Covered Model

Covered Model is Anthropic's term for their top tier releases that come wrapped in mandatory data retention policies. Fable 5 earned the label on launch day in June 2026. The classification requires Anthropic to hold all user prompts and generated outputs for a minimum of 30 days in their systems. This retention supports the sophisticated classifier network that scans every interaction for potential policy breaches around sensitive domains. When those classifiers activate on a Fable 5 session the API call returns successfully but the model performance drops to Opus 4.8 without clear notification in many cases. The data stays behind regardless. Design teams running complex agentic tasks now confront a hard governance choice they did not face with lighter models. The retention applies universally across all subscription levels. No special exemptions exist for creative workflows or design agencies. The policy reflects the capability level of Mythos class systems and the corresponding oversight Anthropic believes they require.

Covered Model is not a description of technical ability or intelligence. The label carries no weight when measuring performance on CursorBench or FrontierCode Diamond benchmarks where Fable 5 set new records. It is not an indicator of context window size or output token limits. Those specs remain separate from the compliance wrapper. The term does not refer to the research only Mythos 5 variant that ships with reduced guardrails for partners in Project Glasswing. Covered Model specifically denotes the public facing version with full policy enforcement. It is not a configurable setting or something you can disable through the Claude dashboard. The 30 day hold is structural. It is not a short term precaution that will vanish once Anthropic collects enough launch day data. Legal and safety reviews baked it in as a permanent condition for accessing these capabilities. It is not equivalent to previous Claude versions that permitted zero data retention configurations for qualifying enterprise customers.

The launch provided immediate concrete examples across the design community. On June 10 2026 multiple Zed users posted in the Hacker News thread about their zero data retention configurations blocking Fable 5 integration entirely. One lead designer at a boutique studio responsible for a fintech mobile app overhaul abandoned plans to use the model for their token migration task after discovering the retention rules. The project involved proprietary payment flow designs that could not legally reside on third party servers for 30 days. GitHub Copilot for Business customers at larger agencies encountered identical blockers forcing them to maintain dual setups with Opus 4.8 for client work. Simon Willison highlighted the retention implications in his benchmark tests where even the low effort pelican SVG runs at 9.67 cents still triggered the full data hold. matheusmoreira described a detailed code review session on internal Lisp tools that got flagged and retained despite being legitimate work. Elie Bakouch amplified the frustration from the research side noting how the policy created friction for teams exploring new LLM frontiers in 2026. The positive counterpoint came from the Stripe engineering group who operated under compliance frameworks that accepted the retention terms. They finished a massive Ruby codebase migration in one day that traditionally required two months of human effort. Designers face parallel situations when they attempt full scale design system conversions or multi file Figma to production code generations that embed client specific patterns and brand secrets. The retention clock starts immediately and runs for the full period regardless of session length or classifier activity. A 2025 healthcare design team at a medical device startup hit the same wall when testing organic UI layouts that referenced biological concepts. Their legal team killed the trial after one session because the 30 day retention violated their data sovereignty clauses established after the 2023 breaches.

Reach for Covered Models in scenarios where your organization controls all the data and the upside from superior reasoning outweighs any governance overhead. These models deliver on long horizon tasks like complete component library refactors that span hundreds of components and require the model to track design decisions across extended sessions. They handle the kinds of ambitious assignments that Karpathy described where you expand the brief and the model simply understands the full scope. Use them for internal capability assessments or when developing your own tools and plugins without external IP concerns. The gains on Terminal Bench at 88 percent and Hebbia Finance benchmarks make them powerful for data heavy design audits if your legal team signs off. Steer clear of Covered Models whenever your projects involve client owned assets or fall under strict regulatory environments. Healthcare design teams mapping patient experiences or banking teams updating compliance dashboards cannot risk the 30 day exposure. Avoid any workflow that might include competitor research or unreleased concepts that must disappear immediately after use. The June 22 2026 deadline created an urgent evaluation period for subscription users. Those who tested agentic design workflows during the free access window quickly identified where the classifiers interfered with biology inspired organic layouts or security focused authentication prototypes. After the cutoff usage credits apply on top of the compliance considerations creating a double tax on adoption. The policy has no planned softening in the near term so design governance teams must build their tool selection matrices accordingly. Teams at agencies like Frog Design and IDEO audited their client contracts in the first week and routed sensitive brand work to Sonnet 4.6 instead.

Covered Model is frontier capability with a 30 day string attached that many design teams cannot afford to pull.