On April 7, 2026, Claude Mythos Preview was officially announced, but it was apparently too dangerous to release. According to Anthropic, Claude Mythos (they claimed that "the fallout—for economies, public safety, and national security—could be severe.")
Instead of releasing Mythos to the general public, they spun up Project Glasswing, a cybersecurity initiative that also involved some big-name companies. The idea was that they'd be able to deploy Mythos for defensive cybersecurity purposes and threw the whole web wide open.
Two months later, Anthorpic released a guardrailed version of Mythos—called Fable—that was fully available to the public... for about three days. Then the U.S. government told them they needed to , which realistically meant suspending access for everyone.
Claude Mythos 5 and Claude Fable 5 were (and still are, just unavailably) Anthropic's most capable models. They aren't designed for everyday use but for maximum performance in tasks like software engineering, scientific research, computer vision, and automating knowledge work. As , the longer and more complex the tasks, the more capable Mythos and Fable are relative to other models.
As a tech journalist, I've been covering AI and security (separately and together) for well over a decade. My read is that while there's a degree of marketing to Anthropic's claims, Mythos and Fable are the most powerful models ever released. When they were available, they jumped to , including Zapier's , which tests models on real-world workflows.
Table of contents:
What are Claude Mythos and Claude Fable?
Claude Mythos 5 and Fable 5 are general purpose that are particularly good at coding, reasoning, and vision-based tasks. The only meaningful difference between them is that Fable 5 has guardrails that prevent it from answering queries about cybersecurity, chemistry and biology, and distillation (, in an attempt to train cheaper models). Anthropic estimates that the guardrails kick in in around 5 percent of queries, and in those cases, it automatically falls back to Opus 4.8.
Mythos has been available since April 7, 2026. At the time, Anthropic claimed it was because it had already uncovered "thousands of high-severity vulnerabilities, including some in every major operating system and web browser."
That's why Mythos Preview was released to Project Glasswing partners. At the time, that included Amazon, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks—though as part of a wider rollout. Initial results suggest the project has been a big success. more than 10,000 high- or critical-severity security bugs as a result. While things didn't go full Skynet, it seems like Anthropic's slow rollout was the right approach.
Anthropic also planned to release a version with the biology and chemistry safeguards removed to scientists as part of a trusted access program in the coming months, but we'll see how that pans out given the current suspension.
Incidentally, in launching Mythos 5 and Fable 5 (the first numbered versions of the two models), Anthropic has made a complete mess of its model numbering convention. Its other three models are currently Opus 4.8, Sonnet 4.6, and Haiku 4.5. It's still better than what OpenAI is doing, but as someone who has to keep track of these things, it's deeply upsetting.
How good is Claude Fable 5?
There are two ways of looking at how good Claude Fable 5 is:
The benchmarks and claimed abilities
What it's like for real-world applications
First, the benchmarks: . In addition to taking the number one spot on Artificial Analysis, it set the highest score on 5 of the 10 underlying benchmarks. It's a solid step up from Opus 4.8, which is in second place.
As for the claims, Anthropic says that Fable 5 (or Mythos 5) was able to migrate a 50-million line Ruby codebase for Stripe compressing "months of engineering into days." It's proficient at both drug design and developing novel hypotheses in molecular biology, and successfully conducted a genomics project where it designed and trained a machine learning model that identified cells performing the same role across animal species. Most impressively, it beat Pokémon FireRed with its built-in vision capabilities and was three times as good at as Opus 4.8.
All this is to say that Fable 5 and Mythos 5 are legitimately setting the new frontier for AI models. But how does that translate into the real world?
For starters, Fable 5 took the lead on Zapier's AutomationBench, which tests AI models on end-to-end workflow execution using 47 real tools across six business functions. Each task drops an agent into an isolated environment (CRM records, inbox threads, calendars, and more) with the kind of ambiguity that makes real work hard. That's solid evidence that Fable translates into the real world.
Did you know? Every time a new major AI model is released, Zapier runs it through AutomationBench, an open benchmark that tests models on real business workflows.
