Claude Fable 5 is here. We've been waiting since strangers broke into it.

Anthropic launched a new model overnight. For most people, that's just a Tuesday. For me, it's the end of a story I've been following since March, when the word "Capybara" turned up in leaked source code and nobody outside Anthropic knew what it meant.

I should be honest about the state I was in by launch day. I had alerts set on the Anthropic newsroom. Plural alerts. I'd been refreshing their blog since May like it owed me money. So when the announcement dropped on 9 June (US time, which meant somewhere around 3am here in Sydney), I was awake for it. I'm choosing not to examine too closely what that says about me. I'm exactly the person this launch was engineered for, and I'm aware that makes me part of the problem.

Here's why I care, and why you might want to. This isn't a normal model launch. We've covered this model through a source code leak, a break-in by strangers who guessed a URL, an NSA deployment, a Pentagon legal war, and a $965 billion IPO filing. All of that happened before the public could type a single prompt into it.

It's here now. It's called Claude Fable 5. And I've got opinions about all of it, including the name.

What Actually Launched

The announcement itself was short, which is what you do when 22 million people are going to read it anyway.

Claude

@claudeai

Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.

Its capabilities exceed those of any model we’ve ever made generally available.

82.3K

9 June 2026

The official line: "a Mythos-class model that we've made safe for general use. Its capabilities exceed those of any model we've ever made generally available."

Read that sentence again, because every word in it was chosen by someone whose job is choosing words. "Made safe for general use" is doing more heavy lifting than a forklift. The capability tier is Mythos-class. What we're getting is the version of that tier that passed the safety review. The unrestricted Claude Mythos 5 stays locked behind Project Glasswing, Anthropic's security partner programme, whose members have reportedly used Mythos to find more than 10,000 software vulnerabilities. So the most capable model Anthropic has built is still one you can't have. The most capable model you can have just changed dramatically, though.

Anthropic clarified the split themselves, presumably after watching the replies fill up with confused people:

Claude

@claudeai

Claude Fable 5 is available everywhere today. Claude Mythos 5 is restricted to Glasswing partners until we expand our trusted access program.

anthropic.com/news/claude-fa…

3.1K

9 June 2026

So that's the structure. Fable 5 is public, everywhere, today. Mythos 5 stays restricted to Glasswing partners "until we expand our trusted access program." Mythos-class is the tier. Fable is the variant you're allowed to touch.

Which brings me to the naming. This model has now had three names: Capybara (the internal codename that leaked in March), Mythos (the class name that leaked through a CMS slip on Anthropic's own website, also in March), and now Fable (the product you can actually buy). Three names before the public got to type a single prompt. Anthropic names models the way fantasy authors name swords, and I say that with genuine affection. At least Fable is easier to spell than Capybara.

One more detail worth sitting with. Anthropic spent five months telling everyone that Mythos-class models couldn't be released until cybersecurity safeguards were in place. Whatever safeguards weren't ready in January are apparently ready now. I'd love to see that Jira board. "Make frontier model safe for the general public" feels like the kind of ticket that gets dragged to Done with a certain amount of ceremony.

How We Got Here

If you haven't been following this saga on our blog (entirely forgivable, you presumably have a life), here's the fast-track version. And yes, this is at least the fourth article we've published that touches this model. I'm aware of how that sounds.

Late March 2026: the Claude Code source leak. Anthropic's own tooling shipped with readable source code, and inside it were animal codenames for unreleased model families, including Capybara, the tier above Opus. Days earlier, the Mythos name itself had slipped out through a CMS error on Anthropic's own website. Two security incidents in one week, from the company about to lecture everyone on cybersecurity safeguards.

Related Article11 min read

512,000 lines of Claude Code leaked via npm packaging error

Anthropic shipped their entire Claude Code source in every npm install. A missing config line exposed KAIROS, 44 hidden features, and a stealth mode...

Read full article

April 2026: strangers got into Mythos Preview by guessing a URL in a third-party vendor environment. Not a sophisticated state-sponsored attack. A guessed URL. The whole thing surfaced through Discord, of all places. In retrospect, that break-in functioned as an unofficial early access programme. Just not one Anthropic was running on purpose.

