Claude Fable 5 went GA on June 9 — simultaneously across the Anthropic API, Amazon Bedrock, Microsoft Azure Foundry, and GitHub Copilot. If you’re on Copilot Pro+, Max, Business, or Enterprise, the most capable model Anthropic has ever shipped is now accessible inside VS Code, Visual Studio, JetBrains, and every other Copilot-integrated IDE.
There is a catch. For enterprise and business accounts, Fable 5 is off by default, requiring explicit admin opt-in in organization settings. Understanding why requires knowing something Anthropic has not loudly advertised: Mythos-class models carry a mandatory 30-day data retention requirement.
The Zero Data Retention problem#
GitHub Copilot’s enterprise tiers have offered Zero Data Retention (ZDR) as a compliance feature. Under ZDR, your code snippets, queries, and AI responses are not stored by GitHub or its model partners — they process the request and discard it. This matters for financial services, healthcare, defense contractors, and any organization with strict data governance requirements.
Fable 5 cannot comply with that contract.
Anthropic requires that interactions with Mythos-class models be retained for 30 days so that safety classifiers can process them. Fable 5’s capability tier — the first Claude to sit above Opus, the model widely understood to be the commercial release of the Mythos-class system behind Project Glasswing — carries new oversight requirements. The safety infrastructure for the most capable models has audit dependencies that ZDR would prevent.
The result: a GitHub Copilot Enterprise customer with ZDR enabled cannot use Fable 5 under that ZDR configuration. An admin must make a deliberate decision to trade ZDR coverage on Fable 5 interactions for access to the model. Importantly, this decision is scoped: Opus 4.8, Sonnet, Haiku, and GPT-5.5 continue to honor existing ZDR settings. Only Fable 5 interactions carry the 30-day retention requirement.
Why “off by default” is the right call#
GitHub and Anthropic made the correct UX decision here. Automatic enrollment would be wrong: enterprise customers who chose ZDR for compliance reasons should not suddenly find Fable 5 interactions being retained without an explicit decision by someone with authority to make it.
Requiring admin opt-in does more than prevent a surprise compliance failure. It ensures that the governance function — legal, security, compliance — knows Fable 5 is in organizational use. Developers who find the model in the picker and start using it without enterprise enablement won’t get access; that’s the filter working as intended.
The tradeoff is explicit by design: read Anthropic’s retention terms, decide whether the capability is worth the governance change, enable it org-wide if yes. That’s a more mature integration than most model additions to Copilot have warranted. Fable 5 merits it.
What GitHub’s CPO said#
GitHub CPO Mario Rodriguez framed the Fable 5 integration in specific terms. This marks the beginning of Copilot moving from “AI assistance” to “AI agents completing spans of engineering work.”
The framing is significant because it’s what every AI coding tool vendor has been claiming for 18 months. The difference with Fable 5 is that the underlying model makes the claim more credible. Fable 5’s agentic capabilities — codebase-scale reasoning, multi-session continuity, deeper tool-use reliability — are the reason Anthropic positioned it as “our most intelligent model” rather than just another capability increment. Rodriguez is attaching Copilot’s narrative to a model that can genuinely execute multi-step engineering work, not just complete the next token.
The structural critique remains unchanged. Copilot’s IDE-embedded architecture means Fable 5 runs inside your editor, not as a true autonomous terminal agent with filesystem control, long-horizon memory, and the full Claude Code skill/plugin/MCP ecosystem. You get Fable 5’s intelligence in the IDE container — which is genuinely powerful — but not Fable 5’s full agentic potential, which requires Claude Code’s architecture to express. A Formula 1 engine in a family sedan is still a better sedan. You haven’t bought the F1 car.
Free access window closes June 22#
For Pro, Max, and Team plan Copilot subscribers (not Business/Enterprise), Fable 5 is free through June 22. Starting June 23, using Fable 5 in Copilot costs GitHub AI Credits at rates that correspond to $10 per million input tokens and $50 per million output — identical to the Anthropic API and Azure Foundry pricing.
Opus 4.8 in Copilot remains on the standard included model tier with no extra credit cost. Teams that adopted the Opus-in-Copilot workflow since the April 23 API default switch are not being pushed toward Fable 5 unless the work justifies it. The two-tier structure (Opus at $5/$25 on the API, Fable 5 at $10/$50) maps directly into Copilot’s credit model. You pay for the step-change, or you keep using the workhorse.
Azure Foundry: the quieter launch#
The Fable 5 Copilot launch ran alongside a simultaneous GA on Microsoft Azure Foundry — the enterprise AI platform that replaced Azure OpenAI Service. This matters more for enterprise teams than the Copilot integration for a specific reason: Foundry offers more deployment control (VNet integration, customer-managed keys, Azure Policy compliance gates) and operates outside the IDE paradigm entirely.
If your organization already has Opus 4.8 deployed via Azure Foundry for internal tooling, Fable 5 is available in the same deployment environment at the same compliance tier. No new contracts, no new onboarding, no Copilot subscription required. That is a faster path to Fable 5 in production for enterprise teams than navigating the Copilot licensing and governance opt-in process.
The direct-to-API alternative#
If data governance is a hard constraint, the options are not binary. “Use Fable 5 with 30-day retention via Copilot” and “don’t use Fable 5” are not the only paths. A third option: deploy Fable 5 directly via the Anthropic API under a Business Associate Agreement, or via AWS Bedrock under HIPAA-eligible configuration, where retention and processing terms are negotiated directly rather than inherited through Copilot’s layered agreements.
For teams already running Claude Code as their primary interface — the terminal-native path where Fable 5’s agentic capabilities fully express — the Copilot integration was never the relevant deployment anyway. /model fable has been live in Claude Code since June 9. If you’re already using Claude Code with Fable 5 for long-horizon tasks, you’re getting the full capability set with data terms governed directly by your Anthropic or Bedrock agreement, not by Copilot’s retention policy.
What to actually do#
Copilot Pro/Max/Team users: Try Fable 5 before June 22 while it’s free. Prioritize high-stakes sessions — architecting systems, debugging non-obvious bugs, long-context review of complex codebases. Benchmark your actual usage cost before June 23 so the credit impact isn’t a surprise.
Copilot Business/Enterprise admins: Read Anthropic’s 30-day retention terms before enabling. If ZDR compliance is non-negotiable for any code-adjacent use, deploy Fable 5 via Azure Foundry under your existing enterprise agreement instead. If the retention terms are acceptable, enable it and communicate to your engineering org that Fable 5 interactions are retained — they should know.
Claude Code users: If you’re not on /model fable yet for your hardest tasks, now is the time. The capability gap between Opus 4.8 and Fable 5 is real, the 2× price is real, and the combined recursive sub-agents (v2.1.172) + Dynamic Workflows architecture means you can actually use Fable 5’s full agentic depth in Claude Code — which Copilot’s IDE container cannot match.
Sources: GitHub Changelog — Claude Fable 5 in Copilot, Coursiv — Claude Fable 5 in GitHub Copilot, Windows News — Fable 5 in Azure Foundry & Copilot