Claude Fable 5 Is Here: Anthropic Just Shipped Its Most Powerful Model to Everyone

On June 9, 2026, Anthropic released Claude Fable 5 — the first publicly available model in its Mythos class, the family of systems the company spent the spring insisting it had no plans to ship. Two months ago, Claude Mythos existed only as a leaked draft blog post that cratered cybersecurity stocks. Today, anyone with a Pro subscription can use a derivative of it for free until June 22. The release is simultaneously the most capable model the public has ever had access to and the most unusual product launch the AI industry has seen: a frontier model wrapped in classifier guardrails, a mandatory data retention policy, and a two-week free window with a hard cliff at the end. This is the launch story — what Fable 5 actually is, what the evidence says about the capability jump, and why Anthropic shipped its scariest technology days after warning the world about it.

From Leaked Secret to Public Product in 75 Days

The path to this launch did not follow any playbook. Claude Mythos was a secret internal model whose existence leaked on March 26, 2026 through an unsecured draft blog post — an operational slip that gave the market its first hint that Anthropic had built something categorically stronger than its public lineup. Cybersecurity stocks cratered on the news, on the inference that a model capable of what the draft implied would rewrite the economics of vulnerability discovery in both directions.

On April 7, Anthropic confirmed it. The company disclosed Claude Mythos Preview publicly, stated it had no plan to release the model, and announced Project Glasswing — a consortium spanning AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks — that would use Mythos to find and fix vulnerabilities in systemically important software before any broader access happened. The results, per Anthropic's June 9 announcement: more than 10,000 high- and critical-severity vulnerabilities found across roughly 50 partners, including vulnerabilities in every major operating system and every major web browser. Mozilla alone patched 271 Firefox vulnerabilities using Mythos Preview. In testing, the model wrote a working exploit for an already-disclosed Windows kernel vulnerability in 31 minutes.

The reaction reached well beyond the tech industry. Treasury Secretary Scott Bessent and Fed chair Jerome Powell convened financial executives within hours of the April reveal, according to press coverage at the time; JPMorgan, Goldman Sachs, and Citigroup began testing the model. Anthropic denied access to the Chinese government, and after European banks were also denied, Mistral AI began building a rival. That full security saga deserves — and gets — its own analysis, but the relevant point for the launch story is this: by June 9, the "patch first, release later" phase was declared sufficient, and Anthropic shipped two models. Claude Mythos 5, the unrestricted version, goes only to vetted cyberdefenders and infrastructure providers in collaboration with the US government. Claude Fable 5 — the same underlying model with safeguards on — goes to everyone.

What Claude Fable 5 Actually Is

Strip away the saga and Fable 5 is a frontier model with a clear spec sheet. It ships with a 1-million-token context window and a 128K maximum output — the largest output ceiling Anthropic has offered. API pricing lands at $10 per million input tokens and $50 per million output tokens: exactly double Claude Opus 4.8's $5/$25, and less than half what Mythos Preview cost the Glasswing partners who had early access.

On benchmarks, Anthropic's claim is sweeping: state of the art on nearly everything tested, with margins exceeding 10% over Opus 4.8 on some evaluations. The more interesting claim is the shape of the advantage. According to the announcement, the longer and more complex the task, the larger Fable 5's lead. Short, well-scoped tasks show modest gains over Opus-class models; long-horizon, multi-step work shows the gap widening. That pattern matters more than any single benchmark number, because long-horizon reliability is precisely the dimension where previous frontier models plateaued — and it is the dimension that determines whether a model can replace a workflow rather than a keystroke.

The customer evidence backs the long-horizon framing. Stripe says Fable 5 "compressed months of engineering into days" — the concrete example being a codebase-wide migration across a 50-million-line Ruby codebase completed in a single day, against an internal estimate of two-plus months for a team of engineers. Cognition reports the highest score ever recorded on its FrontierCode eval. In Anthropic's own demonstrations, the model rebuilt a web application's source code from screenshots alone, and completed Pokémon FireRed end-to-end using only vision input — a stunt benchmark, but one that requires sustained goal-directed behavior over thousands of steps without a human nudging it back on track.

The tool-vendor quotes follow the same theme. GitHub's chief product officer called it "a clear step forward," noting that its "autonomy and reliability exceeded previous benchmarks." Poolside's CTO went further: the model's "reasoning is a clear step beyond Opus 4.8" and "works at senior research scientist grade." On Hebbia's finance benchmark, Fable 5 posted the highest score of any model. None of these are independent academic evaluations — they are launch-day partner quotes, and should be weighted accordingly — but they are consistent with each other and consistent with the Glasswing evidence, which was produced over two months of real operational use rather than a demo cycle.

Beyond Code: The Science Results

The most under-discussed part of the announcement is the science portfolio. Anthropic reports that Fable 5 generated novel molecular-biology hypotheses that working scientists preferred roughly 80% of the time over Opus-class output, and conducted autonomous genomics research that outperformed published models from Science-journal papers — models 100x larger than the specialized systems they replaced. Anthropic's own protein-design experts say the model "accelerated drug design aspects by around ten times." These claims are self-reported and deserve external replication, but they explain something important about the release design: the biology and chemistry capability is strong enough that Anthropic classified it as a dual-use risk, which directly shaped the safeguard architecture discussed below.

The Strangest Release Shape in Frontier AI

A normal model launch is a pricing page and an API endpoint. Fable 5's launch has three structural features no frontier release has had before, and each one tells you something about what Anthropic believes it built.

