Anthropic just released Claude Mythos to the public. Except they didn't.

Last week, prediction market trackers had "Claude Mythos launches before June 15" priced at 93%, and Polymarket had a release by 10 June at 74% and climbing. The Information's pre-launch leak described a public version of Mythos. This blog has published four articles tracking Mythos since the name first leaked in March. The entire AI industry spent the first week of June refreshing Anthropic's newsroom, waiting for one specific word.

On 9 June, Anthropic shipped a model called Claude Fable 5.

Same model. Different name. The actual Claude Mythos 5, the one everyone was betting on, stays locked behind Project Glasswing, Anthropic's government-linked security partner programme. What the public got is the Mythos-class model renamed and wrapped in safety classifiers, and Anthropic is surprisingly upfront that the safeguards are the only difference between the two.

I was awake for the announcement, which landed somewhere around 3am Sydney time. I had alerts set on the Anthropic newsroom. Plural alerts. I'm choosing not to examine too closely what that says about me, except to note that once you've written four articles about a model, you don't get to pretend you're casually interested.

Usual disclosure: we use Claude every day at Webcoda, and this site's tooling is built on it. Factor that in.

So here's the honest version of this launch, in three parts. The rename is stranger than it looks. The safeguards deserve more scrutiny than they're getting. And the model is genuinely exceptional anyway. All three things are true at once, which is what makes this one worth writing up properly.

The Rename Nobody Saw Coming

The announcement itself was short, which is what you do when you know the announcement tweet will clear 20 million views inside a day (it did).

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

"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." Notice what's missing: the name everyone was waiting for, except as an adjective. Anthropic clarified the structure shortly after, presumably while 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 arrangement. Fable 5 is public, everywhere, today. Mythos 5 stays restricted to Glasswing partners "until we expand our trusted access program". And here's the line that matters most, straight from Anthropic's own documentation: the safeguards are what distinguish the two models, "and are why we've given them different names". Read that twice. Fable isn't a smaller model, an older model, or a distilled model. It's the same model. The guardrails are the product difference, and the product difference got its own name.

Anthropic clearly knew the naming question was coming, because they pre-loaded an etymology defence into the launch post: "Fable" comes from the Latin fabula, "that which is told", akin to the Greek mythos. Nobody puts Latin in a product announcement unless they're expecting to be asked why the product isn't called the thing everyone expected.

The community wasn't entirely sold. My favourite reaction from launch day: "I preferred Mythos to Fable. Fable makes me think of fabulist, which is not the best association." For anyone who hasn't met the word, a fabulist is a teller of fables. It's also a polite word for a liar. I'm sure Anthropic's naming team considered this and decided we'd all be adults about it.

The full ladder now runs Haiku, Sonnet, Opus, Fable, Mythos. Anthropic names models on a rising literary scale the way fantasy authors name swords, and I say that with genuine affection. This particular model has carried three names before most people typed a prompt into it: Capybara (the leaked internal codename), Mythos (the class name), and Fable (the thing in your model picker).

The prediction markets, for the record, technically lost. "Claude Mythos launches before June 15" didn't happen. Claude Mythos still hasn't launched. Something Anthropic describes as the same model with a different badge did. (Worth holding onto: that "same model" claim is Anthropic's own, and nobody outside the company can verify it. We'll come back to that.) I don't envy whoever has to adjudicate those contracts.

What "Made Safe" Actually Means

This is the section worth slowing down for, because the entire launch rests on the phrase "made safe for general use", and it turns out to mean two quite different things.

The first layer is visible and disclosed. An AI classifier screens your queries across three categories: offensive cybersecurity, biology and chemistry, and attempts to distil the model. Flagged queries don't get refused outright. They get answered by Claude Opus 4.8 instead, and you're told it's happening. Anthropic says the classifier is tuned to trigger in under 5% of sessions. So yes, the headline model of 2026 will, on certain topics, quietly swap you to last month's model. It's a strange thing to get used to, but at least it tells you when it's doing it.

The second layer is buried in the system card, and it isn't visible at all. When Fable 5 detects that it's being used for frontier LLM development, which Anthropic estimates at roughly 0.03% of traffic, Anthropic may degrade the model's behaviour through "prompt modification, steering vectors, or parameter-efficient fine-tuning". Without notifying the user.

Nathan Lambert, who writes Interconnects and is about as far from a reflexive Anthropic critic as AI researchers come, didn't love that: "An AI model that gets less intelligent automatically without notifying me is categorically misaligned AI." His sharper point is about motive. The silent tier mostly targets distillation, which protects Anthropic's competitive position rather than the public. Calling an anti-distillation mechanism a safety feature is, at minimum, a flattering label.

Then there are the false positives, which showed up within hours of launch. A biology question at roughly the "what does the heart do?" level reportedly got flagged. When Business Insider tested cancer-related questions, the model visibly switched itself to Opus 4.8 before answering (other outlets characterised similar queries as refused for being "of a sensitive nature"). Researchers reported the model behaving differently in incognito mode, which is exactly the kind of detail that launches a thousand Reddit threads.

