Anthropic’s Claude Fable 5 Goes Public, With Mythos 5 Still Gated

Anthropic released Claude Fable 5 on Tuesday. The public version of its most powerful AI, Fable 5 carries a new safety layer that routes high-risk queries to the older Claude Opus 4.8. Pricing sits at $10 per million input tokens and $50 per million output tokens, double the cost of Opus 4.8 and the highest list price on any major frontier model.

Anthropic says the classifier fires in less than 5% of sessions; most users will never notice the rerouting, but a small share will see answers from a less capable model on cybersecurity, biology, chemistry, and model distillation. The same week, Anthropic expanded Project Glasswing to roughly 150 vetted organizations across more than 15 countries, the program that gates the unrestricted Mythos 5. Same weights, two doors.

Two Models, One Set of Weights

Fable 5 and Mythos 5 share a base. Anthropic calls Fable 5 a “Mythos-class” model the company has made safe for general use; Mythos 5 is the same model with some of the safeguards lifted, available only to organizations Anthropic has already approved. The split is about access, not capability: “For most ordinary enterprise and developer tasks… Fable 5 performs effectively the same as Mythos 5,” VentureBeat reported.

Where the two diverge is cyber, biology, chemistry, and model distillation. Fable 5 refuses those questions and falls back to Opus 4.8. Mythos 5 answers them, and is rolling out to the existing Project Glasswing cohort first, with a broader “trusted access” program to follow. Anthropic is clear that the public release is provisional. “To release the model both safely and quickly, we’ve tuned these safeguards conservatively – they’ll sometimes catch harmless requests, though they trigger, on average, in less than 5% of sessions,” the company said in Anthropic’s full launch announcement for Claude Fable 5 and Mythos 5.

The 5% That Don’t Get Fable

In cybersecurity, biology, chemistry, and attempts to distill the model into another system, Fable 5 silently hands the request to Claude Opus 4.8. The user gets an answer. TechCrunch reported the carve-out as “at least 95% of Fable sessions running entirely on the model’s own responses.”

The trade-off is open. Anthropic has tuned the classifier to err on the side of caution, so some benign questions will also be rerouted, and the company has flagged this as known friction it plans to reduce: “We recognize that this will be frustrating to some users, and our aim is to reduce false positives as we update and refine the safeguards after launch.”

Vals AI’s SWE-bench Verified leaderboard shows the cost: on most coding tasks, Fable 5 leads by a wide margin, but on guarded categories its behavior is closer to Opus 4.8. Subscription access follows a tighter schedule:

1. Through June 22, 2026: Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost.
2. On June 23, 2026: Anthropic pulls Fable 5 from those plans; usage requires credits.
3. The API and consumption-based Enterprise plans have Fable 5 available immediately at the listed price.

A 30-Day Tax On The Strongest Public Model

Fable 5 and Mythos 5 share a price. At $10 per million input tokens and $50 per million output, that is double the cost of Opus 4.8 and the highest list price on any major frontier model.

It is also less than half the cost of Claude Mythos Preview, the gated version that ran the early Glasswing work. Same weights, different price tiers. A heavy engineering workflow at that rate can run into four or five figures in a month, and a heavy agent workload can spend the same in a day, with the Stripe example Anthropic cites as a useful reminder: a 50-million-line Ruby migration finished in a day would otherwise have taken “more than two months by hand.”

The other tax is data. Anthropic now requires a 30-day retention on all traffic to Fable 5 and Mythos 5, even for enterprises that previously had zero-retention agreements. The data is not used for training, and Anthropic says it is held only to “defend against complex and novel attacks, including new jailbreaks” and to “identify and reduce false positives.” TechCrunch noted the policy “could set an industry precedent in which access to increasingly powerful models comes with mandatory data-retention policies framed as a safety measure.”

Same price, more rules. Anthropic’s framing is that the friction is the cost of releasing a Mythos-class model in 2026 rather than waiting another year. Other labs will face the same trade-off on the same timeline. Anthropic shipped first.

What The Benchmarks Show

Fable 5’s lead is widest on coding. On Vals AI’s SWE-bench Verified leaderboard, Fable 5 scored 95.00% against Opus 4.8’s 88.60% and GPT 5.5’s 82.60%. On Anthropic’s SWE-bench Pro figure, Fable 5 reached 80.3% against GPT 5.5’s 58.6%, and on Cognition’s FrontierCode Diamond evaluation, Fable 5 hit 29.3% against Opus 4.8’s 13.4% and GPT 5.5’s 5.7%.

The pattern softens on finance. Vals AI’s Finance Agent v2 puts Gemini 3.5 Flash at 57.86%, Fable 5 at 56.31%, and Opus 4.8 at 53.92%; Fable 5 wins against the Claude model, but where it loses it loses by a thin margin against a model that costs less per task.

Stripe, the payments company, in Anthropic’s launch post.

Fable 5 compresses months of engineering into days. In our 50-million-line Ruby codebase, it did in a day what would’ve taken us more than two months by hand.

Cursor called Fable 5 the state of the art on CursorBench, and called it “a class of long-horizon problems that were out of reach for earlier models.” GitHub said Fable 5 “took on complex, long-horizon coding tasks with a level of autonomy and reliability that exceeded previous benchmarks,” and Crosby Legal reported that “in blind review, our lawyers found its redlines matched or beat our current model every time.” Vals AI’s SWE-bench summary puts Fable 5 ahead of Opus 4.8 on coding benchmarks while noting a high refusal rate on bio and cyber-related questions, the same reroute Anthropic describes.

