Anthropic launched two models today. One you can use. One you probably can’t.
Claude Fable 5 is a Mythos-class model. The same beast that’s been powering Project Glasswing’s cyberdefense work. But with safety guardrails bolted on so regular people can use it without, you know, accidentally enabling a nation-state cyberattack.
Claude Mythos 5 is the same underlying model with the guardrails lifted in specific areas. It’s only available through Project Glasswing, in collaboration with the US government. Think of it as the unrestricted version for people who already have the clearance to handle it.
Same weights. Different rules. That’s the whole story in one sentence.
Here’s what you need to know.
The Safety Tradeoff
Fable 5 comes with a set of AI classifiers that watch what you’re asking. If you hit a topic in cybersecurity, biology and chemistry research, or model distillation, the query gets handed off to Claude Opus 4.8 instead.
You’ll know when it happens. The model tells you.
Anthropic says these classifiers trigger in less than 5% of sessions. They tuned them conservatively: better to block a harmless request than let a dangerous one through, and early results from 1,000+ hours of external bug bounty testing found no universal jailbreaks.
This is the same approach they’ve been building toward since Project Glasswing launched in April. The difference is scale. Mythos Preview went to a handful of partners. Fable 5 goes to anyone with a Claude account.
The catch: Anthropic also announced a mandatory 30-day data retention policy on all Fable 5 traffic, reported by TechCrunch. Even for enterprise customers who previously had zero-retention agreements. The data isn’t used for training. It’s for defending against novel attacks, but if you’ve been operating on a no-logs assumption, that just changed.
What Fable 5 Can Actually Do
The benchmarks are impressive. But the real-world stories are better.
Stripe put Fable 5 on a 50-million-line Ruby codebase and asked it to do a codebase-wide migration. The model did in a day what Stripe says would have taken a whole team over two months by hand.
Cognition’s FrontierCode evaluation, which tests whether models can pass hard coding tasks while meeting production-quality standards, has Fable 5 at the top of the leaderboard, even at medium effort.
Hebbia’s Finance Benchmark for senior-level reasoning: Fable 5 has the highest score of any model, with big gains in document-based reasoning, chart interpretation, and problem solving.
IMC tested it on trading-analysis evaluations and said Fable 5 aced them nearly across the board: factual lookup, conceptual reasoning, root-cause analysis, expected-value analysis.
The vision stuff is wilder. Fable 5 can extract precise numbers from detailed scientific figures. It can rebuild a web app’s source code from screenshots alone. Previous Claude models needed complex helper harnesses to play Pokémon FireRed. Fable 5 beat the game with a bare-minimum, vision-only setup.
On the memory side, it stays focused across millions of tokens. Give it persistent file-based memory and its performance on Slay the Spire improves three times more than Opus 4.8’s does. It reaches the game’s final act three times more often.
The Biology Research Is the Real Sleeper
The Mythos 5 side of this launch is doing work that doesn’t get the same press but might matter more in the long run.
Anthropic’s internal protein design team says Mythos 5 accelerated aspects of drug design by about ten times. In one test, the model handled protein design tasks: choosing binding sites, selecting tools, recovering from failures, tasks that normally require a trained scientist. Nine of 14 protein targets yielded strong drug design candidates.
It’s also the first model Anthropic has seen that consistently produces novel, compelling scientific hypotheses. In blinded comparisons, scientists preferred Mythos’s molecular biology hypotheses 80% of the time over Opus-class models. One hypothesis about an E. coli protein was independently corroborated by a lab working on the same problem.
And in genomics, Mythos 5 conducted a week of largely autonomous work: assembled single-cell data for millions of cells across 138 animal species, designed and trained a custom machine learning model to identify cells performing the same role in distantly related organisms. The resulting model outperformed one published in Science despite being 100 times smaller.
That’s not a benchmark. That’s actual science.
The Pricing Story
Fable 5 and Mythos 5 cost $10 per million input tokens and $50 per million output tokens.
That’s double Opus 4.8 pricing, but less than half what Mythos Preview cost. The economics make sense. This is a significantly more capable model at a lower premium than the preview tier.
The rollout has a weird shape though. From today through June 22, Fable 5 is included free on Pro, Max, Team, and seat-based Enterprise plans. On June 23, it switches to usage-credit billing. Anthropic says they’ll restore it as a standard subscription feature “as soon as possible,” which is corporate speak for “we’re still figuring out the pricing model.”
If you’re on Pro or Max, the smart move is to spend the next two weeks figuring out if Fable 5 is worth the premium for your workflow. You’ve got a free trial window.
What This Means
The dual-track model is the part I find most interesting.
Anthropic is shipping the exact same underlying model with the same weights and same capabilities through two different channels with two different safety profiles. One for the public with classifier guardrails. One for government and critical infrastructure with those guardrails lifted.
That’s a first. OpenAI doesn’t do this. Google doesn’t do this. Nobody has tried the “same brain, different operating rules” approach at this scale.
It’s also a real answer to the capability-vs-safety debate that’s been running in circles for years. Instead of holding back the model until safety catches up, Anthropic shipped both versions simultaneously. The public gets the capability with guardrails. The people who need unfettered access get it through a program designed for that.
All of this comes as Anthropic heads toward an IPO, one of two major AI labs that filed confidentially this week. The dual-track launch strategy makes more sense in that context. Public investors want to see both capability and safety. This is Anthropic showing it can deliver both.
Will the classifiers be too aggressive? Probably, for some use cases. The 5% trigger rate is self-reported, and “cybersecurity” is a broad category. If you’re a security researcher writing legitimate exploit analysis, you might find yourself talking to Opus 4.8 more than you’d like.
But the architecture is right. And in an industry where the standard move is “ship first, figure out safety later,” Anthropic choosing a third option is worth paying attention to.
I’ve been covering how the major AI labs approach these tradeoffs. This launch is the clearest signal yet that one of them has a real strategy for walking the line between capability and safety.
