Claude Fable 5 Is Back — and Anthropic Just Rewrote the Rules for How AI Gets Released

On June 12, Claude Fable 5 went dark. On June 30, it came back. But the version of the AI industry that exists on July 1 looks meaningfully different from the one that existed on June 11, and the eighteen days in between are why.

Anthropic confirmed on June 30 that the US government has lifted the export controls that forced the company to suspend global access to both Claude Fable 5 and Claude Mythos 5 earlier this month. Fable 5 is available again starting today — July 1, 2026 — across Claude.ai, Claude Platform, Claude Code, and Claude Cowork for users globally.

What makes this redeployment worth reading carefully isn't the return of access. It's the four detailed commitments Anthropic made in its official statement — commitments that amount to a new operating agreement between one of the world's most powerful AI companies and the US government, arrived at under pressure, and likely to shape how every other frontier AI company operates from here.

Why It Was Pulled: What We Now Know for Certain

The export controls on June 12 were triggered by a specific incident: Amazon researchers had found a method of bypassing Fable 5's safety classifiers — a jailbreak that allowed the model to identify software vulnerabilities and, in one case, produce code demonstrating how a specific vulnerability could be exploited.

Anthropic's official account of what happened next is important to read closely, because it directly addresses the framing that dominated coverage at the time.

The company's own testing confirmed that the jailbreak did not expose any unique Mythos-level capabilities — the kind of advanced offensive cyber abilities that the far more restricted Mythos 5 possesses. More significantly, Anthropic found that when it tested the same jailbreak technique across a range of competing models — including Claude Opus 4.8, GPT-5.5, and Kimi K2.7 — every single one of them could produce the same vulnerability identification. When it tested the specific code-demonstration behavior, every model tested could replicate it, including Claude Haiku 4.5, Sonnet 4.6, and multiple GPT variants.

In plain terms: Fable 5 wasn't uniquely dangerous in the way the export control implied. The behavior was, as Anthropic put it, "a borderline case for Fable 5's safeguards" — a jailbreak that had penetrated the safety margin the company deliberately sets around genuinely harmful behaviors, but had not actually reached those harmful behaviors themselves.

That context matters for understanding both the original shutdown and what has changed since.

What Anthropic Actually Changed

Despite concluding that the jailbreak didn't expose unique dangerous capabilities, Anthropic moved quickly to address the specific technique that triggered the shutdown.

Working with the US government, the company trained a new, improved safety classifier specifically designed to target and block the vulnerability-identification behavior described in the Amazon report. The new classifier blocks the specific technique in over 99% of cases. When a request is blocked, users are notified and the request is automatically redirected to Claude Opus 4.8 instead.

There is a cost to this, and Anthropic acknowledged it honestly: the improved classifier flags more legitimate requests as potentially problematic. Routine coding and debugging tasks will occasionally trigger the block and get routed to Opus 4.8 instead of Fable 5. The company said it will continue refining the classifier to reduce these false positives over time.

The US Department of Commerce's Center for AI Standards and Innovation reviewed both the original and new safeguards and confirmed they are "extraordinarily strong."

The Jailbreak Severity Framework: A Bigger Deal Than It Sounds

The most consequential development to come out of the Fable 5 shutdown isn't the new classifier. It's the industry framework that the shutdown's aftermath is producing.

Anthropic has announced that it is partnering with Amazon, Microsoft, Google, and other Glasswing partners to develop a consensus framework for assessing the severity of AI jailbreaks — and how AI companies should respond to different severity levels.

To understand why this matters, consider the problem it's solving. Right now, there is no agreed standard anywhere in the AI industry for what makes one jailbreak more serious than another. When a researcher discovers a way to bypass an AI model's safety systems, there is no common vocabulary for describing how dangerous that finding actually is — which means companies, governments, and the public have no reliable way to tell whether a given jailbreak is a minor technical curiosity or a genuine national security concern.