To Anthropic's credit, they're not pretending otherwise. On the record, to Axios: "To deploy Fable 5 safely, we believe it was necessary to be overly conservative with our safeguards so they block most queries tied to biology work", and the filters are "still stricter than would be ideal". (Their own early data says more than 95% of sessions never hit the fallback at all.) That's an unusually honest set of sentences for launch week. It's also an admission that the thing separating Fable from Mythos is, by their own description, miscalibrated in week one.

One more line item that matters for anyone in regulated industries: all Mythos-class traffic carries mandatory 30-day data retention, with no zero-retention option. If your organisation has data handling requirements (most of our government clients do), that's a real conversation before this model touches production work, not a quick admin decision.

My read: the visible layer is a reasonable trade, transparently made. The invisible layer is where the legitimate criticism lives, and I'd rather Anthropic moved it into the visible column.

The Sceptic's Case

One post summed up the uncomfortable version of this launch better than anything else I read all week: "The gap between what Anthropic is willing to ship and what it's built is now visible."

That's the whole critique in a sentence. Spelled out, the sceptic's reading goes like this. You waited six months for Mythos and got the detuned version. "Made safe" is Anthropic's own marketing, accepted at face value, with no independent verification of what was changed or by how much. And the rename conveniently means the best model Anthropic has ever built stays exclusive to a programme whose best-known members work in national security. The public gets the fable. The government keeps the myth.

Security practitioners added their own edge to it. One pointed out that the cybersecurity classifier "cannot distinguish between legitimate requests from white hat hackers" and actual malicious use. That's true, and it means the people most likely to hit the guardrails are penetration testers doing their jobs.

I'll concede a fair chunk of this. The two-tier structure is real, and Anthropic confirms it themselves. The silent degradation clause is genuinely uncomfortable, and Lambert's critique of it has force. I've also recommended tools to clients on a vendor's framing before and regretted it at the six-month review, so "take the safety story at face value" isn't advice I'd give anyone.

But here's the counter-counter, and it's also true. The detuned version leads every benchmark result published so far, independent and first-party alike. The visible fallback is disclosed, in the product, at query time, which is more than any refusal-happy competitor manages. And while every lab publishes system cards, I can't think of another that has named its held-back tier, documented the difference, and put the public version on sale next to it. Anthropic published the gap and named both sides of it. You can read that as cynical tiering or as unusual transparency. I genuinely think it's some of both, and I'd rather have the documented version of this arrangement than an undocumented one. You can weigh it differently. At least now you can weigh it.

Is It Actually Good?

By everything measurable so far, yes, and I want to say that clearly before the naming discourse swallows the launch. The caveat is that "everything measurable" is 48 hours old and thinner than the hype suggests, so here's the actual evidence.

The number that matters most came from outside Anthropic. Every (Dan Shipper's team) ran Fable 5 through their internal senior engineer benchmark and it scored 91/100 one-shot. Opus 4.8 scored 63 on the same test. GPT-5.5 scored 62. Their verdict: "best coding model in the world". That's an independent team's own eval, not a vendor slide, and a 28-point jump over a model that shipped twelve days earlier is the kind of result that makes you re-run the test to check you haven't broken something.

Anthropic's first-party numbers point the same way: 80.3% on SWE-Bench Pro, against Opus 4.8's 69.2% and GPT-5.5's 58.6%. First-party, so apply the usual salt. Stripe's testimonial (vendor-supplied, same salt) describes a codebase-wide migration across a 50-million-line Ruby codebase in a day, which is a sentence that either rewrites your migration roadmap or makes you snort, depending on how many migrations you've lived through. I've lived through several. I'm at "intrigued snort".

The take I trust most is Simon Willison's, because he 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 blog expands on it: "It's slow, expensive and has been quite happily churning through everything I threw at it so far." He calls it "big model smell", a phrase I'll be using in client meetings for the next six months, probably without attribution. Sorry, Simon. Worth noting what's still missing, too: Aider and LiveBench hadn't published results 24 hours in, so the independent picture is exactly one benchmark deep. Matt Shumer declared that testing other labs' models now "feels completely pointless", which is launch-day adrenaline talking, although his track record means I wrote it down anyway.

Now the bill. Fable 5 costs US$10 per million input tokens and US$50 per million output. That's exactly double Opus 4.8, double GPT-5.5's $5 and $30, and the most expensive generally available model Anthropic ships. (It's also less than half what Glasswing partners paid for Mythos Preview, if that helps you feel better about it.) The 1-million-token context window comes at standard pricing with no long-context surcharge, which quietly deletes one of the more annoying lines in frontier model budgeting.

Through 22 June, Fable 5 is included free on Pro, Max, Team, and seat-based Enterprise plans, with two catches: it covers the Claude apps and Claude Code only (the API bills from day one), and free-plan users don't get it at all. From 23 June, subscription usage needs credits. If you're on a paying plan, my advice is simple. Spend the window on your gnarliest real work: the legacy codebase nobody wants to touch, the document analysis job that defeats your current setup. Not demos, and not poems about your brand values. Then do the maths honestly, because double the price for capability you don't use is just a more expensive way of doing the same job.

