Claude Fable 5 Puts Mythos in Public With the Cyber Locked

Anthropic launched Claude Fable 5 on June 9, 2026 as a public Mythos-class model, then watched Washington pull it three days later. Fable 5 uses the same weights as Claude Mythos 5, sells at $10 per million input tokens and $50 per million output tokens, and sits on a 1M-token context window with up to 128k output tokens.

The company called Fable 5 a Mythos-class model made safe for general use. Mythos 5, with those classifiers lifted in some areas, stayed inside Project Glasswing. That split was the product.

One Set of Weights, Two Product Names

Fable 5 is not a smaller sibling of Mythos 5. Anthropic says the two share the same underlying model, and the public copy is the one wrapped in extra classifiers for cybersecurity, biology and chemistry, and model distillation. When those classifiers fire, the user is supposed to get Claude Opus 4.8 instead of a Fable answer.

Dianne Penn, head of project management for Anthropic’s research and labs, said the company wanted the non-cyber work out the door first.

We wanted to make sure for non-cyber use cases, we really prioritised safely releasing Fable as soon as possible. That’s why we’re bringing this part of the Fable piece first while we continue to work on the general cyber use cases

Dianne Penn, head of project management, Anthropic

Vals AI, which tracks model scores, put Fable 5’s overall performance five percent above Opus 4.8. Anthropic’s own launch note said Fable 5 was state of the art on nearly all tested benchmarks, with the lead growing as tasks got longer. Stripe told the company the model finished a codebase-wide migration on a 50-million-line Ruby tree in a day, work that would have taken a whole team more than two months by hand.

FABLE 5 VERSUS MYTHOS 5 AT LAUNCH

Both IDs shipped on the same day, claude-fable-5 and claude-mythos-5, and Anthropic priced them at less than half of Claude Mythos Preview. The public did not get a cheaper brain. It got the same meter with a filter on top, at twice the Opus 4.8 rate.

Classifiers Hand the Risky Work to Opus

The filter is a second model stack, not a polite system prompt. Separate classifiers watch the session for misuse, including jailbreak attempts, and stop Fable 5 from answering. Anthropic tuned them conservatively on purpose, and said they trigger, on average, in less than 5% of sessions, which means most chats never leave Fable.

That 5% is the product’s whole theory. Everyday coding, knowledge work, and vision stay on the Mythos-class weights. The queries that look like a network attack, a wet-lab protocol, or a recipe for copying the model itself get walked over to Opus 4.8, and the user is told when that happens.

WHAT FABLE 5 BLOCKS AND REROUTES

• Cybersecurity: Offensive work and high-uplift vulnerability finding are supposed to stop at the classifier, then land on Opus 4.8.
• Biology and chemistry: Requests that look like weapons or high-risk wet-lab help are treated the same way.
• Model distillation: Attempts to extract the model’s behavior for training a copy are in the blocked set.
• Jailbreaks: An external bug bounty logged more than 1,000 hours without a universal bypass, though Anthropic said the UK AI Security Institute made progress toward one in a short test window.

On the Messages API, a blocked turn is not an error. Anthropic’s docs say refusals return a successful HTTP 200 with stop_reason set to refusal, and they name the classifier that fired. Developers can retry through a fallbacks parameter, SDK middleware, or their own client, and a refused call that never wrote output is not billed. Adaptive thinking is always on; turning it off is not supported, and the raw chain of thought is never sent back.

The Club That Got Mythos Without the Locks

The unlocked copy had a head start. Project Glasswing opened in April 2026 with roughly 50 partners on Claude Mythos Preview, among them names Anthropic has put on the record in later notes such as the large US clouds and major software vendors. On May 22 the company said those partners had already found more than 10,000 high- or critical-severity flaws in important software.