That gap is exactly what created the Fable 5 situation. A jailbreak that Anthropic's own analysis says was minor — affecting behaviors well within the safety margin, replicable in weaker models — triggered the first-ever export control on an AI model. A shared severity standard, had one existed, might have produced a very different government response.

Anthropic's proposed framework scores any discovered jailbreak across four criteria. The first is capability gain — how far beyond existing tools does the jailbreak take the user, and does it expose capabilities not available elsewhere? The second is breadth — does the jailbreak unlock a single narrow behavior or a wide range of harmful capabilities? The third is ease of weaponization — how much effort does it take to turn the jailbreak into an actual attack? The fourth is discoverability — is this technique already circulating publicly, or does it require specialist knowledge to find?

This four-dimensional scoring system is, in essence, an attempt to bring to AI the same kind of structured severity assessment that cybersecurity has used for software vulnerabilities for decades. The Common Vulnerability Scoring System has given the security industry a shared language for prioritising patches and communicating risk for years. Anthropic and its partners are trying to build the equivalent for AI jailbreaks, and they're inviting any industry partner or model provider to join the effort.

A new HackerOne program has also launched where security researchers can submit potential cyber jailbreaks in Fable 5 for structured review.

The New Government Partnership

The second major development is a set of formal commitments Anthropic made about how it will work with the US government going forward — commitments that go significantly beyond anything the company had publicly committed to before.

The first commitment is pre-release government access and evaluation. For any model that materially advances capability in areas relevant to national security, Anthropic will give designated government partners expanded early access to both the model and its safeguards before broad public release — with dedicated Anthropic technical staff working alongside government evaluators during testing periods.

The second is rapid information sharing. When significant jailbreaks or misuse patterns are identified, Anthropic commits to quickly investigate, notify appropriate government counterparts, and share new safeguards so they can be independently tested. Threat intelligence will be shared with government partners in advance of publication.

The third is dedicated research resources — dedicated Anthropic teams working on shared government priorities, significant compute allocation to support government testing, and red-teaming expertise made available for joint AI evaluation research.

The fourth is a common industry bar: working with government and industry peers toward a shared, voluntary security and evaluation standard across frontier model providers.

Read together, these four commitments describe something that didn't formally exist before June 12: a structured, ongoing security relationship between Anthropic and US national security agencies — one that gives the government a seat at the table before the next powerful model ships, rather than after.

What About Mythos 5?

Fable 5 is back globally. Mythos 5 — the more powerful, less restricted model restricted to vetted Glasswing partners — is a different story.

Anthropic confirmed that access to Mythos 5 has been restored for a set of US organisations following US government approval on June 26. The company says it is continuing to coordinate with the government to expand access to the broader set of domestic and international Glasswing partners.

The timeline for Mythos 5 reaching international partners remains unconfirmed. What is clear is that Mythos 5 is on a fundamentally different regulatory track from Fable 5 — one where government approval precedes access rather than following it.

What This Means for Everyone Using AI Right Now

For individual users and developers who use Claude for coding, writing, research, and analysis — the practical answer is that Fable 5 is back, and the main change you'll notice is occasional redirects to Opus 4.8 when requests touch on cybersecurity topics. Anthropic has committed to refining those classifiers to reduce unnecessary blocks over time.

For businesses built on the Claude API, access is confirmed for Claude Platform, Claude.ai, Claude Code, and Claude Cowork starting July 1. Re-enablement on AWS, Google Cloud, and Microsoft Foundry is coming as quickly as possible.

For the broader AI industry, the implications are harder to contain in a bullet point. The Fable 5 shutdown and its aftermath has established several things that weren't established before: that the US government will use export controls on AI models themselves — not just the hardware behind them; that the trigger for such controls can be a jailbreak finding; that Anthropic has committed to giving the government pre-release access to frontier models going forward; and that the industry is now moving toward a formal jailbreak severity standard that could govern how all AI companies communicate risk to governments and the public.

Each of those things is genuinely new. The eighteen days between June 12 and June 30 produced all of them. And Fable 5 is back — but the world it's returning to has already moved on.