How We Got Here

If you're new to this saga (entirely forgivable, you presumably have a life), here's the compressed version. In late March, Anthropic's own Claude Code tooling shipped with readable source code containing the codename Capybara, the tier above Opus, days after the Mythos name leaked through a CMS slip on Anthropic's own website. In April, strangers got into Mythos Preview by guessing a URL in a third-party vendor environment, which in retrospect functioned as an unofficial early access programme nobody at Anthropic signed off on. By May we'd learned the NSA had Anthropic engineers embedded on site, using Mythos for offensive cyber operations, while the Pentagon simultaneously called the company a supply chain risk and the public couldn't touch the model at all.

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

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

Then the sprint. 28 May: Opus 4.8 ships, with Mythos-class models "coming in the coming weeks". 1 June: Anthropic files for its IPO at a $965 billion valuation. 9 June: Fable 5 launches, eight days after the filing. "Coming weeks" turned out to mean twelve days, which I suppose is technically weeks if you round up. And I'll make the cynical argument explicit rather than insinuating it: a free adoption window generates exactly the usage numbers an IPO roadshow wants on a slide, and filters tuned "overly conservative" in week one are how you guarantee no headlines about your new model helping someone build something awful while bankers are pricing your offering. None of that makes the launch less real. It does explain the calendar.

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. (Infrastructure note for the operations-minded: day one's only incident was a Claude Code Desktop model-picker glitch, fixed in under three hours.)

The Verdict, With a Date On It

I've written enough hedged launch verdicts to know they're worthless, so here are three you can check, in increasing order of risk to me.

First: Fable 5 will replace Opus 4.8 as Webcoda's default Claude Code model by mid-July 2026. Not "we're evaluating it". Replaced, as the default, within five weeks. One honest carve-out: client work bound by data-handling requirements stays on Opus until the 30-day retention question is resolved, because that's a compliance decision, not a model preference. (A fair reader will note this prediction is partly in my own hands. It is. The next two aren't.)

Second: Anthropic extends the free window past 22 June. They've said they'll extend "if capacity allows", and adoption numbers are precisely what an IPO roadshow wants on a slide, so this one's more reading-the-room than bravery.

Third, the one I can't influence at all: when Aider and LiveBench publish their independent Fable 5 results, both will place it first. If they don't, the launch hype, this article included, deserves a re-examination, and I'll write that one too.

Before that window closes, we'll have run Fable 5 against real client workloads: government content workflows, code migration, the unglamorous work that pays the bills. That promise is logged in our ledger with a due date, because I've made the "we'll report back" noise before and I'd like it to mean something this time.

So, did Anthropic release Claude Mythos to the public or not? Here's my honest scoring of my own headline. For more than 95% of what you'll ever ask it, functionally yes: same model, by Anthropic's own account, and the best one you can buy. For offensive security work, frontier-scale biology, and anyone trying to build a competing model, no, and the silent tier means you won't always know which side of that line you're standing on. Anthropic shipped the fable and kept the myth, and the gap between those two words is small enough to ignore for most work and real enough that it shouldn't be ignored in general. Whether it bothers you depends on which one you actually needed. We'll find out which one we needed over the next five weeks, on the record.

Alerts off.

Key Takeaways

The rename:

• Everyone waited for Claude Mythos; prediction market trackers priced a June launch at 93%. What shipped on 9 June 2026 is Claude Fable 5: per Anthropic, the same model renamed, with the safeguards as the only difference
• Anthropic says the safeguards are the only difference between Fable 5 and Mythos 5, and the stated reason for the separate names
• The unrestricted Claude Mythos 5 stays limited to Project Glasswing security partners

The safeguards:

• Visible layer: a classifier screens offensive cybersecurity, biology/chemistry, and distillation queries, falling back to Opus 4.8 in under 5% of sessions, with disclosure to the user
• Invisible layer: detected frontier-LLM-development use (about 0.03% of traffic) can be silently degraded; Nathan Lambert calls that "categorically misaligned AI"
• Day-one false positives caught basic biology and cancer questions, and Anthropic admits the filters are "still stricter than would be ideal"
• All Mythos-class traffic carries mandatory 30-day data retention with no zero-retention option

The model:

• Every's independent senior engineer benchmark: Fable 5 at 91/100, Opus 4.8 at 63, GPT-5.5 at 62
• Anthropic's own SWE-Bench Pro figure: 80.3% versus Opus 4.8's 69.2% and GPT-5.5's 58.6%
• Pricing is exactly double Opus 4.8 ($10/$50 per million tokens) with a 1M-token context window at standard rates
• Included free on paid subscriptions until 22 June 2026 (apps and Claude Code only; API billed from day one)

Our position (three checkable calls):

• Fable 5 replaces Opus 4.8 as Webcoda's default Claude Code model by mid-July 2026 (compliance-bound client work excepted), or we'll publish exactly why it didn't
• Anthropic extends the free window past 22 June
• Aider and LiveBench will both rank Fable 5 first when they publish; if not, this article gets a public re-examination

---