OpenAI has unveiled a new generation of its models, GPT-5.6. It consists of three models named Sol, Terra, and Luna. Each targets a different combination of performance and price. But there is a catch. For now, only a handful of companies can try them. The US government asked OpenAI to slow down the launch, and the company complied, albeit reluctantly. Let’s first take a look at what GPT-5.6 can actually do. And then at why most people will not be able to access it anytime soon.
A new naming system
This time, OpenAI has changed the logic behind the names. The number indicates the generation (5.6), while the name indicates the performance tier. Sol sits at the top, Terra is the balanced middle ground, and Luna is fast and inexpensive. Each tier can advance at its own pace, and no one has to get as confused as before, when labels such as “Instant” did not say much.
Sol is the new flagship and, according to OpenAI, a significant leap over GPT-5.5. The company describes it as its most capable model, especially when it comes to cybersecurity. It leads in all the benchmark tests OpenAI presented for the new release.
Terra is intended to be the workhorse for routine everyday tasks. Its performance remains on par with the previous flagship, GPT-5.5, but at roughly half the price. We may see this with future releases as well: the quality of the previous top model at a mid-tier price.
Luna is the fastest and cheapest. It targets high-volume workloads, tasks sensitive to response speed, and users keeping a close eye on their budgets. But bear in mind that “cheapest” does not automatically mean “weakest” for every task.
Sol and its capabilities
Sol is the only model to unlock two new settings. The first is called max and gives the model more time to think through complex prompts. The second is ultra; instead of having one model solve the task on its own, it divides the work among so-called sub-agents that operate in parallel. This is useful for long, multi-step tasks that a single agent cannot handle within one context.
And how does Sol perform in tests? It achieved a new best result on Terminal-Bench 2.1, which measures command-line work requiring planning and tool coordination. According to the available data, Sol Ultra scored 91.9 percent. It also improved on GeneBench v1, which focuses on genomics and quantitative biology, while using fewer tokens than GPT-5.5. And on ExploitBench, which tests long-horizon vulnerability discovery, it outperformed older models.
OpenAI spoke extensively about safety. Sol is reportedly trained to refuse prohibited assistance involving cyberattacks, even when users try to disguise their intent or bypass the model’s safeguards. The company also emphasizes that Sol is better at helping to find and fix vulnerabilities than at reliably carrying out attacks. According to OpenAI’s own preparedness framework, Sol reportedly does not cross the critical threshold in the cyber domain.
Model pricing
Prices are per million tokens. Sol costs $5 for input and $30 for output, which is nearly half the price of Anthropic’s competing Claude Fable 5 model ($10 and $50). Terra costs $2.50 for input and $15 for output. Luna is the cheapest, at $1 for input and $6 for output.
We will have to wait for access. The US government has intervened again
OpenAI unveiled a top-tier trio of models and immediately announced that almost no one could use them, much like what happened with the competing Mythos model after its announcement. Access is available through the API and Codex, but only to a small group of vetted partners for now. According to available information, this amounts to roughly twenty organizations. There are currently no public waitlists or application forms.
The reason is a request from the US government. National cybersecurity authorities asked OpenAI to limit the launch until the administration completes a system for testing the safety of new models. OpenAI CEO Sam Altman told employees that during the preview, the government would approve access one customer at a time. Even after OpenAI discussed its plans with senior officials, Commerce Secretary Howard Lutnick called Altman and warned the company not to proceed without approval from other agencies.
The background is an executive order signed by President Trump in early June. It asks leading labs to voluntarily submit their most advanced models to the government for review up to 30 days before release. But OpenAI’s case shows how voluntary this really is. Dean Ball, a former White House adviser on artificial intelligence and future OpenAI employee, says the order has effectively created an involuntary licensing regime for leading models.
OpenAI complied, but made it clear that it was unhappy about doing so. “We do not believe this approach to government involvement should become the long-term norm,” the company wrote in a Friday post. “It keeps the best tools away from users, developers, businesses, cyber defenders, and partners around the world who need them.” OpenAI described the limited preview as a short-term step toward broader availability, which is expected within a few weeks. The company is also working with the administration on a new cybersecurity process and a repeatable procedure for future releases.
This is not the first such intervention. Just a few weeks earlier, competing company Anthropic had to disable access to its Mythos 5 and Fable 5 models because of a government export-control order. By comparison, OpenAI received more lenient terms. On Friday, the government also relaxed one of Anthropic’s restrictions, allowing Mythos 5 to return to a small group of cyber defenders.
OpenAI confirmed to Euractiv that all companies participating in the preview are based in the United States. No European company or authority therefore has access yet.
Sources: thehackernews.com and techcrunch.com
