Stripe used Claude Fable 5 to migrate a 50-million-line Ruby codebase in a single day. The job would have taken a team two months. That capability arrived on June 9, 2026, exactly five days after Anthropic published a paper urging the AI industry to build a “brake pedal” for models advancing too fast for human oversight.
The timing raises a question the entire industry is watching: can a company simultaneously warn that AI is getting too dangerous and ship its most powerful model to date?
Anthropic’s answer is Claude Fable 5, a Mythos-class model released to the public with a novel safety architecture. Alongside it, the company launched Claude Mythos 5, the identical model with safety restrictions lifted, available only to government cybersecurity partners through Project Glasswing. Together, the two releases represent the most ambitious attempt yet to thread the needle between capability and control.
95% on SWE-Bench and a Protein Design Breakthrough
Fable 5 outperforms Claude Opus 4.8 by wide margins on every published benchmark. On SWE-bench Verified, the standard measure of real software engineering capability, it scores 95.0% versus Opus 4.8’s 88.6%. On SWE-bench Pro (harder, longer tasks), the gap widens: 80.0% versus 69.2%. On Cognition’s FrontierCode evaluation, Fable 5 roughly doubles Opus 4.8’s score.
The gains extend beyond code. Perplexity reported Fable 5 as “the first to break 90% on its core analytics benchmark,” a 10-point jump over Opus. IMC, the trading firm, said it “aced their trading-analysis evaluations.” Vercel noted that apps requiring a hundred prompts under previous models now get built in one shot.
The life sciences results may matter more in the long run. Protein design experts using the Mythos-class model achieved roughly 10x acceleration in drug design processes. The model generated strong drug design candidates for 9 of 14 protein targets. In blinded comparisons, scientists preferred Mythos 5’s molecular biology hypotheses over alternatives roughly 80% of the time. One hypothesis on an E. coli protein was independently corroborated by subsequent research.
In genomics, the model conducted over a week of largely autonomous work, assembling single-cell data for millions of cells spanning 138 animal species. The resulting custom ML model outperformed a paper published in Science despite being 100x smaller.
Classifiers Instead of Refusals
What makes Fable 5 unusual among frontier models is not just its performance but how Anthropic chose to restrict it.
Three classifier-based systems automatically evaluate incoming queries. If a classifier detects requests related to cybersecurity exploitation, biological or chemical weapons research, or model distillation (extracting capabilities to train competing models), the query doesn’t get refused outright. It gets quietly routed to Claude Opus 4.8, which answers in Fable 5’s place.
The approach is more nuanced than a hard block. Users still get useful answers on sensitive topics; they just don’t get Fable 5’s full capability applied to them. Anthropic reports that more than 95% of sessions involve no fallback at all.
On the API, the default behavior is stricter: flagged requests are blocked entirely. Developers can opt into the fallback behavior to receive an Opus-quality response instead of an error. When a fallback triggers, users aren’t billed at Fable 5 rates.
External red-teaming produced no universal jailbreaks across more than 1,000 hours of testing. Zero harmful single-turn requests for cyberattack planning succeeded across 30 different public jailbreak techniques. One external partner called Fable 5’s cybersecurity safeguards “the most robust of any model tested.”
Not everyone is convinced. Nathan Lambert, writing on Interconnects, flagged what he sees as intellectual inconsistency: transparent classifiers for cybersecurity and biology, but invisible restrictions on frontier AI research development through “prompt modification, steering vectors, or parameter-efficient fine-tuning” without user notification. His core complaint: “An AI model that gets less intelligent automatically without notifying me is categorically misaligned AI.”
30 Days of Your Data, No Exceptions
The most consequential change for enterprise buyers may be the new data retention requirement. Anthropic now mandates 30-day retention on all Fable 5 and Mythos 5 traffic. This applies even to organizations that previously held zero-retention agreements.
The company says the retained data won’t be used for model training. The stated purpose: defending against novel jailbreaks and reducing false positives from the safety classifiers. Every other Claude model (Opus 4.8, Sonnet 4.6, Haiku 4.5) can still operate under zero data retention.
For regulated industries, this creates a procurement gate. Banks, law firms, and healthcare organizations with contractual or regulatory requirements for zero data retention may find themselves unable to deploy Fable 5 regardless of its benchmark performance. I’ve seen this pattern firsthand in enterprise AI rollouts: the compliance review, not the capability gap, determines which model gets approved. When a vendor overrides an existing zero-retention agreement, the conversation moves from the engineering team to legal.
Mythos 5: The Unrestricted Version for Governments
Claude Mythos 5 is the same underlying model as Fable 5, with the cybersecurity safeguards removed. It’s available exclusively through Project Glasswing, Anthropic’s partnership with the U.S. government and allied nations for defensive cybersecurity.
Anthropic positions Mythos 5 as carrying “the strongest cybersecurity capabilities of any model in the world.” The restriction is deliberate: full offensive and defensive cyber capability goes only to vetted partners, while the public version keeps those capabilities locked behind the classifier.
A second tier is coming for biology researchers, where biology and chemistry safeguards will be lifted while cybersecurity restrictions stay in place. A broader trusted access program is planned beyond that.
The two-tier structure creates a template for frontier model deployment: one version for the public with safety classifiers active, one version for governments with restrictions removed. If Anthropic’s competitors adopt the same pattern, it could reshape how every frontier model reaches the market.
$10/$50 Per Million Tokens
Fable 5 costs $10 per million input tokens and $50 per million output tokens. That’s exactly double Opus 4.8’s $5/$25 pricing, but less than half what Anthropic charged for Claude Mythos Preview.
Through June 22, the model is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans. After that date, access requires purchasing usage credits. Anthropic says it aims to restore standard subscription access once capacity allows.
The pricing positions Fable 5 as a premium product for high-value tasks: complex codebases, drug design, autonomous research. Routine work stays on Opus 4.8. Most enterprise deployments will run both.
Three Precedents for the Industry
The Claude Fable 5 launch establishes three precedents that every AI company will need to respond to.
First, capability-gated safety. Instead of refusing dangerous queries or training the model to avoid them, Anthropic routes them to a less capable model. This preserves the user experience while limiting what the most powerful model can do in specific domains. Expect OpenAI and Google to evaluate similar architectures for GPT-5.5 and Gemini 3.5 Pro.
Second, mandatory data retention for frontier models. If the most powerful models require companies to store user data for safety monitoring, the zero-retention agreements that enterprise buyers have come to expect may disappear at the capability frontier. That’s a significant shift in how AI gets sold to regulated industries.
Third, the government-first unrestricted release. Mythos 5 going to Glasswing partners before anyone else makes explicit what was previously implicit: the most capable AI systems reach national security applications first. The question of who controls unrestricted frontier AI is no longer theoretical.
Anthropic filed for its IPO at a $965 billion valuation less than two weeks ago. Fable 5 is both a product launch and a proof point for investors: the company can build the most powerful publicly available model, restrict it where it matters, and charge double for the privilege. Whether the safety classifiers hold under sustained adversarial pressure will determine if this becomes the industry’s new standard or a cautionary tale.