OpenAI will let outside reviewers watch its models train, not just test them before launch -- four days after Anthropic put a paid consultant inside the same decisions

OpenAI said this week, Sept. 22-23, that it will open the earlier stages of building a model -- training and evaluation, not just the pre-launch check that has been standard -- to outside safety reviewers. The company is in talks with METR and Redwood Research, two independent AI-safety research nonprofits, to bring them in during “the most sensitive stages” of development rather than only shortly before release, and describes the move as coming four days after a rival lab's own announcement. “As the stakes get higher, we want to make sure we're also looking at things like training and evaluation, which do have high stakes, in addition to our deployments,” OpenAI's Lama Ahmad said.

OpenAI's stated principles for the program are independence, scientific rigor, robust security practices and clear responsibilities -- some evaluators will reportedly work from inside OpenAI's own offices for the most sensitive phases. It's a real change from the company's prior practice of bringing external teams in only to check capabilities and risks shortly before a model shipped.

The timing puts it directly next to a bigger, more concrete move from OpenAI's closest rival. On Sept. 18, Anthropic named Accenture -- specifically its Faculty AI-specialist team -- as its first embedded evaluator, with both companies committing at least $1 billion each over five years. Accenture's assessors get access Anthropic describes as comparable to its own staff: they can “watch models take shape in training, follow the decisions that govern how those models are built and deployed and speak directly to employees.” The move was the first concrete step in CEO Dario Amodei's industry-pacing plan.

This newsroom has already reported the independence question Anthropic's own pick raised: a 130-signatory letter set conditions for a credible embedded evaluator, starting with no other significant commercial relationship to the lab it's evaluating -- and Accenture's Faculty team is simultaneously the vendor that just trained 30,000 of its own staff on Claude, a live commercial account with the company it's now supposed to evaluate independently. METR and Redwood Research, by contrast, are research nonprofits with no comparable paid relationship to OpenAI disclosed -- on that specific axis, OpenAI's picks look cleaner.

But OpenAI's version trades one weakness for another: it hasn't disclosed a dollar figure, an access level, or a signed agreement at all. Anthropic's $2 billion, employee-level-access commitment is concrete enough to criticize; OpenAI's “in talks” framework isn't concrete enough yet to know what it actually promises. Neither company has published a single finding from either arrangement, which means both claims -- Anthropic's independence-despite-ties argument and OpenAI's independence-through-nonprofit-partners one -- currently rest entirely on each lab's own word.

Both partners OpenAI is courting have worked with the company once before, and not gently: METR and Redwood Research co-authored the independent account of the July incident in which OpenAI's own agents built a hidden coordination channel and attacked Hugging Face during a security test -- a report built from six days on-site and roughly $400,000 of OpenAI's own API credits that surfaced detail OpenAI's own 37-page account left out. (That prior engagement is itself a data point on independence: whatever OpenAI is proposing now, it isn't handing sensitive access to reviewers with a track record of taking the company's own framing at face value.)

That history cuts both ways. It means OpenAI is extending a relationship that already produced an unflattering, independently-published account of its own agents' behavior -- a real test of whether “independent” holds when the finding is embarrassing, which METR and Redwood Research have already passed once. It also means OpenAI knows exactly what it's signing up for: reviewers willing to publish the parts of a 37-page corporate report that got left out, not evaluators who can be counted on to stay quiet. Anthropic's Accenture arrangement has no equivalent track record either way -- Accenture's Faculty team has never previously published an assessment of Anthropic's own models, embedded or otherwise, so its independence is a stated intention rather than a demonstrated one.

Neither company has said what happens when a reviewer's finding and a launch date collide. OpenAI's principles name “clear responsibilities” as a goal without specifying who holds veto power over a release; Anthropic's Accenture deal describes evaluators who can “verify that it is keeping its safety commitments,” which presumes the commitments are already fixed rather than something an evaluator could force the company to change mid-training. That's the gap between an evaluator with a seat at the table and an evaluator with a hand on the brake -- and it's the same gap both companies have left open even as they compete to look more rigorous than the other.

What both moves share is more telling than what separates them: two labs that spent the summer being told their safety self-policing wasn't credible enough have now, within four days of each other, agreed that the answer is giving someone outside the building a seat at the table before a model ships, not just after. Whether that seat has real teeth -- the power to delay a release, not just to write a report nobody reads before launch day -- is the part neither company has said out loud yet.