Glasswing Grows To 150 Organizations In 15+ Countries

The Mythos 5 release runs through Glasswing. The program has been quietly finding software bugs since April, and Anthropic announced the expansion on June 2, 2026, taking the partner count from roughly 50 to approximately 150 vetted organizations. The new cohort covers sectors underrepresented in the first wave: power, water, healthcare, communications, and hardware.

The original 50-partner cohort was a who’s who. It included Amazon Web Services, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks; the new expansion, per TechCrunch citing the Financial Times, adds Okta, Samsung, SK Hynix, SK Telecom, NATO, and ENISA, with country coverage spanning:

• Australia
• Belgium
• Canada
• France
• Germany
• India
• Italy
• Japan
• The Netherlands
• New Zealand
• South Korea
• Spain
• Sweden
• Switzerland

Anthropic has not published the full roster, and TechCrunch noted it had reached out to confirm the country list. The list is reported, not officially confirmed. Anthropic’s own post covers the country count and the sectors, not the names.

The vulnerability counts are why the queue forms at the door. Anthropic said its initial Glasswing cohort has found more than 10,000 high- or critical-severity flaws since April. Cloudflare identified 2,000 bugs across its critical-path systems using Mythos Preview, with 400 rated high or critical and a false-positive rate the company described as better than that of human testers. Mozilla found and fixed 271 vulnerabilities in Firefox 150, more than 10 times the previous Firefox release’s count using an earlier Anthropic model, and Anthropic’s own scan of over 1,000 open-source projects flagged 23,019 potential vulnerabilities, with 6,202 estimated as high or critical and independent reviewers confirming over 90% of a 1,752-finding sample as valid.

The internal story on the Firefox numbers is at how Mythos found 271 Firefox vulnerabilities in a single pass, and the macro-financial read is at the IMF’s macro-financial cyber warning on Mythos.

Why Mythos Stays Scarce Even After Going Public

Anthropic’s framing is that the public version is the safer version, and the unrestricted Mythos 5 stays rare by design. The company has warned for months that Mythos-class models are close to general release elsewhere, with rivals likely to ship them without comparable safeguards. Fable 5 is the public-track piece; Project Glasswing is the defender-track piece.

Anthropic spelled out the position in the Glasswing expansion post. Same warning, same architecture.

Cheap, fast AI models with powerful cyber capabilities are around the corner.

Anthropic, in its Project Glasswing expansion post on June 2, 2026. The same week, OpenAI released GPT-5.5-Cyber to vetted defenders including Cisco, Intel, SentinelOne, and Snyk, and both sides of the AI cyber loop are now staffed. The launch also lands inside a new policy frame.

The same day as the Glasswing expansion, President Donald Trump signed an executive order titled “Promoting Advanced Artificial Intelligence Innovation and Security.” The order creates a voluntary framework for the federal government to seek up to 30 days of pre-release access to “covered frontier models,” with the National Security Agency holding final designation authority and the Treasury Department leading an AI cybersecurity clearinghouse. It is a narrowed version of an earlier 90-day review window the administration scrapped on May 21, 2026, after industry pushback, and the IPO context sits one day earlier: Anthropic’s $965 billion IPO filing.

The Friction Anthropic Is Willing To Absorb

Anthropic is open about the cost of the architecture. Fable 5 will refuse some benign requests, the 30-day retention will not be popular with privacy-sensitive buyers, and the price will price some teams out. The company’s framing is that these frictions are the price of releasing a Mythos-class model in 2026 rather than waiting another year to ship.

The early user signal is split. On Reddit, one Claude subscriber reported being “switched to Opus 4.8 for that response because of safety concerns,” and independent cybersecurity analysts have argued the classifier can produce false positives on legitimate defensive work. Anthropic’s stated plan is to “reduce false positives as we update and refine the safeguards after launch.” The model ships with the conservative tuning in place, and the next Mythos-tier release will be the first test of whether the architecture holds.

Frequently Asked Questions

How much does Claude Fable 5 cost?

Fable 5 lists at $10 per million input tokens and $50 per million output tokens, the same as Mythos 5. That is twice the cost of Claude Opus 4.8, the previous frontier Claude, and the highest listed rate among the major AI models in VentureBeat’s pricing snapshot. It is also less than half the cost of Claude Mythos Preview, the gated version that preceded it.

What is the difference between Fable 5 and Mythos 5?

Same underlying model, different access tiers. Fable 5 is the public version and routes high-risk queries in cybersecurity, biology, chemistry, and model distillation to Claude Opus 4.8. Mythos 5 is the same weights with those safeguards lifted, available only to organizations Anthropic has approved through Project Glasswing. For most enterprise and developer tasks, Anthropic says, the two perform effectively the same.

When can I use Fable 5 on my Pro, Max, Team, or Enterprise plan?

Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans through June 22, 2026. From June 23, 2026, Anthropic pulls Fable 5 from those plans and usage requires credits; the company has said it plans to restore Fable 5 as a standard subscription feature as soon as possible. The API and consumption-based Enterprise plans have Fable 5 available immediately at the listed price.

What does the 30-day data retention mean for my data?

Anthropic now mandates a 30-day retention on all Fable 5 and Mythos 5 traffic, even for enterprises that previously had zero-retention agreements. The data is not used for training. Anthropic says it is held only to defend against complex and novel attacks, including new jailbreaks, and to identify and reduce false positives. The policy could set an industry precedent for how access to increasingly powerful models is conditioned on data retention.