Anthropic Fable 5.1 and Mythos 5.1 Overview

Model Performance and Cost Analysis

Anthropic's new Fable 5.1 and Mythos 5.1 models represent the current frontier for the company, offering improved reasoning capabilities at the cost of higher token consumption. While Anthropic highlights a 75% price reduction for cache reads—dropping to $0.25 per million tokens—the total cost per task often exceeds that of previous versions because the models require approximately 1.7x more output tokens to reach their peak intelligence levels. Benchmarks show Fable 5.1 achieving a 66 score on the Artificial Analysis leaderboard, outperforming Opus 5 and GPT 5.6, but remaining significantly more expensive per task than models like Grock 4.6 or GPT 5.6.

Enterprise Safeguards and Data Policy

Anthropic introduced Enterprise Frontier Safeguards (EFS) to address zero data retention requirements for corporate clients. This system allows customers to control the cloud infrastructure where their data is stored, though Anthropic retains read access for misuse detection. Additionally, the company implemented new restrictions on API accounts to prevent users from manually editing prior context in multi-turn conversations, a move intended to mitigate distillation attacks where third parties extract model intelligence to train competing models. All outputs from models released after August 2nd now include watermarks to comply with EU AI Act transparency requirements.

Benchmark Improvements

Fable 5.1 shows measurable gains in specialized benchmarks compared to its predecessor:

• Terminal Bench Science: Fable 5.1 achieved a 26% score at an $11 cost, compared to Fable 5's 25% score at $34.
• Cursor Bench: Fable 5.1 reached 73.4% accuracy at $9.64, improving upon Fable 5's 70.5% accuracy at $17.32.
• Humanity's Last Exam: The model reached 65% pass rate at max effort, a slight increase over the 63.8% achieved by Fable 5.