Anthropic and OpenAI are now building industry-specific AI applications and features, a strategic shift that is putting many partner companies in a tough spot — some are suddenly competing head-to-head with their primary AI model supplier. Design platform Canva is the latest firm to encounter this challenge.

Despite these developments, most industry observers believe Anthropic and OpenAI will continue to offer models through their application programming interfaces (APIs), which generate billions of dollars in annual revenue for both firms. As long as that arrangement persists, companies still have a fighting chance against fresh competition from their own vendors. But how sustainable is this setup?

Over the past few weeks, I have spoken with researchers at both frontier AI labs and founders who build applications on top of the latest models, and I have started to question whether this equilibrium can hold. There are growing signs that, whether driven by internal corporate strategy or external government pressure, both Anthropic and OpenAI could weaken the capabilities of their most powerful models available via API.

Why the API access to top-tier models may be restricted

The Trump administration, concerned about potential misuse by cybercriminals and other malicious actors, has pressured both companies to adopt more cautious release processes for certain new models. During this period, clients must go through case-by-case reviews to obtain model access. (The administration recently issued guidance outlining a special testing framework that large AI firms can follow before releasing models.)

Even after models become publicly available through the API, Anthropic and OpenAI may still throttle output performance on certain tasks for the same security reasons. Both companies have already done this — for example, Anthropic's Fable model has imposed limits on cybersecurity-related tasks. Naturally, when it comes to their own in-house vertical applications and features, the two firms may not apply such performance downgrades to their models.

Both companies have warned that future models could reach dangerous levels of capability in areas like hacking and bioweapon development. At some point, AI providers may conclude that preventing malicious users from exploiting models for such purposes has become nearly impossible; and if bad actors break through defenses and cause real harm, the companies would face an overwhelming wave of public backlash. It may sound extreme, but they could decide that completely cutting off external access to advanced models is actually the safer route.

Beyond security: the distillation dilemma

Aside from safety concerns, there are other motivations pushing both firms to tighten API access to advanced models. Both want to prevent knowledge distillation — a technique where competitors, particularly Chinese AI companies, use outputs from commercial advanced models to train their own new models. A former Anthropic researcher told me that completely eliminating distillation is practically impossible.

If U.S. AI firms genuinely worry about malicious actors using AI models to cause global disruption, they would certainly not want to indirectly help Chinese companies develop fully open-weight models that anyone can use for free. Anthropic wrote in its recent report on potentially catastrophic AI risks: "If a distilled model inherits the high-risk capabilities of the original model but lacks the comprehensive safety measures that accompany it, then even if the original model is well-protected, the distilled version could still create secondary risks for the world."

Open-source software advocates, however, argue that open models — including those distilled from Anthropic models — can actually help developers defend against cyberattacks. Hugging Face has experienced intrusions by OpenAI agents, and open-source models can serve a protective role in such scenarios.

Commercial incentives and the path forward

There is also a more pragmatic, commercially driven consideration: Anthropic and OpenAI could cut off API access for business purposes. As previously reported, some investors have already warned developers that Anthropic might withhold its best technology and instead build applications that compete directly with its clients. It is worth noting that if Anthropic, as the absolute leader in the AI API market, were to take such a step, it would almost certainly trigger a massive antitrust investigation.

Of course, OpenAI and Anthropic are unlikely to willingly abandon the API business, which remains a crucial revenue stream. However, the potential returns from proprietary AI applications could even surpass those from the API business. Anthropic has already entered fields like AI-driven drug discovery.

For now, the uncertainty surrounding future API access explains why AI application developers like Harvey and Cursor have already begun training their own internal models instead of relying entirely on OpenAI and Anthropic. More companies are sure to follow their lead.