GitHub confirmed on Monday that Claude Fable 5 is generally available inside GitHub Copilot, reaching Copilot Pro+, Max, Business, and Enterprise subscribers from today. The announcement makes Fable 5 the first model in Anthropic's Mythos class to land in a major integrated development environment, and it carries a condition that sets it apart from every other Claude model GitHub has distributed: Anthropic will hold onto prompts and outputs for up to 30 days.
That caveat matters because all other Claude models in GitHub Copilot, including Claude Opus 4.8, Sonnet 4.5, and Haiku 4.5, operate under Zero Data Retention. Queries are processed and discarded. Nothing is stored on Anthropic's side once a response is returned. Fable 5 is different. The retention window exists so that Anthropic can run the safety classifiers that govern the model's behavior in high-risk areas, and the company says the retained data is never used to train models. After 30 days, the logs are deleted. But the window itself is new ground for an Anthropic model inside a Microsoft-owned platform.
Why Fable 5 Needs a Different Architecture
Claude Fable 5, launched today, is built on the same foundation as Claude Mythos 5 but with safety classifiers that intercept queries in designated high-risk categories. When a request touches cybersecurity, biology, chemistry, or model distillation at a level that crosses Anthropic's threshold, Fable routes it to Claude Opus 4.8 instead, with the user notified that a fallback occurred. Those classifiers require context to operate, which is the technical reason Anthropic needs the 30-day window. Stripping data retention would mean running Fable's capabilities without the safety layer that makes public distribution feasible.
Early numbers suggest the classifiers fire rarely. Anthropic reports that more than 95% of Fable sessions complete entirely on Fable's own responses, with no fallback triggered. The company's red team ran over 1,000 hours of adversarial testing and found no workable universal jailbreaks. The remaining 5% or fewer are the cases where retention is most consequential, since those are the sessions where the model's handling of a query was escalated and logged.
Claude Fable 5 in GitHub Copilot: Key Numbers
- Copilot plans includedPro+, Max, Business, Enterprise
- Data retention window30 days (vs. ZDR for all other Claude models)
- Training use of retained dataNone
- Fable sessions with no fallback>95%
- Red-team jailbreaks found0 (1,000+ hours of testing)
- SWE-Bench Pro score80.3% (vs. 69.2% for Opus 4.8)
Enterprise Admins Must Opt In
GitHub has set the Claude Fable 5 policy to off by default for Copilot Business and Enterprise accounts. Organization administrators need to go into Copilot settings and explicitly enable it. GitHub's framing is that enabling the policy constitutes acknowledgement of the data retention requirement, which puts the decision in the hands of the teams that manage data governance rather than individual developers. Smaller Copilot plans, Pro+ and Max, do not require an admin step.
The distinction between plan types reflects the different sensitivity of enterprise code. A solo developer on a personal Copilot subscription and a software engineer at a regulated financial institution are in very different positions when it comes to third-party data retention, even a 30-day window that is not used for training. The opt-in architecture gives enterprises a way to evaluate the trade-off before their entire engineering org starts sending queries through a model that retains logs.
"Unlike other Claude models in GitHub Copilot, Claude Fable 5 requires data retention to operate safety classifiers. Anthropic retains prompts and outputs for up to 30 days to detect harmful or abusive use. After 30 days, it deletes the prompts and outputs." GitHub Copilot changelog, June 9, 2026
Performance and the Microsoft Context
The timing adds a layer of context that is hard to ignore. As GitHub ships Fable 5 to millions of Copilot users, Microsoft internally told its own engineers to stop using Claude Code by June 30, citing cost management and a push toward GitHub Copilot CLI. The two decisions sit within the same Microsoft umbrella but reflect different calculations: the internal budget decision is about which tool employees use for their own work, while the Copilot distribution is about what GitHub can offer as a platform to its external developer base.
From a benchmark standpoint, the upgrade is meaningful. Fable 5 scores 80.3% on SWE-Bench Pro, the standard agentic coding evaluation, compared to 69.2% for Opus 4.8 and 58.6% for GPT-5.5. GitHub's internal testing found that Fable 5 completed equivalent autonomous coding tasks with fewer tool calls and lower token consumption than prior Opus-tier models, which has practical implications for the usage-based billing structure GitHub Copilot switched to on June 1. With per-token costs now visible to users and organizations, a model that reaches conclusions with fewer steps is a better economic fit.
What Changes for Developers
Fable 5 appears in the model picker in Visual Studio Code across all Copilot modes: chat, ask, edit, and agent. The rollout is gradual rather than instant for all accounts, so not every eligible user will see it immediately. There is no additional subscription charge; access is included at the plan tiers where it is available. Developers who prefer to stick with ZDR semantics can continue using Opus 4.8, Sonnet 4.5, or Haiku 4.5, all of which remain available and unaffected.
The data retention question will get more attention in the coming weeks as enterprise teams work through their own assessments. Some organizations with strict data residency requirements, particularly in the European Union and in regulated industries, will need to consult their legal and compliance teams before enabling the policy. GitHub has said it will provide additional documentation for enterprise administrators on how retained data is handled, including storage location and access controls.
For developers without those constraints, Fable 5 in Copilot is the most direct path to Mythos-class capability in a coding environment. The model's performance on long-horizon tasks, complex refactors, and visual code understanding is materially ahead of what Anthropic's previous generation offered. The 30-day retention window is the price of that capability in this distribution channel, and GitHub and Anthropic have decided it is a price they can openly disclose rather than obscure. See the full overview of what Claude Fable 5 offers across all platforms or browse the Claude model family for a comparison of tiers.
