Anthropic's Claude Fable 5 Launched, Rewrote Coding Rules, Then Vanished

Anthropic's most capable public model ever launched in GitHub Copilot — then got pulled by a US government export control directive three days later.

·
·
·
Read8 min
TypeNews
  • Launched then suspended: Claude Fable 5, Anthropic's most capable public model ever, launched in GitHub Copilot on June 9 and was pulled by US government export control directive on June 12.
  • New capability tier: Fable 5 is the first public Mythos-class model, scoring 80.3% on SWE-bench Pro and completing a 2-month codebase migration in one day at Stripe.
  • Classifier-based safety: Dangerous queries (cyber, bio, distillation) fall back to Opus 4.8 rather than hard-refusing; over 95% of sessions saw no fallback at all.
  • Data retention required: Unlike all other Claude models in Copilot, Fable 5 required 30-day prompt/output retention for safety classifiers — not for training.
  • Government dispute: Anthropic complied but publicly disagreed, arguing the alleged jailbreak is narrow, non-universal, and available in other deployed models including GPT-5.5.
  • Sovereignty risk surfaced: The episode exposed that frontier API dependencies can vanish overnight due to export controls, with no advance notice to customers.

Anthropic launched Claude Fable 5 on June 9, 2026, and it immediately landed in GitHub Copilot. Three days later, it was gone. The model's brief window of availability was enough to make clear that something genuinely new had arrived , and that the industry's assumptions about deploying frontier AI are about to get a lot more complicated.

A new tier above Opus

Fable 5 is the first publicly available model in Anthropic's new Mythos class , a capability tier that sits above Opus. Fable 5 is described as a Mythos-class model made safe for general use, with capabilities exceeding any model Anthropic has ever made publicly available, showing state-of-the-art performance on nearly all tested benchmarks across software engineering, knowledge work, vision, and scientific research. The longer and more complex the task, the larger Fable 5's lead over previous models.

The key benchmarks tell the story:

  • 80.3% on SWE-bench Pro and 29.3% on FrontierCode Diamond , the latter being Cognition's evaluation of whether models can pass difficult coding tasks while meeting production codebase standards
  • Highest score of any model on Hebbia's Finance Benchmark for senior-level reasoning
  • First model to break 90% on complex, long-running analytical tasks at one early-access partner
  • Highest-performing model on ViBench, Replit's end-to-end vibe-coding benchmark

The Stripe use case is the most concrete illustration of what "long-horizon" actually means in practice: in a 50-million-line Ruby codebase, Fable 5 performed a codebase-wide migration in a single day that would have taken a full engineering team over two months by hand. GitHub's internal benchmarks showed it completed equivalent autonomous coding work with fewer tool calls and lower token consumption than previous Opus-tier models.

How it actually works: safeguards as architecture

Fable 5 and its unrestricted sibling, Claude Mythos 5, share the same underlying model. What differentiates them is a layer of safety classifiers , separate AI systems that run in parallel and intercept requests before the main model responds. This is not a fine-tuned refusal system baked into the weights; it is an external gating mechanism.

Claude Fable 5 includes safeguards that limit its performance in specific areas where misuse risk is elevated. Harmful prompts related to cybersecurity, biology, chemistry, and health fall back to receive a response from Opus 4.8 instead. Critically, when a query falls back to Opus 4.8, you are billed at Opus 4.8 rates rather than Fable 5 rates. Anthropic's early data showed that more than 95% of Fable sessions involved no fallback at all.

The three classifier domains are:

  1. Cybersecurity , covers both exploit development and broader offensive cyber tasks like reconnaissance and lateral movement. Anthropic ran an external bug bounty that produced no universal jailbreaks in over 1,000 hours of testing.
  2. Biology and chemistry , broadly covers requests that could assist in designing dangerous pathogens or viruses. Mythos 5 outperformed dedicated protein language models on predicting viral shell assembly properties, which is why this domain is gated.
  3. Distillation , detects large-scale attempts to extract Claude's capabilities to train competing models, particularly in adversarial contexts.

A universal jailbreak , a prompt or harness that bypasses safeguards broadly, not just in narrow edge cases , is the threat model Anthropic is defending against. Non-universal jailbreaks, which only work in specific limited contexts, are considered acceptable residual risk. Some attacks only become visible across multiple requests: best-of-N jailbreaking, for example, sends hundreds of slight variations of a prompt in the hope that one will work. This is why Fable 5 introduced a new data retention requirement.

The data retention trade-off

This is the detail that matters most for enterprise teams. Anthropic retains prompts and outputs for up to 30 days to operate safety classifiers that detect harmful or abusive use, and after 30 days it deletes the prompts and outputs. Retained data is not used to train Anthropic's models.

