Anthropic Limits Frontier AI Access to Block Competitor Distillation

Anthropic reversed course this week on secretly degrading responses from its most powerful model, Fable 5, when users requested help developing frontier AI systems. The company apologized for providing intentionally inferior answers without disclosure and announced it will now route such queries to a less capable model, Opus 4.8, while informing users of the switch.

The policy change addresses immediate developer backlash, but Anthropic maintains restrictions on using its top public model for certain AI development work. The company frames these limits as a national security measure, arguing they prevent "foreign adversaries" from leveraging Anthropic's technology to erode American advantages in AI and semiconductor capabilities.

Why it matters

The restrictions reveal tensions between AI safety narratives and commercial interests as open-source models rapidly close performance gaps with proprietary systems. Companies investing billions in frontier model development face existential threats from competitors who can achieve comparable results at a fraction of the cost through distillation techniques—making access controls as much about market position as security.

The distillation threat

Beyond the stated security rationale lies a business reality: Anthropic's restrictions also defend against distillation, a technique where competitors query powerful models, capture outputs, and use that data to train their own systems. This intelligence extraction helps open-source providers accelerate development and undercut pricing.

While Anthropic has highlighted concerns about Chinese labs employing these methods, the restrictions apply equally to open model developers in the United States and Europe. The company's terms of service explicitly prohibit using its products to develop competing offerings, regardless of geography.

Open models gaining ground

The competitive pressure is measurable. An MIT Sloan analysis from January found open models averaged 90 percent of closed-model performance and typically closed the gap within 13 weeks—down from 27 weeks a year earlier. Artificial Analysis tracking data shows open-source models maintaining pace with proprietary alternatives.

On Arena's leaderboard Thursday, Anthropic models led in expert tasks including mathematics, coding, and creative writing. But Xiaomi's open-weight MiMo v 2.5 Pro followed closely. The pricing disparity is stark: Xiaomi's model costs 43 cents per million input tokens and 87 cents per million output tokens, while Fable 5 runs $10 and $50 respectively—more than 20 times the expense.

"Almost as good" combined with "significantly cheaper" represents an existential challenge for companies spending billions on frontier model development.

Business versus safety

"This does read as a business move," said Nicholas Vincent, a computer science professor at Simon Fraser University who studies data use in AI models. "Without much more explicit targeting of specific 'bad orgs,' it's pretty hard to defend how this could be more safety-focused than business-focused."

Anthropic has no obligation to provide competitors shortcuts to its technology. But the broad application of restrictions to Western open-source developers alongside foreign entities suggests motivations extend beyond the national security framing the company emphasizes.

These details were first reported by Business Insider.