On June 9, 2026, Anthropic released Claude Fable 5, the first Mythos-class model anyone can actually use. Until today, Mythos was locked behind a cybersecurity program called Project Glasswing, available to a small set of trusted partners. Now a safeguarded version is live on the Claude app and the API for everyone.
This is a big deal for a simple reason. Anthropic has been sitting on its most capable model for two months, refusing to ship it broadly because of misuse risk. Fable 5 is how they finally let it out: same brain as Mythos, with hard limits bolted on in a few dangerous areas.
If you build software, do research, or just want to know where the AI race stands in mid-2026, this one matters. Let me break down what it is, what changed, and what it means for you.
What is Claude Fable 5?
Claude Fable 5 is a Mythos-class AI model from Anthropic, released to the public on June 9, 2026. It’s the safeguarded twin of Claude Mythos 5, built on the same underlying model but with blocks on high-risk topics like cybersecurity and biology. Anthropic says its capabilities beat every model the company has previously made generally available.
“Mythos-class” is the new tier above Anthropic’s Opus models. The first one, Claude Mythos Preview, came out in April 2026 through Project Glasswing. Today’s launch adds two more: Fable 5 (public, with guardrails) and Mythos 5 (restricted, guardrails partly lifted).
The naming has a story. Anthropic says Fable comes from the Latin fabula, meaning “that which is told,” which links back to the Greek word mythos. So Fable and Mythos are the same model wearing two different names. The safeguards are the only thing separating them.
Here’s the short version of how the two compare:
| Claude Fable 5 | Claude Mythos 5 | |
|---|---|---|
| Who can use it | Anyone (public) | Glasswing partners + select bio researchers |
| Cyber safeguards | On | Lifted |
| Bio/chem safeguards | On | On (lifted for trusted bio program) |
| Underlying model | Same | Same |
| Pricing | $10 in / $50 out per million tokens | $10 in / $50 out per million tokens |
The story behind Mythos
To get why Fable 5 has guardrails, you need the backstory.
In April 2026, Anthropic unveiled Claude Mythos Preview and immediately said it would not release the thing broadly. The model was too good at finding and exploiting software vulnerabilities. Letting anyone use it could hand serious cyberattack capability to people who shouldn’t have it.
So Anthropic created Project Glasswing, a program that gave Mythos only to cyber defenders and critical infrastructure providers. The idea: use the model’s offensive skill to find security holes before attackers do. Last week, the company expanded access to hundreds of organizations across 15 countries, again focusing on groups that manage critical infrastructure.
The whole time, Anthropic said the eventual goal was to get Mythos-level capability to everyone, as long as they could build safeguards strong enough to prevent misuse. Fable 5 is that promise cashing out. The safeguards are now good enough, in Anthropic’s view, for a general release.
Key features of Claude Fable 5
Anthropic claims Fable 5 tops nearly every benchmark it tested. The pattern they point to: the longer and more complex the task, the bigger Fable’s lead over older models.
Coding that runs for hours
This is where Fable 5 shows off. During early testing, Stripe reported that Fable 5 compressed months of engineering into days. In one example, it handled a codebase-wide migration in a 50-million-line Ruby codebase in a single day, work that would have taken a whole team over two months by hand.
The benchmarks back the hype. On SWE-Bench Pro, a tough agentic coding test, Fable 5 scored 80.3%, against 69.2% for Opus 4.8, 58.6% for GPT-5.5, and 54.2% for Gemini 3.1 Pro. That’s a wide gap.
Reasoning at research-scientist level
Anthropic positions Fable 5 as strong on hard analytical work. On a knowledge-work benchmark called GDPval-AA, Fable 5 scored 1932, compared with 1890 for Opus 4.8, 1769 for GPT-5.5, and 1314 for Gemini 3.1 Pro. The finance and trading testers Anthropic quoted said it topped their senior-reasoning evaluations.
Vision good enough to skip the scaffolding
Fable 5 is Anthropic’s best vision model so far. It can pull precise numbers out of detailed scientific figures and rebuild a web app’s source code from screenshots alone. The fun demo: earlier Claude models needed a complex helper harness to play Pokémon FireRed, and Fable 5 beat the game with a vision-only setup.
Memory that holds across millions of tokens
Long tasks are where Fable pulls ahead. It stays focused across millions of tokens and improves its own work using notes it keeps. When Anthropic had it play the card game Slay the Spire with file-based memory, the memory boost helped Fable three times more than it helped Opus 4.8.
