Anthropic's Most Powerful Public Model Is Back After a 19-Day Government Shutdown

What You'll Learn

• The exact 19-day timeline, from the June 9, 2026 launch to the July 1 relaunch, with the legal mechanism the government used to force a shutdown in a single afternoon.
• What Fable 5 actually is versus Mythos 5, and why almost no Executive reading this will ever touch Mythos 5 directly.
• The specific vulnerability the government cited, and why cybersecurity experts publicly disputed how serious it was.
• When to reach for Fable 5 over Sonnet 5 for daily work, and the one case where Mythos 5 is worth requesting instead of either one.
• The single-vendor risk lesson every business running AI-dependent operations should take from this, with concrete steps.

At 5:21 p.m. ET on June 12, 2026, Anthropic received a letter from Commerce Secretary Howard Lutnick and had to pull its most capable public model offline for every customer on earth, including the hundreds of millions of people already using it. Not a bug. Not a service outage. A direct order from the US government, using export control authority, the same legal tool historically used to stop sensitive technology like encryption software or missile components from reaching foreign buyers, now applied to a chatbot subscription.

If any part of an Executive's operation leaned on Fable 5 for research, coding, or analysis during those 19 days, that capability simply vanished with zero notice. No status page warning. No 30-day migration window. Just gone by that evening.

Fable 5 is back as of today, July 1, 2026, available to everyone worldwide again. What follows is the precise timeline of why it disappeared, what the government actually claimed happened, why security experts split into two camps over whether the risk was real, and the specific guidance an Executive needs on when this model is worth reaching for now that it has returned.

June 12, 2026

Export control order shuts down both models after jailbreak report

~19 Days Offline

Open letter, partial thaw for vetted organizations, no public restoration date

June 30-July 1, 2026

Commerce lifts the order; Fable 5 restored globally

What Fable 5 and Mythos 5 Actually Are

Anthropic launched both models on June 9, 2026, as a matched pair built on the exact same underlying system. Mythos 5 is the raw version, with most of its safety filters removed. Fable 5 is the same brain wrapped in what Anthropic calls "guardrails": automated filters that watch every request and reroute anything that looks like it could be misused for hacking or bioweapon research to a different, less capable Anthropic model (Claude Opus 4.8) instead of answering.

Think of guardrails as a bouncer standing between the user and the model's full power. Ask Fable 5 to debug a spreadsheet formula or draft a contract clause, and the bouncer waves the request straight through. Ask it something that reads like an attempt to write malware or design a dangerous virus, and the bouncer redirects the conversation to the weaker model before it can help.

Anthropic built this fallback deliberately cautious, and it barely matters in practice: the company reports that more than 95% of Fable 5 sessions get zero interference, meaning normal business use runs at full Mythos-level strength without ever touching the guardrail.

Mythos 5 has no such bouncer for cybersecurity tasks. Anthropic describes it as capable of finding and exploiting software vulnerabilities better than any other model, and better than all but the most skilled human security specialists. That is exactly why it never opened to the public. It first reached roughly 200 companies in April 2026 through an invite-only program called Project Glasswing, including Apple, Amazon, Google, Cisco, Nvidia, and Microsoft, who used early access to stress-test it against their own products and hunt for flaws. Mythos 5 formally launched on June 9 alongside Fable 5, restricted to that same vetted list. On a standard Claude subscription, Mythos 5 was never available and is not becoming available now. Fable 5 is the model that matters for daily business use, and it is the one back today.

The 19 Days, Hour by Hour

Three days after launch, on June 12, the government sent Anthropic a directive under export control authority: federal power originally built to stop sensitive American technology from reaching foreign adversaries by regulating who can buy or use it. Applying it to an AI subscription legally required Anthropic to block every foreign national anywhere in the world, including foreign employees working inside Anthropic itself, from using either model.

Anthropic's public statement was direct about what happened next: with no reliable way to check a user's nationality in real time across every cloud platform and integration where Claude runs, the company had no option but to shut both models off for absolutely everyone, American customers included. Access to every other Claude model, including Sonnet 5 and Opus 4.8, was untouched throughout.

