What Claude Fable 5 Is—and Why It Matters
Claude Fable 5 is Anthropic’s new "Mythos-class" AI model that targets Mythos-level performance while adding stronger safety guardrails and a revised data retention approach for enterprise users. Positioned above Claude Opus, Fable 5 is designed to deliver top-tier results in software engineering, complex knowledge work, vision tasks, and scientific research, while remaining "safe for general use" through strict AI model security controls. Anthropic says Fable 5 scores at or near the top of most capability benchmarks and can handle long, complex tasks better than its earlier public models, giving enterprises access to advanced AI without exposing them to the higher-risk behavior associated with Mythos itself. Alongside Fable 5, Anthropic is making Claude Mythos 5 available in private to selected Glasswing partners, drawing a clear line between an aggressively capable research model and a safety-focused system intended for broad deployment.
Safety Architecture: Classifiers, Failover, and Claude Fable 5 Safety
Anthropic’s safety design for Claude Fable 5 centers on classifiers and controlled failover. When prompts touch on high‑risk areas—cybersecurity, biology, chemistry, or model distillation—Fable 5 does not answer directly. Instead, it routes those requests to Claude Opus 4.8, described as the "next‑most‑capable" model but less likely to enable harmful misuse. Anthropic explains that "Fable 5 is able to handle requests itself roughly 95% of the time," with cautious guardrails that can occasionally block benign queries. These AI model security controls are backed by Project Glasswing testing, including a bug bounty program in which no white‑hat tester could find a universal jailbreak after 1,000 hours of attempts. For enterprise AI safeguards, that combination—specialist classifiers, automatic fallbacks, and intensive red‑teaming—signals a move toward shipping powerful systems only when they can reliably refuse the highest‑risk instructions.
Benchmark Power Without Mythos-Level Risk
On benchmarks, Fable 5 aims squarely at Mythos-level performance without exposing users to Mythos-level risk. Anthropic reports that Fable 5 and Mythos 5 outperform Mythos Preview, Opus 4.8, OpenAI’s GPT‑5.5, and Google’s Gemini 3.1 Pro across agentic coding, knowledge work, spatial reasoning, tool use, legal tasks, multidisciplinary reasoning without tools, biology, cybersecurity, and health. Mythos Preview only stays ahead in computer use and multidisciplinary reasoning with tools. In practice, Anthropic claims that Fable 5 completed a coding project in one day that would otherwise have taken a team more than two months, rebuilt a web app’s source code from screenshots, and played complex games such as Slay the Spire and Pokémon FireRed far better than previous Claude models. For enterprises, this suggests Mythos-class throughput on difficult workstreams while the safety system filters out the most dangerous use cases.
Anthropic Data Retention Policy: What Changed for Zero-Data Clients
Alongside Fable 5, Anthropic updated its data retention policy in a way that directly affects organizations that previously relied on strict zero data retention. The company now keeps prompts and outputs from Mythos-class models for 30 days for trust and safety, even when customers configure zero data retention in Claude Console or access the models through cloud partners. According to Anthropic, "Prompts submitted to, and outputs generated by, Mythos-class models are retained for 30 days for trust and safety purposes, on every platform where these models are offered." Consumer plans, such as Claude Free, Pro, and Max, already operated under a retention policy, so they remain unchanged. Anthropic stresses that this log data will not be used to train new models, framing the shift as a necessary trade‑off: short‑term logging to detect and investigate misuse, balanced against enterprise expectations of privacy and compliance.
Dual-Tier Release and the Future of Enterprise AI Safeguards
The split between publicly available Fable 5 and privately tested Mythos 5 under Project Glasswing highlights a dual-tier release strategy that is becoming central to responsible AI deployment. Mythos-class capability is now accessible, but only the safer tier reaches the general market; the more hazardous, less constrained model stays in controlled partnerships with governments and selected organizations. For enterprises, this offers a clearer risk envelope: Fable 5 aims to maximize productivity while built‑in classifiers, failover behavior, and 30‑day logging form a defensive perimeter. At the same time, the new Anthropic data retention policy forces security and legal teams to weigh stronger safety monitoring against stricter internal data rules. Claude Fable 5 safety features suggest where the sector is heading: powerful models paired with explicit enterprise AI safeguards, and a norm that the most capable systems will be gated, audited, and continuously monitored rather than fully open.