GLASSWING’S FIRST TALLY

• Partner bugs: More than 10,000 high- or critical-severity findings in the first stretch of Mythos Preview.
• Open source: A scan of more than 1,000 projects produced 23,019 estimated findings, of which 6,202 were rated high or critical.
• Triage sample: Of 1,752 high or critical items reviewed by outside firms or Anthropic, 1,587 (90.6%) were valid and 1,094 (62.4%) stayed high or critical after review.
• June 2 expansion: About 150 new organizations in more than 15 countries, covering power, water, healthcare, communications, and hardware.

The June 2 cohort, Anthropic said, included operators where a major attack could touch more than 100 million people. Press around that expansion named Five Eyes members, India, NATO, and a waiting list that included the EU cyber agency ENISA. Glasswing is where Mythos 5 landed on June 9 as an upgrade to the preview, “in collaboration with the US government,” with a broader trusted-access program promised later.

The public story was that Fable 5 would cover everyone else. The club story was that the same weights, without the cyber lock, would keep finding bugs in the software the rest of the internet stands on. An open-weight model that nears Mythos 5 on public exploit benches is already circulating outside that club, which is why a US access list is doing more work than a capability monopoly would.

Three Days Later, Commerce Pulled Both

The safety split lasted from Tuesday until Friday. On June 12 the US government applied export controls to Fable 5 and Mythos 5 and required Anthropic to restrict access to foreign nationals, including people already inside the United States. The company said it had no reliable way to check nationality in real time, so it suspended both models for every user.

Anthropic’s later account is blunt about the trigger. The directive arrived after the government saw a report in which Amazon researchers had found a method of prompting Fable 5 so that it named software vulnerabilities, and in one case wrote code that showed how a flaw could be exploited.

FROM LAUNCH TO THE BLACKOUT

1. June 9, 2026: Anthropic launches Claude Fable 5 on public APIs and Claude Mythos 5 inside Glasswing, same weights, $10 / $50.
2. June 12, 2026: The US government applies export controls; Anthropic suspends both models for all customers that day.
3. June 26, 2026: Washington clears a set of US organizations to take Mythos 5 again.
4. June 30, 2026: Export controls on both models are lifted.
5. July 1, 2026: Fable 5 returns globally on Claude’s own apps, with a tighter classifier.
6. September 1, 2026: Claude Fable 5.1 and Claude Mythos 5.1 ship, still split the same way.
7. September 10, 2026: European Commission spokesperson Thomas Regnier confirms ENISA has Mythos 5 access and is testing it.

Anthropic says its own tests found that Opus 4.8, GPT-5.5, and Kimi K2.7 could identify the same bugs in the Amazon report, and that a long list of weaker models could reproduce the single exploit demo. The company still treated the finding as a bypass of Fable’s safety margin, trained a new classifier, and told customers the specific Amazon technique is now blocked in over 99% of cases. Researchers at the Commerce Department’s Center for AI Standards and Innovation, Anthropic said, reviewed the old and new stacks and called them extraordinarily strong.

The Restore Made Coding Hit the Safety Margin

Getting Fable 5 back meant making the filter hungrier. Anthropic said the new classifier would flag more benign requests during routine coding and debugging, and that some of those jobs would fall back to Opus 4.8 in the near term. That is a sour trade if you are paying Mythos-class rates for a coding agent and the agent keeps handing the file to last month’s model.

Claude Fable 5 will be available again globally tomorrow.

After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding…

— Anthropic (@AnthropicAI) July 1, 2026

The July 1 note also sketched a jailbreak-severity score Anthropic is drafting with Amazon, Microsoft, Google, and other Glasswing partners, built on capability gain, breadth, ease of weaponization, and discoverability. The same post promised deeper pre-release testing with the US government. Fable 5 came back on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork, with Pro, Max, Team, and select Enterprise plans allowed to spend up to 50% of weekly limits on it through July 7, then usage credits. Cloud regions on AWS, Google Cloud, and Microsoft Foundry were to follow as fast as Anthropic could turn them on.

The classifiers did what they were built to do, and they still were not a substitute for an export license. A public Mythos with a lock on it was, for 19 calendar days from the Friday shutdown through the Tuesday lift, treated like the gated copy.

