Access to some of the most capable AI systems ever built vanished almost as soon as it appeared. Then, quietly, it began to return. But not to everyone.
On June 9, Anthropic launched Claude Fable 5, a model the company described as a “Mythos-class” system made safe for general use. The same day it introduced Claude Mythos 5, the same underlying technology with certain safeguards removed, reserved for a narrow group of cyberdefenders and infrastructure operators through its Project Glasswing program. Within days both models were suspended. A U.S. government export control directive forced Anthropic to pull access while officials reviewed risks.
Now the picture has shifted again. As of this week Mythos 5 has been restored to a select group of trusted users. Fable 5 could follow within days if negotiations with federal agencies conclude successfully. The episode captures a new reality in artificial intelligence development. The strongest models no longer debut solely on technical merit. They arrive only after regulators sign off.
According to reporting in Digital Trends, the U.S. Commerce Department cleared the return of Mythos 5 for limited trusted customers. Commerce Secretary Howard Lutnick told Anthropic the company had made meaningful progress on government concerns. Final approval for broader Fable 5 access still requires input from the Pentagon and National Security Agency. Discussions continue.
The distinction between the two models matters. Fable 5 carries new safety classifiers. When a query touches high-risk topics such as certain cybersecurity techniques, biological weapons or chemical synthesis, the system falls back to Claude Opus 4.8. Anthropic says these triggers fire in fewer than 5 percent of sessions and the rate is falling as classifiers improve. Mythos 5 lifts some of those restrictions for approved partners working on defensive applications.
“For us, it’s really around what we call ‘race to the top,’ being able to provide this technology in a valuable fashion, and at the same time providing the right safety guardrails so that it can do asymmetrically more benefits than harm,” Dianne Penn, Anthropic’s head of product management for research, told CNBC.
Performance claims are striking. Fable 5 posted more than 10 percent higher scores than Claude Opus 4.8 on several benchmarks covering software engineering, complex knowledge work, vision tasks and scientific research. The advantage grows as problems become longer and more intricate. Early users described compressing months of engineering effort into days. One example cited by Anthropic involved migrating a 50-million-line codebase at Stripe in a single day.
In molecular biology the model generated novel hypotheses that human experts preferred 80 percent of the time in head-to-head tests. In genomics it assembled datasets across 138 species, trained a predictive model and produced results that outperformed a published paper in Science. Drug design tasks ran roughly 10 times faster while matching skilled human performance.
Yet capability is exactly why access remains gated. The same systems that accelerate legitimate research could, without controls, assist in offensive cyber operations or hazardous biological work. Anthropic built dedicated classifiers for cybersecurity exploits, agentic hacking, biology and chemistry pathways, and model distillation attempts. Red-teaming found the safeguards robust. No universal jailbreaks emerged.
The company also imposed a new 30-day data retention policy for business customers. Logs are kept for review but not used for training outside safety research. These steps helped persuade officials to begin restoring service.
Project Glasswing, the initiative that first received Mythos 5, partners with organizations protecting critical infrastructure. IBM participates and has highlighted the layered safety approach. “The capabilities of models like Fable 5 and Mythos 5 have the potential to do profound good for the world. We’ve seen the beginnings of this in Project Glasswing,” the official Anthropic announcement stated.
This isn’t isolated to one lab. OpenAI limited its GPT-5.6 preview to trusted partners while navigating similar government reviews. The pattern suggests frontier AI has joined a short list of technologies subject to pre-deployment national security scrutiny. Launches now involve engineering, safety testing and regulatory clearance in equal measure.
Anthropic priced the new models at $10 per million input tokens and $50 per million output tokens, less than half the rate of its earlier Mythos Preview. The company reported a revenue run rate of $47 billion and a valuation near $965 billion at the time of the June announcement. It has filed confidentially for an IPO.
Enterprise customers appear ready. Professional services giant Cognizant, one of Anthropic’s largest clients, plans to roll Claude across its 350,000 employees. Other large organizations have integrated the models into coding platforms, financial analysis and research workflows. Feedback from users at Cursor, GitHub and finance teams repeatedly praised autonomy, fewer hallucinations on complex tasks and the ability to reflect on and validate its own outputs.
But the suspension exposed fragility. Developers who had begun building on Fable 5 suddenly faced broken integrations. Some scrambled to revert to Opus 4.8. The episode served as a reminder that even paid enterprise access carries sovereign risk when models cross certain capability thresholds.
So what happens next? Anthropic says it intends to expand the trusted access program for Mythos 5 to additional biology researchers with appropriate safeguards lifted. Fable 5, once fully cleared, will roll out more broadly to paid subscribers and enterprise plans. The company continues refining classifiers to reduce false positives without weakening protections.
The broader industry watches closely. Rapid iteration remains the norm. Anthropic released multiple Opus and Sonnet updates throughout 2025 and 2026, pushing context windows to one million tokens and improving agentic performance. Yet each leap in capability invites tighter oversight.
Regulators appear to have settled on a strategy. Approve the version with stronger guardrails first. Monitor real-world use. Then consider relaxing controls for verified defensive applications. It is a deliberate pace that prioritizes containment of downside risk even as upside potential grows.
Anthropic’s models are making a comeback. The conditions attached to that return will shape not only who gets to use the technology but how the next wave of powerful AI reaches the market. The engineering race continues. The regulatory one has only begun.
WebProNews is an iEntry Publication