OpenAI's IPO Is Gated on Model Safety Promises, Says Altman

What Altman actually said

For months, people have asked when OpenAI will go public. According to The Verge's report, CEO Sam Altman now says it won't happen until the company can make better promises about model safety, and there is no firm timeline in sight.

Altman also said OpenAI intends to continue with AI progress. He tied the safety question to the recent surge in model capability. That framing matters: the safety bar is being described as something that has to catch up with the models, not as a reason to slow them down.

Why an IPO makes safety promises harder

This is our reading, not something Altman spelled out in the coverage we could confirm. A public listing turns statements about a product into statements to investors. Claims about how a model behaves, and how reliably, become things a company has to stand behind in regulatory filings. If the underlying models are still changing quickly, those claims are hard to make with confidence. Waiting until the promises are stronger is a coherent position, whatever else motivates it.

What it means if you build on frontier models

If you run LLMs inside trading systems, fintech workflows or developer tooling, the takeaway is practical. A vendor's safety posture is a dependency that can change under you, and it is not a guarantee you can build a control plan around.

• Own your guardrails. Put deterministic limits (position caps, allow-lists, human approval on irreversible actions) outside the model, so a behaviour change upstream cannot widen what the system is allowed to do.
• Pin and test model versions. Re-run your own evaluation set whenever a provider ships a new model, instead of assuming safety properties carry over.
• Keep a second option. Abstract the provider behind an interface so you can swap or fail over if terms, behaviour or availability shift.
• Map to a risk framework. The NIST AI Risk Management Framework gives you a vendor-neutral vocabulary for documenting the risks you accept.

What we don't know yet

The report gives no timeline and, from the excerpt we could verify, no concrete criteria for what counts as a good enough safety promise. OpenAI publishes its own Preparedness Framework describing how it tracks severe risks, but we would not treat any single document as the finish line Altman has in mind. Watch for specifics: named thresholds, external audits or evaluation results the company is willing to commit to publicly.