Why Anthropic added strong safeguards
Here’s the part that makes Fable different from a normal model launch.
A model this capable is dangerous in the wrong hands. Anthropic says that without safeguards, Fable 5’s cybersecurity skills could be misused to cause serious damage. So they shipped it with safety classifiers that watch for risky requests.
When the classifiers catch a request about cybersecurity, biology, chemistry, or model distillation, the response gets handled by Claude Opus 4.8 instead, and users are told whenever this happens. So you don’t get a flat refusal. You get a still-capable answer from a slightly less powerful model.
How often does this kick in? Anthropic says the fallback triggers in less than 5% of sessions on average, and more than 95% of Fable sessions involve no fallback at all. The company admits it tuned the safeguards conservatively, so they’ll sometimes catch harmless requests too. They plan to loosen them as they reduce false positives.
The testing was serious. Anthropic ran an external bug bounty with more than 1,000 hours of testing, and nobody found a universal jailbreak. One government testing body did make some early progress toward one, which Anthropic disclosed openly.
There’s also a new data rule. Anthropic now requires 30-day retention for all traffic on Mythos-class models. They say they won’t use it to train models, only to defend against novel attacks and cut false positives, and the data gets deleted after 30 days in almost all cases.
More Read
Claude Fable 5 vs Claude Opus 4.8
If you’re already using Opus 4.8, here’s how the upgrade looks. The numbers come from Anthropic’s own published comparison, so treat them as vendor benchmarks, not independent results.
| Benchmark | Claude Fable 5 | Claude Opus 4.8 |
|---|---|---|
| SWE-Bench Pro (agentic coding) | 80.3% | 69.2% |
| GDPval-AA (knowledge work) | 1932 | 1890 |
| Long-running, complex tasks | Bigger lead | Strong baseline |
| Cyber/bio safeguards | Yes (active blocking) | Standard |
| Price (input / output, per M tokens) | $10 / $50 | $5 / $25 |
| Availability | Public, phased on subscriptions | Public |
The headline: Fable 5 is clearly stronger, especially on long agentic work. But it costs double. Opus 4.8 runs $5 input and $25 output per million tokens; Fable 5 is $10 and $50. For quick everyday tasks, Opus 4.8 may still be the smarter spend.
What this means for students
For students, Fable 5 is a serious research and study companion, with one honest caveat: it’s expensive, and the free tier won’t include it long-term.
The stronger reasoning helps with dense work. Think working through a hard proof step by step, breaking down a research paper, or pulling structured notes out of a 200-page PDF. The improved memory means it can hold context across a long project without losing the plot.
If your university already runs Claude for Education, you may get access through that. Either way, the same study habits apply: use it to understand the reasoning, not to skip the thinking. (You can read more in our guide to free Google AI tools for students and our breakdown of the best AI study tools.)
What this means for developers
This is the audience Fable 5 was built for.
The long-horizon coding is the real story. Tasks that used to break older models, multi-step migrations, large refactors, debugging across a big codebase, are now in reach. Fable 5 is available through the Claude API as claude-fable-5, and also on AWS, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry.
The catch for devs: the cyber safeguards are blunt. If you do legitimate security research, expect the classifiers to block some work and fall back to Opus 4.8. Anthropic knows this is frustrating and says it’ll narrow the guardrails over time. For now, plan around it.
What this means for businesses
For companies, Fable 5 points at automating real knowledge work, not just chat.
The finance and legal testers Anthropic quoted were blunt about the gains. One legal team said its lawyers found Fable’s redlines matched or beat their current model in blind review. A spreadsheet-focused tester said it finished runs 25 to 30% faster than Opus 4.8. That kind of speed-up adds up across a team.
The trade-off is cost and the guardrails. The model you deploy is deliberately limited in a few domains, and the strongest cyber and bio capabilities belong to Mythos 5, which you can’t buy. For most business work, that’s fine. For regulated or security-heavy work, test before you commit.
Pricing and availability
Both Fable 5 and Mythos 5 cost $10 per million input tokens and $50 per million output tokens, which Anthropic says is less than half the price of Claude Mythos Preview. Even so, it’s still the most expensive of the major AI models available globally.
The subscription rollout is staged, so read this carefully if you’re on a paid plan:
- From June 9 through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost.