The second feature is data retention. All Mythos-class traffic — Fable 5 included — carries a mandatory 30-day retention policy, and it applies even to enterprises that previously negotiated zero-retention agreements. Anthropic says the retained data is used for safety and security review only, not training, and that human access is logged. That is a meaningful trust posture, but it is also a unilateral change to existing contractual expectations, and compliance teams in regulated industries noticed immediately. For organizations weighing adoption, this is not a footnote; it is a data processing agreement review.

The third feature is the access window. Fable 5 is available via API immediately at the $10/$50 rate, and it is included free on Pro, Max, Team, and seat-based Enterprise plans from June 9 through June 22. On June 23 it moves to usage credits until capacity allows unmetered access to be restored. Two weeks of free frontier capability followed by a metering cliff is an aggressive adoption funnel — it gets the model into every workflow it can improve, then asks organizations to price what they have already become accustomed to. Teams that have lived through the enterprise token-budget collapse will recognize the pattern: capability first, bill later.

Why Ship the Scary Thing Now?

TechCrunch's framing of the launch was pointed: Anthropic released its most powerful model "days after warning AI is getting too dangerous." The tension is real. This is the company that spent April telling the world Mythos was too consequential to release, convening with regulators, and restricting access to a vetted consortium. Seventy-five days later, a safeguarded version is free with a Pro subscription. What changed?

Anthropic's stated answer is that the patch-first strategy worked: 10,000+ vulnerabilities fixed across the software that matters most, Glasswing expanding from ~50 partners to roughly 150 organizations across 15+ countries, and a safeguard stack that survived 1,000+ hours of external jailbreak attempts. On its own terms, that is a defensible sequence — harden the ecosystem against the capability before distributing the capability. Anthropic also reports that Mythos 5's misalignment levels are comparable to Opus 4.8, which is the company's argument that raw capability rose without a corresponding rise in autonomous misbehavior.

The business context supplies the rest of the answer. Anthropic confidentially filed its IPO S-1 on June 1 at a reported $965B valuation, riding a revenue run-rate of roughly $47B — up from about $10B a year earlier. A company heading into the public markets needs its growth story at maximum credibility exactly when Fable 5 shipped. We covered the filing dynamics in our analysis of the Anthropic and OpenAI IPO race, and the Fable 5 launch slots into that story precisely: it converts a secret capability overhang into recognized revenue and visible market leadership eight days after the S-1 went in.

Competitive pressure points the same direction. Mistral AI is building a Mythos-class rival after European banks were denied access — and every month Anthropic sat on the model was a month for competitors to close the gap without facing it in the market. There is also an argument Anthropic itself has made in other contexts: capability that exists will diffuse, and the party that diffuses it first gets to set the norms — the safeguard router, the retention policy, the vetted-access tier for the unrestricted version. Whether you find that reasoning principled or convenient probably depends on your prior about the company. Both readings can be true at once.

What's Genuinely New vs. Launch-Day Gloss

Every frontier launch arrives wrapped in superlatives, and a disciplined reader should separate the claims that come with operational evidence from the ones that are marketing by another name. Here is our honest sorting.

The honest summary: the long-horizon coding claims are the best-evidenced part of the launch, because Glasswing forced the model through two months of adversarial, operational, external use before anyone wrote a press release. The science and benchmark claims are plausible but currently rest on Anthropic's own reporting. The thing we are most confident about is the shape of the curve — every credible data point says the advantage compounds with task length, which is exactly what changes the economics of AI-assisted engineering. If your mental model of AI coding tools was set during the era we documented in the Claude Code vs. Codex tool war, Fable 5 is the first release since then that plausibly moves the frontier of what those tools can attempt, not just how fast they attempt it.

What This Launch Means If You Build Software

For working teams, three practical conclusions fall out of the launch shape. First, the June 9–22 free window is a genuine, time-boxed evaluation opportunity: every Pro, Max, Team, and seat-based Enterprise account can run Fable 5 against its hardest real workloads for two weeks at zero marginal cost. The right move is to point it at the tasks where Opus-class models currently fail — the multi-day refactors, the cross-repo migrations, the long-context analysis jobs — because that is where the model's lead is claimed to be largest and where 2x pricing would most easily justify itself.

Second, the safeguard router is a product consideration, not just a policy curiosity. If your application sits anywhere near security tooling, computational biology, or content that resembles either, a fraction of your traffic will be silently served by Opus 4.8 — under 5% of sessions on average, but not uniformly distributed across use cases. Anything that depends on deterministic model identity needs to account for that.

Third, the 30-day retention policy means regulated teams cannot treat this as a drop-in upgrade. If your zero-retention agreement was a compliance precondition, Fable 5 currently breaks it, full stop, until your legal team reviews the new terms. None of these caveats change the headline — the most capable model ever released is now broadly available — but they define the difference between adopting it deliberately and adopting it by accident.

Conclusion: The Overhang Is Public Now

For two months, the AI industry operated under a known capability overhang: everyone knew Mythos existed, almost no one could touch it, and every roadmap conversation carried an asterisk. That asterisk is gone. Claude Fable 5 puts Mythos-class capability — safeguarded, metered, retention- logged, but real — in the hands of anyone with an API key. The launch is simultaneously a triumph of the patch-first release strategy and a reminder that no frontier lab, however safety-branded, leaves its best model on the shelf when an IPO and a rival are both in motion.

The next two weeks will produce the first independent read on whether the long-horizon claims survive contact with the public's workloads. Our advice: be part of generating that evidence. The free window ends June 22, the usage-credit cliff arrives June 23, and the teams that ran structured evaluations during the window will be the ones making informed decisions when the meter starts.