I also tested Fable 5 myself, while it was still available, with a moderate code audit and refactor. I won't bore you with the details, but essentially, I'm making a WhatsApp pipeline I vibe-coded work reliably in 99% of cases—not just the easy 80%. With two prompts, it created a solid plan, implemented the first phase of a three-phase plan, and blew through my whole session usage and more than 10% of my weekly limit.
Bear in mind, I'm on the $100/month Claude Max (5x) plan, and this is what happened with two prompts. I can't say the work wasn't good—and it operates successfully with a degree of autonomy that no other model yet matches—but using it for anything but the most complex tasks is a massive waste of money.
Fable 5 took around an hour to burn through my session allowance. It was noticeably slower to make its plan than Opus 4.8. It might have gone a bit deeper on things, but not so much that it made a difference for the work I was doing.
My experience tracks with what I've seen from other AI experts. found that it was slow, expensive, and "happily churning through everything" he threw at it. found it broke their engineering benchmarks and "completed ambitious projects with little supervision," though testers burned through tokens and hit usage limits.
(For what it's worth, Artificial Analysis also found it was the most expensive model they'd ever tested on some of its benchmarks. On Humanity's Last Exam, Fable 5 scored 53.3% and cost $2,174 to run; third-place Gemini 3.1 Pro scored 44.7% and cost just $406.)
How dangerous is Claude Mythos?
with the launch of Glasswing, cybersecurity researchers at Anthropic detailed Mythos's cybersecurity capabilities.
As I said, I've been , and their summary seems legitimately concerning. In some of the worst exploits, under the right circumstances, a hacker could use Mythos to go from an unauthenticated visitor on a website to having admin access on the server.
The big takeaways are that Mythos sounds to be very good at finding and exploiting subtle vulnerabilities. It's also able to combine multiple vulnerabilities into more powerful exploits. In testing, it's found "zero-day" (or undiscovered) exploits in every major operating system and web browser that have been sitting there unfixed for years. The oldest bug it found was from the '90s in OpenBSD.
Mythos is significantly better at this than Opus 4.8 or Sonnet 4.6. The researchers go into a deep technical explanation of some of the exploits Mythos found, and it's worth reading if you're technically inclined.
The researchers stress that Mythos wasn't trained to have these capabilities. They emerged as an unintended consequence of its improvements in coding, reasoning, and autonomy. It's better at finding and fixing vulnerabilities too, but it's now good enough to exploit them as well.
While this information comes from Mythos Preview rather than Mythos 5, it's safe to assume its capabilities are only improving. The also back them up, with partners finding more than 10,000 serious security bugs in its first few weeks.
I'm not a biochemist, so I can't assess whether the fears about Mythos being used to develop novel viruses and biochemical weapons are overblown, but AI models have shown a huge amount of promise with drug development, protein folding, and other adjacent fields, so I wouldn't rule it out. There's also the fear that by lowering the barrier for entry, they'd essentially enable bad actors to vibe-compound dangerous chemicals.
I don't think Mythos 5 is going to , but there are certainly arguments that it needs stronger guardrails for general use—and the U.S. government certainly thinks so. It's also worth noting that Anthropic claims to have deliberately set the guardrails high with a view to reducing the number of false positives in the future after they gather more data.
How to try Claude Fable 5
Fable 5 was available for Claude Pro, Max, Team, and some Enterprise subscribers in the Claude app, Claude web app, and Claude Code CLI—as well as through the API (though at $10 per million input tokens and $50 per million output tokens, it cost twice as much as Opus 4.8).
But with the recent suspension, it's hard to say when it will become available again.
For now, you can , adding the power of Anthropic's AI into all your existing workflows. And with , you can go even further, giving Claude structured access to the tools and context it needs to take action across your stack. That means turning Claude from a smart assistant into an active part of your systems, triggering workflows, updating records, and orchestrating multi-step processes directly from your chat window, securely and at scale. Learn more about .
Related reading:
This article was originally published in April 2026. The most recent update was in June 2026 (twice!—once when Fable 5 was released and once shortly after access was suspended).