The government's letter did not spell out the specific national security concern in detail. Anthropic's own account says its understanding was that officials had learned of a technique for "jailbreaking" Fable 5, a prompt engineered to trick the model's filters into bypassing its guardrails. The report reaching the government reportedly came from Amazon researchers, who found a way to prompt Fable 5 into reading a codebase and identifying, then in one case demonstrating, a software vulnerability.

Anthropic pushed back hard on the framing. Its statement said the flaws surfaced by this technique were minor and already discoverable by other publicly available models without any bypass required, and that this was not a "universal jailbreak," industry shorthand for a technique that opens up a model's full range of dangerous capabilities rather than one narrow behavior. The company later confirmed that Claude Opus 4.8, GPT-5.5, and Kimi K2.7 could all reproduce the same findings, and that models going back to Claude Haiku 4.5 could replicate the exploit demonstration.

The dispute went public fast. On June 14, more than 80 cybersecurity CEOs, executive directors, and security engineers signed an open letter to Lutnick and National Cyber Director Sean Cairncross asking for the export controls to be lifted, arguing no one had shown the jailbreak produced genuinely dangerous new capability. Coverage from CyberScoop and other outlets in the days that followed reflected a real split among security experts: some backed the government's caution, others called the underlying vulnerability routine and the shutdown disproportionate to the actual risk.

80+

Cybersecurity CEOs, executive directors, and security engineers signed an open letter asking the government to lift the export controls, arguing no one had shown the jailbreak produced genuinely dangerous new capability.

The first partial thaw came on June 26, when Lutnick sent a follow-up letter to Anthropic stating that appropriate safeguards were now in place to let certain trusted partners back into Mythos 5. By June 27, access was restored to more than 100 vetted US organizations, including government agencies and Fortune 500 companies, many with the same pre-shutdown access from the original Glasswing program. Fable 5, the public-facing model, stayed dark with no restoration date announced.

The full resolution landed on June 30. Commerce formally lifted the export control restrictions on both models that day. In exchange, Anthropic committed to a new package of safety measures:

• Giving the government pre-release access to future high-capability models and their safeguards for independent testing.
• Sharing information rapidly when new jailbreaks or misuse patterns surface.
• Dedicating staff and compute to joint government security research.
• Working toward a shared industry standard for scoring how serious a given jailbreak actually is.

Anthropic also disclosed it had trained an improved safety classifier that blocks the specific technique from the Amazon report in more than 99% of cases, verified independently by the Commerce Department's Center for AI Standards and Innovation.

Fable 5 goes live globally today, July 1, across the Claude Platform, Claude.ai, Claude Code, Claude Cowork, and now Amazon Bedrock on AWS. Anthropic says it is still working to restore access on Google Cloud and Microsoft Foundry as quickly as possible. No policy restriction is left on who can use Fable 5 itself on any surface where it already runs, today's gap on 2 of 5 delivery channels is a deployment step still in progress, not a new access gate. Counting from the June 12 suspension through the July 1 relaunch, that is 19 calendar days a frontier AI model spent completely dark by government order, the first shutdown of its kind in the industry.

What Fable 5 Is Actually Good For

Anthropic's launch data showed Fable 5 leading on nearly every capability benchmark it publishes, with the biggest gains showing up on long, complex tasks rather than quick single-turn questions. Stripe reported it completed a codebase-wide migration across 50 million lines of Ruby code in a single day, work that would have otherwise taken a full engineering team more than two months. Hebbia's finance benchmark for senior-level reasoning and IMC's trading-analysis evaluations both put Fable 5 at the top of the field.

For an Executive, that translates into a specific use case: reach for Fable 5 over the cheaper Sonnet 5 (covered elsewhere in this issue as the daily-use default) when a task involves multi-step reasoning across a large amount of material, such as reviewing a long contract for inconsistencies, analyzing a dense financial model, or running a research task that needs to hold context across dozens of documents. For quick drafting, routine email, or fast back-and-forth work, Sonnet 5 remains the more cost-effective choice.

