OpenAI pledges early access to next model for U.S. safety testing

Sam Altman is trying to put out some fires in Washington, and his newest fire extinguisher is an early look at OpenAI’s next frontier model.

Altman announced late Wednesday on X that OpenAI is partnering with the U.S. AI Safety Institute to provide early access to its next major generative AI model for safety evaluations. The pledge mirrors a similar arrangement OpenAI made with the UK’s safety body back in June.

Rebuilding trust after a safety exodus

While details on the actual testing protocol remain thin, the timing is hard to ignore.

Back in May, OpenAI effectively dissolved its superalignment team, which was built to steer superintelligent systems. That move sparked severe backlash after internal reports suggested safety research was being sidelined for faster product rollouts. The team’s co-leads, Ilya Sutskever and Jan Leike, both walked out. Leike joined rival Anthropic, while Sutskever launched his own startup, Safe Superintelligence.

OpenAI attempted damage control by voiding restrictive non-disparagement agreements that gagged departing staff, creating a new safety committee, and recommitting 20% of its compute power to safety work. Yet skeptics stayed unconvinced, largely because the safety committee was composed entirely of company insiders, including Altman himself.

Lawmakers have taken notice. A group of five senators, led by Hawaii Democrat Brian Schatz, recently fired off a letter demanding answers on OpenAI’s internal safety policies. In response, chief strategy officer Jason Kwon insisted the company adheres to rigorous protocols at every stage.

Big lobbying bets and regulatory capture

Teaming up with the government also coincides with OpenAI’s broader legislative push in Washington.

Earlier this week, the company endorsed the Future of Innovation Act, a Senate bill that would officially codify the U.S. AI Safety Institute—housed under NIST within the Department of Commerce—as a formal rulemaking and standards body. Critics see this dynamic as a textbook bid for regulatory capture, cementing OpenAI’s influence over federal AI governance.

OpenAI’s spending backs up that strategy. The company dropped $800,000 on federal lobbying in just the first six months of 2024, compared to $260,000 across all of 2023. Altman also sits on the Department of Homeland Security’s AI Safety and Security Board.

The U.S. AI Safety Institute is currently working alongside an industry consortium that includes OpenAI, Anthropic, Google, Microsoft, Meta, Apple, Amazon, and Nvidia to carry out President Biden’s AI executive order guidelines on red-teaming, watermarking, and risk management.

Why it matters

For developers and enterprise users anticipating OpenAI's next frontier model, this agreement likely adds a formal pre-release buffer. While this might stretch launch timelines, it also provides an external safety stamp that enterprise legal teams increasingly crave before deploying cutting-edge AI.

OpenAI is playing smart defense here. By volunteering its next model to federal inspectors, the company deflects mounting congressional heat and counters Anthropic’s hard-won reputation as the industry’s safety-first darling.

Yet the structural concerns remain. OpenAI’s internal safety committee is still completely insular, and Altman’s reaffirmed 20% compute commitment will be judged against past promises that went unfulfilled. Even so, early government evaluation looks like it is rapidly becoming the mandatory cost of doing business at the frontier.