OpenAI is delaying the full public launch of its next flagship system, GPT‑5.6. This comes after the US government requested that initial access be limited to a small, vetted group of partners. This marks a significant increase in Washington's involvement in releasing cutting-edge artificial intelligence models.

The company introduced the GPT‑5.6 family on Friday, which includes three models: Sol, its most powerful model to date; Terra, a mid-tier model comparable to GPT‑5.5; and Luna, a faster, more affordable option designed for high-volume use. Instead of making the models widely available right away, OpenAI plans to start with a limited preview for trusted partners. A broader release is expected "in the coming weeks."

In a blog post about this decision, OpenAI stated that it had informed the government of its plans and the capabilities of the new models before the launch. The restricted rollout happened "at their request”. The company emphasised that this arrangement is temporary. They believe that such government access should not become the standard long-term practice, as it prevents users, developers, businesses, and global partners from accessing valuable tools.

This situation follows an executive order signed by President Donald Trump earlier this month. It established a voluntary framework for AI developers to offer certain frontier models to the federal government for review for up to 30 days before releasing them to trusted partners. OpenAI did not reveal which companies received early access.

Sol launches with what OpenAI describes as its most robust safety stack to date, after the company said it had spent multiple weeks finding weaknesses and pressure-testing the system against real-world attacks. Central to the new release is a fresh reasoning capability called max reasoning effort, which gives Sol additional time to work through complex problems, alongside a new "ultra mode".

On coding, OpenAI said Sol sets a new state of the art on Terminal-Bench 2.1, a benchmark that tests command-line workflows requiring planning, iteration and coordination across tools. In biology, the model is said to show broad gains on GeneBench v1, a benchmark for long-horizon genomics and quantitative-biology analysis, outperforming GPT-5.5 while using fewer tokens to do so.

The most striking claims, however, concern cybersecurity. OpenAI described Sol as its most capable model yet in this area, saying it shifts the balance between performance and efficiency on long-horizon security tasks such as vulnerability research and exploitation. On a benchmark called ExploitBench, the company said Sol performs competitively with Anthropic's Mythos Preview model while using roughly a third of the output tokens. On ExploitGym, a benchmark developed by researchers at UC Berkeley in collaboration with OpenAI and other frontier labs, all three GPT-5.6 models were said to show strong gains in cyber capability as reasoning effort increases.

Sol, it said, is better suited to helping people find and fix vulnerabilities than to reliably carrying out a complete attack from start to finish. The company said the model does not cross the "Cyber Critical" threshold under its internal preparedness framework, and that in evaluations involving the Chromium and Firefox browsers, it was able to identify bugs and exploitation primitives, the basic building blocks of an exploit, without autonomously assembling a fully functional exploit chain under the conditions tested. OpenAI acknowledged, though, that benchmark results cannot account for every way a model might be used or combined with other tools, which it said was part of the reasoning behind pairing Sol's expanded capabilities with stronger safeguards and a phased rollout.

OpenAI said it had dedicated more than 700,000 A100-equivalent GPU hours to automated red-teaming focused on finding universal jailbreaks, attacks designed to work across many prompts and contexts rather than a single narrow case. This was supplemented by extensive human expert red-teaming conducted with third-party testers, a process the company said would continue throughout the preview.

On pricing, GPT-5.6 is billed per million tokens: Sol costs $5 for input and $30 for output, Terra costs $2.50 and $15, and Luna costs $1 and $6 respectively. The release also introduces more predictable prompt caching, with explicit cache breakpoints and a 30-minute minimum cache life. OpenAI said it would launch Sol on Cerebras hardware in July at speeds of up to 750 tokens per second, initially limited to select customers as capacity expands.