Claude Fable 5: Anthropic Doubles Costs for 5.7% Gain

The latest artificial intelligence model from Anthropic, Claude Fable 5, has recently taken the lead in the Artificial Analysis Intelligence Index, surpassing competitors like GPT-5.5. With a score of 64.9 points, it outperforms the best non-Anthropic model by about five points. However, this technological advancement comes at a significant cost.

The model shows a performance improvement of 5.7% compared to its predecessor Opus 4.8, but this is accompanied by a substantial increase in costs. Token prices have doubled, reaching $10 and $50 per million input and output tokens respectively, compared to $5 and $25 for Opus 4.8. Thus, a full execution of the index now costs nearly $10,000, which is double the previous price.

Despite these high costs, Claude Fable 5 sets records in several benchmarks. On AA-Omniscience, it scores 40 points, surpassing the previous leader, Gemini 3.1 Pro Preview, by seven points due to increased accuracy. However, its hallucination rate remains average, with a score of 55%.

In agentic tasks, Fable 5 continues to strengthen Anthropic's position. On GDPval-AA, it achieves an Elo of 1,932, a 2.2% increase over Opus 4.8. It also dominates Terminal-Bench Hard for agentic coding and Tau2-bench Telecom for tool usage. On Humanity's Last Exam, the model scores 53%, more than seven points ahead of Opus 4.8. A single HLE execution with fallback costs about $2,200, the most expensive of all models tested by Artificial Analysis.

Fable 5 uses the same base model as Claude Mythos 5, according to Anthropic, with additional protections for queries related to cybersecurity, biology, chemistry, and model distillation. When a filter is triggered, a fallback mechanism redirects the request to Opus 4.8. These redirected requests still count towards billing, thus increasing total costs.

Fable 5 retains the same context window of one million tokens as Opus 4.8. Pro, Max, Team, and Enterprise subscribers can use it until June 22, with usage counting at double the rate of Opus. After this date, it will switch to credit-based billing, making it even more expensive than the token rates suggest. Anthropic has stated that it will restore subscription access once capacity allows.

My colleague Max recently analyzed the strengths and weaknesses of Fable 5 and found that security filters block a large number of harmless requests, ranging from medical physics questions to basic security exams. Anthropic's system map also revealed invisible throttling that degrades Fable's performance when users attempt to build competing cutting-edge models, although Anthropic has since mitigated this.