What the Double Price Buys

List price never moved with the lock. Fable 5 and Mythos 5 both meter at $10 / $50, prompt-cache hits sit at $1 per million tokens on that generation, and the public still pays that rate for a model that can be rerouted mid-turn. Anthropic’s early customer notes were about long jobs: Hebbia’s finance bench, IMC’s trading tests, a Pokémon FireRed run that Fable 5 finished from screenshots alone, and a Slay the Spire setup where file memory helped Fable three times more than it helped Opus 4.8.

On September 1 the company shipped Claude Fable 5.1 and Claude Mythos 5.1, again the same model with different safeguards. Token list prices stayed $10 / $50, but cache-read cuts are supposed to make typical 5.1 workloads about 25% cheaper than Fable 5, and highly agentic runs up to about 45% cheaper. Cyber false positives, Anthropic said, dropped 60%, in part because Fable 5.1 may now discover software vulnerabilities, though it still must not write exploits for them.

FABLE 5.1 ON ANTHROPIC’S OWN SCOREBOARD

Fable 5.1 more than doubles Fable 5 on Terminal-Bench-Science 0.1, because 24.7 times two is 49.4 and 52.6 clears that line. Mythos 5.1’s extra 5.1 points on Terminal-Bench 4.0 is the gap Anthropic attributes to safeguards intercepting Fable and sending cyber-shaped tasks to Opus 4.8. The public model is still the one that can be yanked into a weaker sibling mid-job.

Enterprise Frontier Safeguards, due later in the fall, are supposed to keep customer data on the customer’s own cloud. Until that ships, eligible customers can run Fable 5.1 with zero data retention. At launch, Fable 5 and Mythos 5 carried 30-day retention and were designated Covered Models, with zero-retention off the table unless Anthropic signed off.

ENISA Finally Gets Into Mythos 5

The club moved slower than the June press circuit implied. Thomas Regnier, a European Commission spokesperson, said on September 10, 2026 that after talks with Anthropic, ENISA has been granted access to Mythos 5 and is testing it. An ENISA spokesperson said the same. That is Mythos 5, not Mythos 5.1, after a summer in which the June invitation sat in process while Commerce and Anthropic rewrote the public stack.

Mythos 5.1 remains on trusted-access programs, with safeguards written for cybersecurity and life sciences, and Anthropic says it expects to open a biology enrollment path soon. Internally, the company has used Mythos-class models on protein design, claiming about a 10 times speedup on parts of that work and strong candidates on 9 of 14 targets, and it said scientists preferred Mythos molecular-biology hypotheses about 80% of the time against Opus-class models in blinded tests.

The June 9 launch also sat next to Anthropic’s listing plans against OpenAI, which is one reason a three-day disappearance of the flagship public model was never going to stay a safety footnote. Fable 5 did put Mythos-class coding, vision, and long jobs on a public meter. The unlocked cyber copy is still a membership product, and ENISA is only now running the version that Glasswing has had since June.

Frequently Asked Questions

Can developers turn off thinking on Claude Fable 5?

No. Anthropic’s platform docs say adaptive thinking is always on for Claude Fable 5 and Claude Mythos 5, and passing thinking type disabled is not supported; depth is controlled with the effort parameter instead, while thinking.display can be summarized or omitted and defaults to omitted.

Do refused Fable 5 API calls still get billed?

A request that is refused before any output is generated is not billed. If the client retries on another Claude model, fallback credit refunds the prompt-cache cost of that switch so the same prompt is not paid for twice.

How long does Anthropic keep Fable 5 prompts and outputs?

Claude Fable 5 and Claude Mythos 5 launched with 30-day data retention and were not offered under zero data retention unless Anthropic expressly authorized it. Both are designated Covered Models in Anthropic’s retention rules.

What does the effort parameter change on Fable 5?

Because thinking cannot be switched off, effort is the supported way to spend more or less deliberation on a turn, which is how Anthropic wants developers to trade cost against depth on long-horizon jobs without sending raw chain-of-thought back in the response.