OpenAI limits GPT-5.6 Sol preview to U.S.-cleared customers
OpenAI says the staggered launch follows a Trump administration request tied to security concerns around advanced AI models.
By Maya Lindqvist · Senior Technology Correspondent
4 min read
OpenAI is limiting early access to GPT-5.6 Sol, its most capable model, to customers cleared through a U.S. government review, the company said Friday. The decision gives Washington a direct role in the rollout of a frontier AI system that OpenAI says has stronger cybersecurity and coding abilities than its earlier models.
The Information first reported that the Trump administration had asked OpenAI to delay broad access because of security concerns. OpenAI said the restricted preview is voluntary and that it expects to make Sol, along with two related models, more widely available in the coming weeks.
OpenAI described Sol as the top model in a new advanced tier. The company said the tier also includes Terra, a more efficient model, and Luna, a lower-cost version.
According to OpenAI, Sol can complete 50% of long-running professional tasks and outperforms all earlier OpenAI systems on coding. The company said Sol showed its biggest improvements in cybersecurity work, including vulnerability discovery and exploitation.
Government review before wider release
OpenAI said it previewed the models and their capabilities to the U.S. government before launch. In a blog post, the company said that, at the administration’s request, it was beginning with a limited preview for trusted partners whose participation had been shared with officials.
The company said the first users are customers approved by the U.S. government and that the list will grow next week. OpenAI described the process as sharing customer names with the government and receiving feedback.
OpenAI tied the temporary approach to President Donald Trump’s June 2 executive order, which directed federal agencies to create a framework for AI companies to voluntarily give the government early access to powerful models for up to 30 days before public release. OpenAI said that framework has not yet been completed.
The company said it does not want this kind of government access process to become the long-term norm. OpenAI said it is working with the administration on a repeatable process for future launches under the cyber executive order framework.
Cybersecurity safeguards
OpenAI said Sol will include two new modes, “max” and “ultra,” that allow longer reasoning and coordination among agents for specific tasks. The company previously said Sol was competitive with Anthropic’s Mythos on a key cybersecurity benchmark, while using about one-third as many tokens as Mythos.
OpenAI said Sol did not reach the “Cyber Critical” threshold in its Preparedness Framework. In tests using Firefox and Chrome, the company said the model identified the beginnings of an exploit but did not create a working version.
The company said it spent 700,000 GPU hours testing its own systems for weaknesses. OpenAI also said human testers will conduct two additional weeks of checks before launch.
OpenAI said the preview will use its most extensive safeguards to date. For higher-risk prompts, the company said a larger model may review the conversation and block a response if it determines the request violates policy.
Wider debate over AI controls
The restricted rollout follows another recent U.S. government intervention involving a frontier AI lab. In early June, the Commerce Department issued export controls on Anthropic that forced the company to cut off foreign access to two leading models, citing national security concerns, according to the department; Anthropic disputed the order.
Critics cited by Fortune have warned that case-by-case government involvement could create an informal licensing system for AI. Jonathan Iwry, a fellow at the Wharton Accountable AI Lab, previously told Fortune that the government was using existing legal powers as a “backdoor licensing regime.”
Dean Ball, a former Trump administration AI adviser, has argued that the U.S. has had an informal AI licensing system since the Mythos episode, with unclear limits on government power and public transparency. Critics say a system without published criteria or an appeals process could let officials decide which companies or customers get market access.
OpenAI listed Sol at $5 per million input tokens and $30 per million output tokens. The company priced Terra at $2.50 per million input tokens and $15 per million output tokens, while Luna costs $1 and $6, respectively.
This story draws on original reporting from Fortune.