OpenAI introduced new GPT-5.6 models. But almost no one will get access to the most powerful ones

The race for the most powerful artificial intelligence continues to accelerate, and OpenAI has just sent a new generation of its GPT-5.6 language models into it. However, you won’t get access to the most capable ones anytime soon – OpenAI has made them available only to a narrow group of vetted companies due to a request from the US government. A similar situation occurred with Anthropic, whose Mythos 5 was also not publicly available. Even the released Claude Fable 5 model had to be shut down by order of the US government shortly after its introduction.

What can the new GPT-5.6 models do?

OpenAI introduced three models at once. The most powerful is Sol, which the company targets for demanding tasks in programming, computer work, scientific research, and cybersecurity. Alongside it is the more balanced Terra for more common deployments, and the fastest and cheapest Luna.

Sol also received two new reasoning modes. In “max” mode, the model takes more time to answer and thinks it through more thoroughly, which is useful for more complex tasks. The “ultra” mode goes even further and involves so-called subagents – the model divides the task among several of its instances, which work in parallel.

Prices vary by model and are calculated per million tokens (simply put, for the volume of processed text). Sol costs 5 dollars per million input tokens and 30 dollars per million output tokens, Terra costs 2.50 and 15 dollars, and the cheapest Luna costs 1 and 6 dollars. According to OpenAI, Terra will offer performance comparable to the earlier GPT-5.5 at roughly half the price.

How does GPT-5.6 compare to Claude?

A direct comparison is difficult for now. OpenAI has not published most of the official results for Sol – for example, scores in demanding tests like SWE-bench Pro or LiveCodeBench are missing. The company only states that Sol set a new record in the Terminal-Bench 2.1 test, which evaluates command-line work.

OpenAI itself published comparative figures for this test: Sol, with 88.8%, leaves the publicly available Claude Opus 4.8 (78.9%) behind and is roughly on par with the most powerful Mythos 5.

OpenAI introduced new GPT-5.6 models. But almost no one will get access to the most powerful ones Adam Kurfürst News Adam Kurfürst News

Furthermore, the comparison is complicated by the fact that both Claude Fable 5 and Mythos 5 remain unavailable to regular users – the US government ordered Anthropic to suspend access shortly after their introduction, and it has not yet been broadly restored.

Why is GPT-5.6 almost unavailable for now?

The GPT-5.6 models are currently in a so-called limited preview. Only about twenty companies approved by the US government have gained access to them via the API interface and the Codex tool, and they are not in ChatGPT at all during the preview. OpenAI plans broader availability in ChatGPT, Codex, and API within weeks, according to its statement.

The restriction came at the request of the Trump administration. As TechCrunch reports, authorities now require companies to submit advanced models for review up to 30 days before release. OpenAI itself adds that, in its opinion, a similar government approval process should not become a long-term standard.

Should the government decide who gets access to the most powerful artificial intelligence?

Sources: OpenAI, TechCrunch, 9to5Mac