Pricing sits at $10 per million input tokens and $50 per million output tokens, less than half the cost of the earlier Mythos Preview model. On Pro, Max, Team, and select Enterprise plans, Fable 5 is included at up to 50% of weekly usage limits through July 7, after which it draws from standard usage credits. This is not a general capability downgrade compared to Mythos 5. It is the same underlying model with a narrow set of cyber and biology related requests rerouted elsewhere.

Fable 5's Safety Layer Still Shapes How It Responds

No access restriction does not mean no filtering. Anthropic's own developer documentation confirms Fable 5 ships with safety classifiers that can decline a request outright rather than answer it, a behavior the API reports as a formal refusal rather than an error. When that happens, the request can automatically fall back to Opus 4.8 instead of dead-ending. Mythos 5 carries no such classifier layer at all: the same underlying model, the same 1 million token context window, the same $10 per million input and $50 per million output token pricing, just without the safety filter. That is the real difference between the two models today, not who can reach them, but how each one is allowed to respond once you do.

Mythos 5 vs. Sonnet 5: A Decision Almost No Executive Needs to Make

For the roughly 100 vetted Project Glasswing organizations that can request Mythos 5 through their Anthropic, AWS, or Google Cloud account team, the honest comparison is not Mythos 5 versus Sonnet 5. It is Mythos 5 versus Fable 5, and Anthropic's own guidance settles it directly: organizations without Mythos 5 access can use Fable 5, which offers the same underlying capabilities. Mythos 5 earns its cost for exactly one narrow job: dedicated offensive or defensive cybersecurity testing, where a legitimate security request might otherwise get flagged and rerouted by Fable 5's classifier, the same category of work the Amazon researchers were doing when they found the vulnerability that triggered this whole shutdown.

Outside of dedicated security research, Sonnet 5 (covered elsewhere in this issue) beats both Fable 5 and Mythos 5 on cost for routine drafting, analysis, and quick turnaround work: $2 to $3 per million input tokens against $10 for either Fable 5 or Mythos 5, a real gap for anything that does not need frontier-level, long-horizon reasoning. Reach for Fable 5 when the task itself demands that depth. Reach for Mythos 5 only if the business is a vetted Glasswing partner doing security work Fable 5's own safeguards would otherwise slow down.

The Business Risk Lesson

A frontier AI model going fully dark for 19 days by government order, with zero advance notice and no way for the vendor to soften the blow, is a genuinely new category of business risk. It is not a server outage that resolves in hours, nor a price hike with 30 days notice. It is a legal directive that took effect the same afternoon it arrived, with no path for Anthropic to appeal in real time.

Any Executive who has built a workflow, a product feature, or an internal tool that depends entirely on one specific frontier model, from one specific vendor, with no fallback, just watched a live demonstration of what happens when that single point of failure gets triggered. It took a government letter this time. It could just as easily be a vendor's own safety incident, a pricing change, or a regional access restriction next time.

The practical fix is not complicated. Know which models power which workflows in the business today. Keep at least one alternate model tested and ready to swap in during an unplanned outage. Document a "if this model disappears tomorrow" contingency for anything customer-facing. Nineteen days is a survivable disruption for most businesses, but not if the dependency gets discovered for the first time during the outage itself.

Action Steps Summary

1. Confirm Fable 5 access today. Check the Claude Platform, Claude.ai, Claude Code, or Claude Cowork account for Fable 5 availability, since it relaunched globally on July 1, 2026.

2. Match the model to the task. Route long, document-heavy, multi-step work to Fable 5, and keep quick daily drafting and routine replies on the cheaper Sonnet 5.

3. Watch the usage window. Fable 5 is included at up to 50% of weekly limits through July 7 on Pro, Max, Team, and select Enterprise plans, after which usage credits apply.

4. Skip Mythos 5 unless you do dedicated security research. It remains limited to roughly 100 vetted organizations, and Fable 5 now offers the same underlying capability, with no access restriction, for everything else.

5. Map the single-vendor risk. List every workflow that depends on one specific frontier model with no fallback, and identify a tested alternate for each within the next 30 days.