OpenAI unveils GPT-5.6, but you can't use it yet

On July 9, 2026, OpenAI unveiled GPT-5.6 and calls it the company’s most capable system to date. The naming convention is also new. The number represents the generation, while the suffixes Sol, Terra, and Luna denote three tiers of sustained performance. Sol is the top-of-the-line model, Terra is the balanced option for everyday use, and Luna is the fast and affordable tier. However, anyone who wants to try out GPT-5.6 right away will be disappointed.

For now, it’s only available to a few partners, not in ChatGPT

During the preview, GPT-5.6 runs exclusively via the application programming interface (API) and the Codex developer environment, and only for a small group of selected partners and organizations. The model is not available in ChatGPT at this stage. There is no public waitlist or sign-up option. Regarding widespread availability in ChatGPT, Codex, and the API, OpenAI has only provided a vague timeframe, namely the coming weeks. No specific date has been given. Until then, GPT-5.5 remains the go-to option for most users, and anyone who wants to know how ChatGPT, Gemini, and Claude perform in everyday use can find a direct comparison there.

The U.S. government is involved

The reason for the delayed launch is unusual. OpenAI demonstrated the model’s capabilities to the U.S. government in advance and, at the government’s request, is initially releasing it only to a small group of partners designated by the government. This is due to the model’s significantly enhanced capabilities in the area of cybersecurity. According to OpenAI, GPT-5.6 is particularly good at identifying vulnerabilities and fixing bugs, which helps defenders but poses risks if it falls into the wrong hands. OpenAI itself emphasizes that such a government-led access process is not intended to be a permanent arrangement, as it keeps good tools out of the hands of users and developers.

What Sol, Terra, and Luna can do

For developers, GPT-5.6 brings noticeable improvements in programming, biological analysis, and security tasks. Terra is said to match the performance of its predecessor, GPT-5.5, while costing only half as much. New features include a “max” level for particularly lengthy reasoning and an “ultra” mode that distributes subtasks among multiple sub-agents. Prices per million tokens are $5 for input and $30 for output on Sol, $2.50 and $15 on Terra, and $1 and $6 on Luna. For regular users, these figures are of secondary importance, but they do show just how clearly OpenAI differentiates between the pricing tiers.

Speed boost through specialized hardware

One detail is likely to interest tech enthusiasts. In July, OpenAI will also launch GPT-5.6 Sol on hardware from chip specialist Cerebras, with speeds of up to 750 tokens per second. That’s many times faster than the typical speed on traditional graphics cards and shows that the industry is increasingly looking for alternatives beyond the well-known GPU market leaders. This option will also initially be available only to select customers while Cerebras expands its capacity.

The bottom line is that GPT-5.6 represents a major leap forward, though most people won’t experience it for themselves for a few weeks. Until the model is integrated into ChatGPT, little will change in everyday use.