OpenAI Delays GPT-5.6 Release After Trump Admin Review Request

The headline writes itself as a product launch, but the more interesting detail is buried near the middle: the Trump administration asked OpenAI to hold back a major model release while the government builds a framework to review advanced AI before it reaches the public.

OpenAI agreed to the limited rollout, at least for now. The company also made clear it is not treating government review as a permanent or default condition for future launches, which suggests the arrangement is more of a one-time accommodation than a new operating norm. That tension between cooperation and resistance is doing a lot of quiet work in a press release that mostly wants to talk about benchmarks.

The model itself is called GPT-5.6, and it ships in three tiers:

• Sol: the highest-performance option, with a new “ultra” mode that uses sub-agents for complex tasks, priced at $5 input and $30 output per million tokens
• Terra: mid-tier for everyday workloads, at $2.50 input and $15 output
• Luna: the fastest and cheapest option, at $1 input and $6 output

OpenAI is pointing to coding, genomics analysis, and long-horizon security tasks as Sol’s strongest areas. The cybersecurity framing comes with a specific qualifier: the company says the model is better at finding and fixing vulnerabilities than at executing full attacks, which is the kind of distinction that sounds reassuring but is also exactly what you would say regardless.

Safety and Access

The safety system in GPT-5.6 is described as OpenAI’s strongest yet, with specific attention to biology risks, jailbreak attempts, and repeated misuse patterns. The model is currently available through the API and Codex for select partners, with broader access to ChatGPT and the wider API coming later.

The pricing structure makes Luna genuinely cheap for high-volume applications. Whether that matters more than the government review question depends on who is asking.