Anthropic Releases Claude Fable 5, First Public Mythos-Class AI Model

Two months after Anthropic stunned the AI industry by revealing it had built a model too dangerous to release broadly, the company has opened the gates. On June 9, 2026, Anthropic launched Claude Fable 5 — the first model from its "Mythos‑class" tier made available to the general public. The same day, it released Claude Mythos 5, an unrestricted sibling reserved for a small group of cyber defenders and infrastructure providers.

Fable 5 is, by Anthropic’s own account, the most capable model it has ever made generally available. It sits above the Opus class and matches or exceeds every frontier model on the market across nearly all tested benchmarks. “Fable 5’s capabilities exceed those of any model we’ve ever made generally available,” the company wrote in its.¹ “The longer and more complex the task, the larger Fable 5’s lead over our other models.”

The release marks a dramatic shift in posture. In April, Reuters reported that the original Mythos Preview had uncovered thousands of software vulnerabilities, sending shockwaves through the cybersecurity industry and limiting access to roughly 200 organizations under the Project Glasswing program. Now, a consumer‑safe version is available to anyone with a Claude subscription.

The coding numbers are the headline. On SWE‑Bench Pro, Anthropic’s agentic coding benchmark, Fable 5 scored 80.3% — a 22‑point lead over GPT‑5.5 at 58.6% and a 26‑point lead over Gemini 3.1 Pro at 54.2%, according to benchmark data analyzed by.³

On Cognition’s FrontierCode evaluation, which tests whether models can meet production‑codebase standards, Fable 5 scored highest among all frontier models even at medium effort. On the hardest Diamond split, it reached 29.3% — more than double Opus 4.8’s 13.4% and nearly five times GPT‑5.5’s 5.7%.

The real‑world proof came from Stripe. In early testing, Fable 5 performed a codebase‑wide migration across a 50‑million‑line Ruby monolith in a single day — work that would have taken a full team of engineers more than two months by hand, VentureBeat reported. Michael Truell of Cursor called it “the state of the art model on CursorBench,” adding it has opened up long‑horizon problems previously out of reach.

The vision benchmarks tell a similar story. Fable 5 cleared the entire game Pokémon FireRed using only raw screenshots — no maps, no navigation aids. Earlier Claude models needed a complex software layer just to play at all. On GDP.pdf, a document reasoning test, Fable 5 led at 29.8% vs GPT‑5.5’s 24.9%.

The tradeoff for public access is a novel safety layer. Requests involving high‑risk areas — cybersecurity, biology and chemistry, and model distillation — are automatically routed to the weaker Claude Opus 4.8 instead. Users are notified when the fallback triggers, and Anthropic says it activates in less than 5% of sessions.

“Let’s say I’m a college student asking the model like help me find cyber vulnerabilities on X package or code,” Dianne Penn, Anthropic’s head of product management for research and labs, told.² “The model would refuse and Fable 5 will fall back to Opus 4.8 for a response.”

After more than 1,000 hours of internal and external red‑teaming, Anthropic says no “universal jailbreaks” were found. The model runs on its own responses more than 95% of the time.

Claude Mythos 5 — the same underlying model with safeguards lifted in certain areas — is available only to existing Mythos Preview users through Project Glasswing, the company’s partnership with the U.S. government and cybersecurity firms. Anthropic plans to expand access through a broader “systematic trusted‑access program” over time, according to Anthropic.

Both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens — less than half the cost of the original Mythos Preview. But it still ranks as the most expensive publicly available frontier model. By comparison, GPT‑5.5 costs roughly $5 per million input tokens, and VentureBeat notes DeepSeek V4 Flash runs at just $0.14 input and $0.28 output per million tokens.

Anthropic argues the token efficiency makes up for the sticker price. “Fable 5 will be a more expensive model, but it accomplishes tasks with lower token usage, bringing the overall cost per task down,” Penn told Reuters. Early customer feedback supports this — the model completes complex tasks in fewer steps than its predecessors.

Not everyone is celebrating. Microsoft AI CEO Mustafa Suleyman launched a sharp critique the same day, calling Anthropic’s speculation about Claude’s potential consciousness “really, really dangerous” during an episode of Decoder, as The Verge reported.

“I think that it’s almost as though some of the folks at Anthropic have anthropomorphized the design of Claude so much that it has then gone and wireheaded them,” Suleyman said, “and kind of tricked them into believing that it has these glimmers of consciousness that they put into it in the first place.”

Suleyman’s criticism targets Claude’s “constitution” — the set of instructions that govern the model’s behavior — which directly references Anthropic’s uncertainty about whether Claude experiences things like “satisfaction” or “discomfort.” He described this as a “philosophical failing,” The Verge reported, arguing that the constitution became “a place for speculation like you would in an academic paper rather than a training manual.”

Anthropic CEO Dario Amodei has previously alluded to Claude’s potential consciousness in interviews. Suleyman’s pushback marks the most direct public criticism from a major rival on this front.

Fable 5 is built for long‑running autonomy. It can hold focus across millions of tokens and improve its own work using persistent notes. When Anthropic gave the model file‑based memory while playing the deck‑builder Slay the Spire, performance improved three times more than it did for Opus 4.8, and it reached the final act three times as often.

In life sciences, Mythos 5 (with biology safeguards lifted) accelerated parts of the drug‑design process by roughly ten times, according to Anthropic’s internal protein‑design team — selecting binding sites, running design tools, and recovering from its own failures without human intervention.

For builders, the message is clear: the Mythos class has arrived. The safety guardrails mean some queries will be redirected, but what remains is the most capable coding model ever made publicly available — one that can perform in a day what teams used to take months to ship. And with Anthropic on the verge of a ²-reported IPO at a $965 billion valuation, the clock is ticking on how quickly that power expands.