This data retention applies only to Claude Fable 5. All other Claude models in GitHub Copilot, including Claude Opus 4.8, Sonnet 4.5, and Haiku 4.5, continue to operate under Zero Data Retention (ZDR). For Copilot Enterprise and Business admins, the Fable 5 policy is off by default and must be explicitly enabled , and enabling it constitutes acknowledgment of the retention requirement.

Anthropic employees cannot access conversations unless they are flagged for potential serious harm or upon a customer's written request, and every instance of access is recorded in a tamper-proof log that reviewers cannot suppress or modify.

Where it was available

Before the suspension, Fable 5 was accessible to Copilot Pro+, Max, Business, and Enterprise users across a wide surface area:

  • VS Code (chat, ask, edit, and agent modes), Visual Studio, JetBrains, Xcode, Eclipse
  • Copilot CLI, GitHub Copilot cloud agent, GitHub Mobile (iOS and Android)
  • github.com and the GitHub Copilot app

Pricing was set at $10 per million input tokens and $50 per million output tokens , less than half the price of the previous Mythos Preview model. It was included free on Pro, Max, and Team plans through June 22.

What it was genuinely good at

Fable 5 is the kind of model that changes how an agent feels when the task is underspecified. It directs exploration well: first learning the environment, then identifying what files, tools, and constraints are available, then building from that grounded picture. It does not spend much time narrating what it is about to do , if it has enough context, it starts building.

The memory improvements are particularly notable for long-running agent workflows. When Anthropic tested the model on the deck-building game Slay the Spire, giving Fable 5 access to persistent file-based memory improved its performance three times more than it improved Opus 4.8's , a proxy for how much better it is at maintaining state across extended autonomous tasks. One developer reported that Fable 5 reverse-engineered an entire 1989 DOS game executable overnight , work that had taken six months with earlier models.

What it was bad at

In coding task benchmarks, Fable 5 often kept working until the harness cut it off. That makes the model feel capable, but it also makes it expensive and slower in agent workflows that do not have strong stop rules. The recommendation from early testers was not to swap it in everywhere: use Fable 5 when the job is to explore, plan, and build, especially when the task can take longer in exchange for a more thorough implementation. For precision-focused tasks like code review, where comment volume and accuracy matter more than thoroughness, earlier models held their ground.

The government suspension

On June 12, three days after launch, the US government issued an export control directive ordering Anthropic to suspend all access to Fable 5 and Mythos 5 for any foreign national , inside or outside the United States. Because Anthropic cannot filter users by nationality in real time, it disabled both models globally to ensure compliance.

Anthropic's understanding is that the government believes it has become aware of a method of bypassing, or jailbreaking, Fable 5. Anthropic disputes the severity of the finding, stating that it reviewed the demonstration and found the level of capability displayed is widely available from other models, including OpenAI's GPT-5.5, and is used routinely by security defenders. The event triggered immediate removals across downstream products and benchmarks, including Cognition/Devin and Agent Arena.

Anthropic is complying with the directive while publicly disagreeing with it. Their core argument: if a narrow, non-universal jailbreak is sufficient grounds to recall a commercially deployed model, it would effectively halt all new model deployments across the entire frontier AI industry.

What this changes for everyone building on AI APIs

Engineers quickly reframed this as a sovereignty risk rather than a pure policy story. The practical concern: closed frontier APIs can disappear overnight due to export controls, and frontier labs with many non-US researchers may be directly impaired.

The Fable 5 episode surfaces several assumptions that now need updating:

  • API availability is not guaranteed. A model you depend on in production can be pulled within hours, with no advance notice and no technical failure on your part.
  • Safety architecture is now a product differentiator. Anthropic's classifier-based approach , where dangerous queries fall back to a capable but less powerful model rather than hard-refusing , is a meaningfully different design philosophy than baking refusals into weights.
  • Data retention is the new enterprise negotiation. The ZDR vs. retention trade-off is no longer just a compliance checkbox; it is now directly tied to which models you can access.
  • Jailbreak resistance is probabilistic, not binary. Anthropic's public position , that perfect jailbreak resistance is impossible, and the goal is to make jailbreaks narrow and expensive , is a more honest framing than the industry has generally offered.

Anthropic says it is working to restore access as soon as possible. In the meantime, Claude Opus 4.8, Claude Sonnet 4.6, and Claude Haiku 4.5 remain fully available across Claude.ai, the API, and all Anthropic products. The model that briefly showed what post-Opus AI looks like in a coding context is currently offline , but the architectural decisions Anthropic made to get it to general availability are already reshaping how the industry thinks about deploying capable models responsibly.

Trending
  • No trending articles

Comments

avatar

Next Reads