Trump AI restrictions push enterprises toward open-source models

Key Takeaways

• Anthropic withdrew its Mythos 5 and Fable 5 models entirely rather than verify user eligibility under new restrictions
• OpenAI now requires federal government approval for GPT-5.6 access, adding uncertainty for enterprise customers
• Combined token share for Google, Anthropic, and OpenAI on OpenRouter dropped from 55% in January to 33% by June

The Trump administration's new restrictions on access to frontier AI models from Anthropic and OpenAI are pushing enterprises toward open-source alternatives. The shift is measurable: on OpenRouter, the combined token share consumed by Google, Anthropic, and OpenAI dropped from 55% in January to 33% by June 2026. China's DeepSeek now leads the platform.

The restrictions caught developers off guard. The White House has broadly favored tech deregulation, so limiting access to cutting-edge models marked a sharp reversal. For startups that built products around Anthropic's Claude or OpenAI's GPT, the policy created immediate operational risk.

What triggered the enterprise exodus?

In early June, Anthropic received orders to block non-U.S. users from its flagship Mythos 5 and Fable 5 models. Rather than build complex user verification systems, the company withdrew both models entirely. OpenAI followed a different path: it agreed to let the federal government approve customers seeking access to its newest GPT-5.6 model.

Both outcomes point to the same problem. Access to proprietary AI became unpredictable. Oren Michels, CEO of Barndoor AI, told AFP that businesses built around a single proprietary model face serious reliability questions when access can be interrupted without warning.

Haitham Mengad, co-founder of AI music startup Stems Labs, described Anthropic's Fable model as transformative for his company. Its sudden withdrawal convinced him that open-source alternatives deserved serious consideration. "It was a pivotal moment," he said.

How are Chinese open-source models filling the gap?

China's Zhipu AI released GLM-5.2 around the same time the restrictions hit. The model reportedly approaches the performance of Anthropic and OpenAI's leading offerings on several industry benchmarks. Because it's open-weight, enterprises can download it, customize it, and deploy it on their own servers without licensing restrictions.

AI analyst Andrew Curran told AFP that GLM-5.2 simultaneously lowers costs and reduces dependence on commercial frontier providers. Organizations running open models locally retain full control over their data, which has eased initial security concerns about Chinese-developed AI.

The competitive shift is stark. DeepSeek, another Chinese open-source model, now commands the largest share on OpenRouter by a substantial margin. Six months ago, U.S. providers dominated the platform.

Will open-source AI remain unrestricted?

The current appeal of open-source AI rests partly on the assumption that once a model is released, governments cannot easily withdraw it. But that assumption may not hold indefinitely.

“If Mythos-level models are considered risky, China will also not want them to be open.”

— Ethan Mollick, Professor at University of Pennsylvania

Mollick's point applies beyond the U.S. If frontier-level models become classified as national security risks, governments worldwide could move to restrict public release of their most capable open models. The window for open-weight distribution of cutting-edge AI may narrow.

What should enterprises consider now?

The restrictions have made multi-model strategies more attractive. Companies that rely exclusively on one provider face concentrated risk. Those that can swap between proprietary and open-source models, or between U.S. and non-U.S. providers, have more operational flexibility.

Cost was already driving interest in open-source AI before the restrictions. Running models on your own infrastructure avoids per-token API fees. The administration's actions added a second incentive: predictability. Open models you download and host cannot be revoked by a distant policy decision.

Frequently Asked Questions

Why did Anthropic withdraw its Mythos 5 and Fable 5 models?

The Trump administration ordered Anthropic to block non-U.S. users. Rather than build complex verification systems, the company chose to withdraw both models entirely.

What is the difference between open-source and closed AI models?

Closed models like ChatGPT are controlled by their developers through APIs and subscriptions. Open-source models release their underlying weights, allowing anyone to download, modify, and run them on their own infrastructure.

Which open-source AI models are gaining enterprise adoption?

China's DeepSeek leads the OpenRouter platform, while Zhipu AI's GLM-5.2 has drawn attention for approaching the performance of leading proprietary models on key benchmarks.

Can governments restrict open-source AI models after they're released?

Once released, open-weight models are difficult to withdraw. However, governments could restrict future releases of frontier-level open models if they classify them as national security risks.

How are enterprises responding to the AI access restrictions?

Companies are diversifying away from single-provider dependence, exploring multi-model strategies that include both proprietary and open-source options from U.S. and non-U.S. developers.

Source: PYMNTS | / PYMNTS