What is GPT-5.6, and why is it only in a limited preview?

Yesterday, Friday, June 26, 2026, OpenAI introduced GPT-5.6, a new generation of its models, but almost nobody can use it yet. Instead of the usual launch to ChatGPT subscribers and developers, OpenAI is opening it first to a handful of partners it trusts. It says it did so at the request of the US government, after showing officials its plans and the models' abilities ahead of time.

That makes this launch unusual in two ways. It brings a new naming system and a new top model. And it is the clearest sign yet that the US government now has a hand in when the most powerful AI models reach the public.

Three tiers with names

GPT-5.6 comes in three versions. In OpenAI's new naming scheme, the number marks the generation, and the names Sol, Terra and Luna mark lasting tiers of capability that can be updated separately.

A token is a small chunk of text, often part of a word; developers pay separately for what they send in and what the model writes back. Sol's price matches GPT-5.5's.

During the preview, partners reach the models through the API and Codex, OpenAI's coding agent. OpenAI plans to bring all three to ChatGPT, Codex and the API more broadly soon, and to offer a very fast version of Sol, up to 750 tokens per second, on Cerebras hardware for selected customers next month.

What Sol can do

OpenAI calls Sol its strongest model yet, with gains in coding, biology and cybersecurity. Two new settings stand out:

• A max reasoning effort, which gives Sol the most time to think before answering.
• An ultra mode, which goes beyond a single agent by splitting complex work among subagents, separate copies of the model that each handle a part of the job at the same time.

On results, OpenAI says Sol sets a new record on Terminal-Bench 2.1, a test of multi-step command-line work, and beats GPT-5.5 on GeneBench, a set of long genetics and biology analyses, while using fewer tokens. TechCrunch reports OpenAI's claim that Sol is slightly better than Anthropic's Claude Mythos 5 at coding workflows. OpenAI will publish fuller test results at the broad release.

Developers also get more predictable prompt caching, which stores the unchanging start of a long prompt so repeated requests cost less. It now supports explicit cache breakpoints and keeps the cache for at least 30 minutes.

Why access is limited

Sol is OpenAI's most capable model for cybersecurity, and that is the heart of the caution. OpenAI says Sol does not reach the "Critical" cyber level of its Preparedness Framework, the internal scale it uses for serious risks. In tests on the Chromium and Firefox browsers, it found bugs but did not produce a complete working attack on its own. Still, OpenAI admits that tests cannot capture every possible use.

Its safety approach has several layers:

1. The model is trained to refuse prohibited cyber help, even when the request is disguised.
2. Real-time classifiers, automatic checkers for cyber and biology misuse, read the answer as it is written; for higher-risk cases they can hold the answer while a bigger reasoning model looks over the whole exchange.
3. Flagged activity can trigger an account-level review across conversations.
4. Different users get different levels of access, and use is monitored.

OpenAI says it spent over 700,000 GPU hours, measured in A100 equivalents (a standard data-center graphics chip), on automated attempts to find universal jailbreaks, tricks that would switch off safeguards across many requests. It warns that during the preview some legitimate requests may be blocked or slowed.

The government request comes against a tense background. TechCrunch reports that after Anthropic released its Fable 5 model this month, the administration ordered it to cut off access for any foreign national, and Anthropic took the model offline entirely. A recent executive order invites leading labs to hand their strongest models to the government for review, on a voluntary basis, as much as 30 days ahead of launch. Dean Ball, who previously advised the White House on AI, told TechCrunch he believes this has become a licensing system in practice.

Why it matters

OpenAI complied, but said clearly that it disagrees with the arrangement as a lasting rule. "We don't believe this kind of government access process should become the long-term default," its announcement says, adding that such delays keep useful tools from developers, businesses and cyber defenders. It calls the preview a short-term step while it works with the administration on a cybersecurity framework under an executive order and on a standard procedure it can reuse for later models.

For the public, the practical effect is a delay. For the industry, the question is now who decides when a frontier model is ready, and by what standard. Ball's worry is that without clearly defined safety standards, reviews could lead to open-ended delays. The next few weeks will show whether OpenAI's promised wider release arrives on schedule.

When will most people be able to use GPT-5.6, and how should they get ready?

OpenAI says it expects broad availability in ChatGPT, Codex and the API "in the coming weeks," but gives no date, and the timing may depend on its talks with the government. Until then, GPT-5.5 remains OpenAI's top model for subscribers and developers.

The best preparation is to know which tier fits which job. Luna suits quick, high-volume tasks such as sorting messages or short summaries. Terra suits everyday writing, analysis and coding. Sol, with its max effort and ultra modes, is for long, difficult projects where quality matters more than cost.

Developers can prepare prompts so they are easy to cache: put fixed instructions first and the changing details last.

[Fixed instructions, the same every time]
You are a support assistant for our bike shop. Answer in two short
paragraphs, and say when you are not sure.

[Changing part, last]
Customer question: Can I bring my e-bike in for a tune-up on Saturday?