May 2026: we learned that roughly half a dozen Anthropic engineers were embedded inside the NSA, using Mythos for offensive cyber operations, while the Pentagon simultaneously maintained that Anthropic was a supply chain risk. Pentagon tech chief Emil Michael called Mythos "a separate national security moment." So the model you couldn't access was busy attacking foreign cyber defences. Typing that sentence still feels pretty insane.

Related Article11 min read

Anthropic filed a $965 billion IPO. The same week, the Pentagon said it still can't be trusted.

Anthropic is filing an IPO valued at $965 billion while simultaneously being designated a Pentagon supply chain risk. The same week they filed, the...

Read full article

28 May 2026: Opus 4.8 shipped, and Anthropic said Mythos-class models were "coming in the coming weeks." That turned out to mean twelve days, which I suppose is technically "coming weeks" if you round up.

1 June 2026: Anthropic filed for its IPO at a $965 billion valuation.

9 June 2026: Fable 5 launched publicly. Eight days after the IPO filing.

About that timing. When I wrote about the IPO last week, I quoted an observer who described Anthropic's recent run of moves as "not a leadership strategy, a launch window." I said at the time that was probably unfair. Shipping your biggest model ever eight days after filing to go public does rather support the theory.

Honestly, watching Anthropic this year has felt like watching an F1 team sandbag through pre-season testing. Everyone knew they had more pace. They just wouldn't show the lap. Now we've seen the lap.

The Early Verdict

Launch-day verdicts are launch-day verdicts, and I'll get to the caveats. But here's what the first 24 hours produced.

The most credible early take comes from Simon Willison, who tests everything, publishes his working, and doesn't do hype:

Simon Willison

@simonw

Wrote up my initial impressions of Claude Fable 5 - it has a big model smell: slow, expensive and capable of crunching through pretty much everything I threw at it simonwillison.net/2026/Jun/9/cla…

90

10 June 2026

His full write-up expands on it: "It's slow, expensive and has been quite happily churning through everything I threw at it so far." That's the most useful sentence anyone has written about Fable 5 so far, and I'll be using "big model smell" in client meetings for the next six months, probably without attribution. Sorry, Simon.

The enthusiasm end of the spectrum is louder. Matt Shumer, who's reliably early on major model releases, wrote: "I keep throwing harder and harder tasks at Claude Fable and it just keeps nailing them." He followed up by saying that testing other labs' models now "feels completely pointless." That's launch-day adrenaline talking, but Shumer's track record means it isn't nothing either.

Then there are the numbers doing the rounds. The most-shared figure on launch day was a 91/100 on a "senior engineer benchmark," and the most-shared version of it was posted in all caps by an aggregator account, which is rarely a great sign for data quality. The underlying number turns out to be real, though the attribution kept getting mangled: it's Every's internal senior engineer benchmark (Dan Shipper's team), not an official Anthropic eval. Their reported comparison: Fable 5 at 91, Opus 4.8 at 63, GPT-5.5 at 62. Anthropic's announcement carries it as a customer testimonial. A 28-point jump on someone's internal benchmark is genuinely striking, but it's still one team's internal benchmark. I'll wait for SWE-bench results and the independent eval shops before treating any single number as load-bearing.

On pricing: Anthropic's pricing page confirms Fable 5 at $10 per million input tokens and $50 per million output. That's exactly double Opus 4.8's $5 and $25, and less than half what Mythos Preview cost the Glasswing partners. The context window is 1 million tokens by default at standard pricing, with no long-context surcharge, which quietly removes one of the more annoying line items in frontier model budgeting.

The more interesting detail is the free access window, and it has a catch worth knowing about. Through 22 June, Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans, covering the Claude apps and Claude Code. It is not free on the API, which is billed from day one, and free-plan users don't get it at all. From 23 June, usage on subscription plans requires credits. So "free until 22 June" really means "free if you're already paying for a subscription." Still: giving every paying subscriber a two-week taste of your most capable model before the meter starts is a confident move. It's also precisely what you'd do three weeks before an IPO roadshow. Both things can be true.

