Anthropic releases Claude Fable 5, its first public Mythos-class model (2026)

Anthropic on Tuesday released Claude Fable 5, the first Mythos-class AI model made available to the general public — ending months of restricted access and bringing the company's most capable model tier to paying subscribers and API customers worldwide. 1

Simultaneously, the company began upgrading existing Project Glasswing partners to Claude Mythos 5 — the same underlying model but with cybersecurity safeguards removed for vetted defenders and infrastructure operators. 2

The launch comes eight days after Anthropic confidentially filed an S-1 registration statement with the SEC, positioning Fable 5 as a commercial inflection point ahead of a prospective IPO. 3

What Fable 5 can do

Anthropic says Fable 5's capabilities "exceed those of any model we've ever made generally available" and that it sits at or near the top of nearly every tested benchmark. The gap over prior Claude models widens with task length and complexity — by design, Fable 5 is optimized for work that runs for hours rather than minutes.

Software engineering is the clearest showcase. During early testing, Stripe reported that Fable 5 compressed months of engineering into days: in a 50-million-line Ruby codebase, the model completed a codebase-wide migration in a single day that would have taken a full team more than two months by hand. 1 On Cognition's FrontierCode evaluation — which tests whether models can pass difficult coding tasks while meeting production-codebase standards — Fable 5 scored highest among frontier models even at medium effort.

Benchmark table comparing Claude Fable 5 and Mythos 5 against other leading frontier models across coding, reasoning, vision, and knowledge tasks. — Benchmark comparison across frontier models. 1

Early customers offered concrete numbers across other domains:

Finance: Fable 5 set the top score on Hebbia's senior-level Finance Benchmark, with substantial gains in document-based reasoning and chart interpretation.
Analytics: One analytics platform said Fable 5 was the first model to break 90% on its core benchmark of complex, long-running analytical tasks — a 10-point jump over Opus 4.8.
Legal: Harvey reported that in blind review, Fable 5's contract redlines matched or beat their prior model every time.
Physics: One research team said Fable 5 nearly reached the output of GPT-5.5 (after four days of compute) in just 36 hours, while using a third of the reasoning tokens.

Vision received a significant upgrade. Fable 5 can extract precise numbers from complex scientific figures, rebuild a web app's source code from screenshots alone, and — in a demonstration Anthropic published — beat the video game Pokémon FireRed start-to-finish using only raw game screenshots, without any mapping tools or helper harnesses that prior models required.

Long-horizon memory is described as Fable 5's defining trait for agentic work. Dianne Na Penn, Anthropic's head of product management for research and labs, told Fortune: "Now you could use a model like Fable 5 to run very long-horizon projects, potentially overnight." 4 In internal tests using the deck-building game Slay the Spire, persistent file-based memory improved Fable 5's performance three times more than for Opus 4.8.

Safeguards that made general release possible

Mythos-class models were first announced in April 2026 and immediately restricted through Project Glasswing over concerns about their cybersecurity capabilities — specifically, their ability to find and exploit software vulnerabilities with minimal human guidance. 2

Fable 5 ships with a new layer of safety classifiers: separate AI systems that intercept potentially dangerous queries and reroute them to Claude Opus 4.8 instead. Three categories trigger automatic fallback:

Cybersecurity — any request related to offensive hacking, exploit development, or reconnaissance.
Biology and chemistry — most requests involving research in areas where Mythos-level capability could assist weapons design.
Distillation — queries flagged as attempts to extract Claude's outputs to train competing models.

Anthropic says more than 95% of Fable 5 sessions involve no fallback at all. The company also ran an external bug bounty program in which no tester found a universal jailbreak in over 1,000 hours of testing. UK AISI made partial progress toward one in a brief initial window, but no full bypass was confirmed. 1

Penn acknowledged the classifiers are deliberately conservative at launch: "We recognize that there might be some benign requests that end up being blocked initially. We're working actively on making those safeguards improvements post-launch." 4

Results of running cyber evaluations with Fable 5 — showing the classifiers prevent Fable from making any progress on offensive cybersecurity tasks. — Cybersecurity evaluation results: Fable 5 with classifiers active shows near-zero progress on offensive cyber tasks. 1

A new data retention policy accompanies the release: all traffic on Mythos-class models — on both first- and third-party surfaces — is subject to 30-day retention. Anthropic says the data will not be used for model training and will be deleted after 30 days in almost all cases; the purpose is detecting novel jailbreak patterns and reducing false positives.

Mythos 5: the unrestricted track

Claude Mythos 5 is the same model as Fable 5 but with its cyber classifier disabled. Access today is limited to organizations already enrolled in Project Glasswing — roughly 150+ cybersecurity firms and critical infrastructure operators across 15+ countries. Anthropic plans to expand Mythos 5 access through a broader trusted access program developed in consultation with the US government. 2

A parallel trusted access track for biology is planned: a small group of biomedical researchers will gain access to Fable 5 with its biology and chemistry classifiers removed. Anthropic demonstrated the underlying potential here: Mythos 5 accelerated aspects of the drug design process by approximately 10x internally, matching skilled human operators on protein design tasks without human assistance across 9 of 14 targets in one study. 1

Pricing and availability

Both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens — less than half the price of Claude Mythos Preview, which they replace. 1

Claude Fable 5 and Mythos 5 — pricing and availability

Input tokens (per 1M)	$10.00
Output tokens (per 1M)	$50.00
Price vs. Mythos Preview	< 50%
Sessions with no fallback	95%

Loading stats card…

Subscription plan access is rolling out in stages. From today through June 22, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23, access on those plans will shift to usage credits. Anthropic says it intends to restore Fable 5 as a standard subscription tier once compute capacity allows. API and consumption-based Enterprise customers have full access immediately via the model ID claude-fable-5.

IPO timing and competitive context

The Fable 5 launch is Anthropic's largest commercial release since Claude 3 and arrives at a moment of concentrated public company activity in AI. Anthropic filed a confidential S-1 with the SEC on June 1; OpenAI filed its own IPO paperwork the following week; SpaceX's market debut is scheduled for Friday. 3

Anthropic's run-rate revenue was reported at $47 billion in conjunction with its $65 billion Series H funding round closed in late May, suggesting the company is entering the IPO process with evidence of demand at the high end of enterprise AI. Fable 5 — at double the per-token cost of Opus 4.8 — is positioned as the product that justifies that pricing tier.