Anthropic Releases ‘Fable 5’: A High-Performance Powerhouse with a New Privacy Trade-off
Table of Contents
The Arrival of the ‘Mythos’ Class
Anthropic has officially expanded its model lineup with the release of Claude Fable 5, a new high-performance AI designed to push the boundaries of software engineering and scientific research. Positioned as a ‘Mythos-class’ model, Fable 5 sits above the existing Claude Opus in terms of raw capability, signaling Anthropic’s intent to compete more aggressively with the top-tier offerings from OpenAI and Google.
The release is split into two tracks: Fable 5 is now generally available to the public, while the more potent Claude Mythos 5 remains restricted to ‘Glasswing’ partners. This distinction is more than just nomenclature. According to Anthropic, the difference between Fable and Mythos lies primarily in the safety architecture. Fable is the ‘neutered’ version, equipped with aggressive guardrails and a failover mechanism that redirects high-risk queries—specifically those involving chemistry, biology, or cybersecurity—back to the more stable Claude Opus 4.8.
The Price of Performance
For developers and enterprises, the most immediate impact is the cost and efficiency. Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens. While this represents a significant price drop from the previous Mythos Preview (which cost $25/$125 respectively), it remains a premium product compared to smaller, faster models. Anthropic claims Fable 5 is more token-efficient, which may offset some of the higher costs for heavy users.
The real-world utility of the model is being highlighted through a high-profile case study involving Stripe. Anthropic claims that the payment giant used Fable 5 to migrate a massive 50-million-line Ruby codebase in a single day—a task the company estimates would have taken a human engineering team roughly two months to complete manually.
The ZDR Controversy: No More ‘Zero’ Data Retention
While the technical benchmarks are impressive, a significant shift in Anthropic’s privacy policy is drawing scrutiny from enterprise clients. The company is effectively ending ‘Zero Data Retention’ (ZDR) for Mythos-class models. Previously, organizations using Claude via AWS Bedrock, Google Cloud, or Microsoft Foundry could opt for ZDR to ensure their proprietary data never touched a log file.
Under the new policy, prompts and outputs for Mythos-class models will be retained for 30 days across all platforms. Anthropic frames this as a necessary security measure to combat misuse and track potential bad actors. The company insists that this data will not be used to train future models, but for organizations with strict compliance and data sovereignty requirements, the shift from ‘zero’ to ’30 days’ is a critical change in the trust model.
Benchmarking Against the Giants
In internal testing, Anthropic reports that Fable 5 significantly outperforms Claude Opus 4.8, OpenAI’s Codex 5.5, and Google’s Gemini 3.1 Pro across a variety of knowledge-work and coding benchmarks. Interestingly, the company omitted Google’s more recent Gemini 3.5 Flash from its comparison, leaving a gap in the current competitive analysis.
On the safety front, the model’s resistance to prompt injection—the act of ‘tricking’ an AI into ignoring its rules—is remarkably high. According to the official model safety card, Mythos 5 saw a success rate of only 4.8% over 100 attempts, putting its security on par with the highly regulated Opus 4.8.
For now, Fable 5 is available at no extra cost for Pro, Max, and Team subscribers until June 22. After this window, the model will shift to a usage-credit system, though Anthropic suggests it may return to standard subscription plans later, pending server capacity.