Now the caveats. Every major model launch since GPT-4 has followed the same arc: 48 hours of euphoria, then a more textured reality. Opus 4.8 is the perfect recent example. The launch-day chatter said it had cured Claude's laziness; within a week, an independent eval firm found it performed worse than its predecessor on their agentic benchmarks. Both observations were true. They were just measuring different things.

I've also personally recommended tools to clients on launch-day enthusiasm before. The six-month review meetings that follow are character-building, and I'd rather not repeat them. So my official position on Fable 5's capability is: extremely promising, according to people I trust, pending actual work.

What This Means for Australian Businesses

Four practical things.

Use the free window. If you're on a Pro, Max, Team, or Enterprise plan, you've got until 22 June to throw your real workloads at the most capable public model on the market without paying anything beyond your existing subscription (API use is billed from day one, so do your testing through the Claude apps or Claude Code). Not demos. Not "write me a poem about our brand values." Your gnarliest actual problems: the legacy codebase nobody wants to touch, the document analysis job that defeats your current setup, the migration you've been putting off since 2023. Two weeks of free access to a frontier model is a genuinely rare evaluation opportunity. Don't waste it asking it to plan your meals.

Do the maths before you commit. Double Opus pricing means the "is it worth it" calculation actually matters now. For a lot of day-to-day work, Opus 4.8 or even Sonnet remains the right answer, and paying double for capability you don't use is just a more expensive way of doing the same job. The question isn't "is Fable 5 better?" It almost certainly is. The question is "is it better at the things I'm actually paying for?"

Think about the trust question, briefly. This is the same model class the NSA has been using for offensive cyber operations, with Anthropic engineers embedded on site. For business use, that cuts both ways. On one hand, "the intelligence community's offensive tooling" isn't the cosiest brand association. On the other, capability is exactly what you're paying for, and there's no stronger capability endorsement than a signals intelligence agency picking it for the hard stuff. I won't pretend this changes my purchasing decision. It probably shouldn't change yours. But you should at least know it.

We'll test it properly and report back. Webcoda will be running Fable 5 against real client workloads over the coming weeks: government content workflows, code migration tasks, the unglamorous stuff that pays the bills. I'll write up what we find honestly, including the parts where it's not worth double the money. If you'd like a practical read on whether it fits your team's work before the free window closes, get in touch.

Was It Worth Six Months of Drama?

The arc that started with a codename in leaked source code ends with a public launch. Capybara became Mythos became Fable, and along the way we got a break-in, a Pentagon ban, an NSA deployment, a federal judge using the word "Orwellian," and an IPO filing. As product launches go, it's the most dramatic one this blog has covered, and we once published an article about a Tamagotchi hidden in a coding tool.

Was it worth the wait? My honest answer: ask me in a month, when the hype has settled and the invoices have arrived. Launch-day me isn't a reliable narrator, and I know this about myself.

But I'll say this much. I've spent most of today throwing things at Fable 5 that I expected it to fumble, and it didn't fumble them. After six months of leaks, lawsuits, and locked doors, the most surprising plot twist of the entire saga might be that the model is genuinely as good as the drama implied.

I'll let you know when I'm sure. For now: alerts off. Job done.

Key Takeaways

The launch:

• Claude Fable 5 launched on 9 June 2026: a Mythos-class model that Anthropic has "made safe for general use"
• Claude Mythos 5 itself remains restricted to Project Glasswing security partners
• Anthropic says Fable 5's capabilities exceed any model it's ever made generally available

The backstory:

• The model first surfaced through the Capybara codename in the late-March source leak
• Strangers accessed Mythos Preview in April by guessing a URL in a third-party vendor environment
• The NSA was using Mythos for offensive cyber operations while the public waited
• Fable 5 launched eight days after Anthropic's $965 billion IPO filing

The practical bits:

• Pricing is exactly double Opus 4.8: $10/$50 per million tokens, with a 1M-token context window at standard rates
• Included free on Pro, Max, Team, and Enterprise subscriptions until 22 June 2026 (apps and Claude Code only; API billed from day one)
• Early verdicts are strongly positive on capability, with consistent notes on speed and cost
• The widely shared 91/100 benchmark is Every's internal eval, not Anthropic's; wait for independent benchmarks before treating launch-day numbers as gospel

---