agentic workflows that require massive codebase analysis without human intervention.

Autonomous Coding Benchmarks

The model demonstrates a significant performance gap over previous generations in autonomous coding evaluations, specifically excelling in tasks that demand multi-step reasoning.

Benchmark	Fable 5	Opus 4.8	GPT-5.5
SWE-Bench Pro	80.3%	69.2%	58.6%
FrontierCode (Diamond)	29.3%	13.4%	Not specified
CursorBench (Max Effort)	72.9%	Not specified	Not specified

Fable 5 more than doubles the performance of Opus 4.8 on the Diamond tier of Cognition’s FrontierCode evaluation. The 72.9% score on CursorBench represents an 8-percentage point increase over the previous best-performing model.

Effort Levels and Reasoning Budgets

These performance gains rely on a new control system that determines the model’s internal chain-of-thought token budget before it generates a final answer. Users select from five Effort Levels.

• None: 0 tokens for fast execution.
• Low: 1,000 to 2,000 tokens.
• Medium: 4,000 to 8,000 tokens.
• High: 16,000 to 32,000 tokens.
• Max: 32,000+ tokens.

The 72.9% CursorBench high score requires the Max setting. If you build autonomous systems, you must manage token budgets carefully to balance latency against reasoning depth.

Safety Classifiers and the Mythos Tier

Fable 5 is the first model in Anthropic’s “Mythos” classification, a tier designated for autonomous operational capability. This designation requires real-time safety classifiers to monitor for high-risk tasks involving cybersecurity or biology. If a request triggers the classifier, the system may refuse execution or route the prompt to the more restricted Claude Opus 4.8 model.

Anthropic also maintains Claude Mythos 5, a twin model with identical underlying intelligence but without the routing classifiers. This uncensored version remains available exclusively to approved organizations in the Project Glasswing program.

Pricing and Transition Timeline

Following a temporary global suspension in June 2024 due to US export controls, Anthropic restored Fable 5 access on July 1, 2026. Paid subscribers retain promotional access up to 50% of their weekly limits through July 19. On July 20, Fable 5 becomes standard for Max and Team Premium plans at the 50% limit threshold. Pro users will transition to usage-based billing, receiving a one-time $100 credit.

API pricing is set at $10 per million input tokens and $50 per million output tokens. The model includes a 90% discount on input tokens for cached context.

When evaluating and testing AI agents on Fable 5, baseline your prompts against the Medium effort level first. The jump to Max effort consumes a minimum of 32,000 tokens per interaction, making it cost-prohibitive for standard autocomplete tasks but necessary for complex, multi-file refactoring.