Anthropic Claude Fable 5 and Mythos 5 Overview

Performance and Reasoning Scaling

Fable 5 and Mythos 5 represent a new class of models from Anthropic that demonstrate significant improvements in software engineering and reasoning tasks. On the Frontier Code Benchmark, Fable 5 achieved a 46% score on the full benchmark, marking a substantial leap over previous models. The models exhibit a strong correlation between performance and test-time compute, where increasing the reasoning budget leads to improved outcomes. This aligns with the concept of scaling test-time compute to unlock capabilities that are otherwise latent in current model architectures.

Safety, Privacy, and Operational Constraints

Anthropic has implemented a safety classifier for Fable 5 that routes sensitive queries—particularly those related to cybersecurity—to the less capable Claude Opus 4.8 model to mitigate risks. This classifier has a reported false-positive rate of 5%. Furthermore, the release introduces a mandatory data retention policy for Fable and Mythos models, requiring that all traffic on these surfaces be retained for 30 days to facilitate defense against novel jailbreaks and complex attacks. While Anthropic states this data will not be used for training, the policy marks a departure from previous privacy standards.

Pricing and Access

Fable 5 is priced at $10 per million input tokens and $50 per million output tokens. While expensive, this remains lower than the pricing for OpenAI's GPT-5.5 Pro models. Access for subscription users is currently available but will be restricted after June 2022, at which point usage will require paid API credits. This shift signals the end of inclusive access to frontier-class models within standard monthly subscription tiers, reflecting the high computational costs of serving these larger, more capable models.