- On June 23, Anthropic removes Fable 5 from those plans. After that, using it needs usage credits.
- The company says it aims to restore Fable 5 as a standard subscription feature once it has enough capacity.
On the API and consumption-based Enterprise plans, Fable 5 is fully available now.
Industry reaction and why the launch matters
The timing tells you a lot. Fable’s launch comes as Anthropic prepares to enter the public markets, alongside OpenAI and SpaceX, and follows the company’s own plea urging major AI labs to set up a coordinated brake on frontier development.
That’s the tension worth watching. Anthropic warned that AI is advancing dangerously fast, then shipped its most capable public model days later. The company’s answer is the safeguard split: ship the capability, block the dangerous edges. Whether that holds up is the open question of 2026.
On the competitive side, the benchmark gaps over GPT-5.5 and Gemini 3.1 Pro are large enough that rivals will respond. The AI race isn’t slowing down. (We covered the broader picture in our piece on the 2026 AI model race.)
Potential challenges and concerns
A few honest concerns balance out the wins:
The price is steep. At double Opus 4.8, Fable 5 only makes sense for work that actually needs the extra capability. For routine tasks, you’re overpaying.
The guardrails are admittedly clumsy. Anthropic says benign requests will sometimes get blocked, and security researchers will feel this most. The fallback to Opus 4.8 softens the blow, but it’s still a worse experience for legitimate edge cases.
And the subscription situation is messy. Free access ends June 22, and the credit system after that adds friction. If you build a workflow on Fable 5 this week, budget for the June 23 change.
The future of the Mythos family
Fable 5 is the start, not the finish. Anthropic is already shipping Mythos 5 to existing Glasswing partners as an upgrade, with cyber safeguards lifted, and says it has the strongest cybersecurity capabilities of any model in the world.
The roadmap includes two expansions: a broader trusted-access program for cybersecurity organizations, and a new trusted-access program for biology, aimed at speeding up biomedical research with the bio safeguards removed for vetted labs. Expect Mythos-class capability to spread slowly and deliberately, gated by who Anthropic trusts.
Key takeaways
- Claude Fable 5 is Anthropic’s first public Mythos-class model, released June 9, 2026.
- It’s the safeguarded twin of Mythos 5, same model, with blocks on cyber, bio, chemistry, and distillation.
- It tops Anthropic’s benchmarks, with a wide lead on long-running coding and knowledge work.
- It costs $10/$50 per million tokens, double Opus 4.8, and is the priciest major model out there.
- Free subscription access ends June 22, after which you’ll need usage credits.
- The launch puts Anthropic ahead on benchmarks while testing whether “ship it with guardrails” is a safe enough answer for frontier AI.
FAQ
What is Claude Fable 5?
Claude Fable 5 is Anthropic’s first publicly available Mythos-class AI model, launched June 9, 2026. It’s the company’s most capable public model, built on the same base as Claude Mythos 5 but with safeguards on high-risk topics.
What does Mythos-class mean?
Mythos-class is Anthropic’s tier of models above its Opus class in capability. The first was Claude Mythos Preview in April 2026. Fable 5 and Mythos 5 are the next two.
Is Claude Fable 5 available to everyone?
Yes. It’s live on the Claude app and the Claude API. Free access on Pro, Max, Team, and Enterprise plans runs through June 22, after which it needs usage credits. Mythos 5, the unrestricted twin, stays limited to Project Glasswing partners and select biology researchers.
How is Claude Fable 5 different from Claude Opus 4.8?
Fable 5 is stronger, especially on long, complex coding and reasoning tasks, scoring 80.3% to Opus 4.8’s 69.2% on SWE-Bench Pro. But it costs double ($10/$50 vs $5/$25 per million tokens) and has active blocking on cyber and bio topics.
Why did Anthropic add safeguards?
Because Mythos-class models are strong enough at cybersecurity and biology to cause real harm if misused. The classifiers redirect risky queries to Opus 4.8 instead of answering, and trigger in under 5% of sessions.
Can students use Claude Fable 5 for studying?
Yes, for research, step-by-step problem solving, and working through long documents. The main limits are cost and the fact that free access ends June 22. Use it to understand reasoning, not to shortcut the learning.
Published June 9, 2026. Pricing, availability, and benchmark figures are from Anthropic’s launch announcement and may change. Benchmark numbers are Anthropic’s own published results.
