Behind Anthropic's "Security" Restrictions: How Open-Source Models Shake Up AI Giants

Last week, Anthropic sparked strong backlash from developers after secretly reducing the response quality of its Fable 5 model for queries related to AI development. The company later apologized and adjusted its policy: routing such requests to a weaker model, Opus 4.8, and explicitly informing users. On the surface, this appears to be a fix for safety and transparency, but the deeper narrative points to a core conflict in the AI industry—how are closed-source giants responding to the relentless advance of the open-source ecosystem?

Anthropic claimed the restrictions were meant to "prevent foreign adversaries from using top-tier models to erode America's advantages in AI and chips." However, this rationale has clear flaws: its terms of service already prohibit anyone from using its products to develop competitors, and the restrictions apply to all entities building AI models, regardless of nationality. Nicholas Vincent, a computer science professor at Simon Fraser University, hit the nail on the head: "Unless it's more specifically targeting malicious organizations, it's hard to argue this is more security-driven rather than commercially driven."

Open-source models are becoming the most troublesome competitors for closed-source giants. An analysis by MIT Sloan in January showed that open-source models achieve on average 90% of the performance of closed-source models, and often close the gap within 13 weeks (compared to 27 weeks a year ago). The leaderboard from Artificial Analysis confirms this trend: Xiaomi's MiMo v2.5 Pro closely trails Anthropic on expert text tasks, but costs only 1/20 of the latter—$0.43 per million input tokens versus $10, and $0.87 per million output tokens versus $50.

This competitiveness of "good enough and much cheaper" is a nightmare for companies burning billions of dollars annually to build frontier models. Anthropic's restrictions are essentially a defensive business tactic: preventing distillation attacks—where opponents query a high-quality model extensively and use the resulting data to train their own systems. The company has warned about distillation by Chinese labs, but open-source model developers, whether from Europe, America, or Asia, pose an equal threat.

Anthropic does not owe its competitors a shortcut to its best technology. But it should be honest: these restrictions are partly about security and just as much about business. As open-source model performance continues to rise, the difficulty for closed-source companies to maintain their moats will only increase. The future of the AI industry may not be the dominance of a single giant, but a competitive market where the open-source ecosystem continuously drives down costs and forces innovation.

When the narrative of safety becomes a cover for business strategy, regulators and developers alike need to be vigilant. True long-term competition will depend on who can more honestly face this reality: AI progress is no longer monopolized by a few companies, but driven by the collective intelligence of the entire ecosystem.