OpenAI Restricts Access to New GPT-5.6 Sol Model at US Government's Request

OpenAI has begun a limited rollout of its new GPT-5.6 model family, Sol, Terra and Luna. Access to the most capable of the three, Sol, has been granted to a narrow circle of about 20 vetted organizations, after the company consulted its deployment plans with the US government in advance. The caution stems from the model's results on offensive hacking capability tests.

Three model tiers

GPT-5.6 is a new family of three models built for different purposes. Sol is the flagship, described by OpenAI as the strongest in coding, biological and cybersecurity capabilities. Terra is a mid-tier variant, cheaper and faster but still capable of complex tasks. Luna is the fastest and cheapest, aimed at everyday use.

The company says Sol performs especially well on long, multi-step tasks involving vulnerability research in code. On the ExploitBench test, the model matched the results of competing systems while using roughly a third of the output tokens its rivals required.

Critical threshold and government oversight

All three models crossed the 'high' threshold on internal cyber-offensive capability tests, but none crossed the 'Cyber Critical' threshold defined under OpenAI's Preparedness Framework. In tests against the Chromium and Firefox browsers, Sol was able to identify bugs and exploit building blocks, but did not independently produce a fully functioning attack chain under test conditions.

The model does not cross our internal Cyber Critical threshold - OpenAI, GPT-5.6 Sol system card

Even so, OpenAI opted for an unusually cautious launch. The company briefed the US government on the model's capabilities ahead of release and, at the government's request, began distribution with a narrow group of trusted partners whose list had previously been shared with government officials. Broader availability is expected within the coming weeks.

Safeguards and criticism

OpenAI describes GPT-5.6 Sol as the model with its most robust safety stack to date, featuring trained refusals for risky cybersecurity requests, real-time classifiers that pause and verify suspicious responses, and protections against repeated abuse attempts. External experts also worked on the model's safety testing, and the automated red-teaming tests alone consumed more than 700,000 GPU compute hours.

Restricting access to the model at the request of the government administration has also drawn criticism. Some observers note that such a model reaches select entities approved by the government first, rather than the open market, potentially giving privileged partners a head start in accessing the most powerful AI-based cybersecurity tools.

Implications for businesses and developers

For IT security and penetration testing teams, the announcement means that the most powerful AI vulnerability-detection capabilities will for now be available only to select institutions, mainly in the United States. Polish companies and security teams will have to wait for broader access to Sol or rely on the cheaper Terra and Luna variants, which will reach customers before the flagship model.

Sol's pricing, $5 per million input tokens and $30 per million output tokens, places it in the premium model segment, comparable to the most expensive offerings from competitors. OpenAI also announced that Sol will run on Cerebras infrastructure at speeds of up to 750 tokens per second, aiming to give customers near real-time responses on complex analytical tasks.

The coming weeks will show how quickly OpenAI expands access beyond its narrow circle of partners, and whether other companies developing models with high cyber-offensive capabilities adopt a similar controlled-release approach in cooperation with government authorities.

Sources: Previewing GPT-5.6 Sol: a next-generation model (openai.com), OpenAI unveils GPT-5.6 Sol, Terra and Luna models (venturebeat.com), OpenAI limits GPT-5.6 Sol access to Trump administration partners (foxnews.com), OpenAI Reveals GPT-5.6 Sol Cybersecurity Model, Restricts Early Access (infosecurity-magazine.com)