In the early hours of June 26, 2026, OpenAI opened a preview of GPT-5.6 Sol along with its two siblings, Terra and Luna. This time the company did not do a broad launch. It ran a limited test through the API and Codex for a group of trusted partners, after first briefing the US government on the model's capabilities. A preview handled this cautiously is probably worth paying attention to. For programmers who use AI to write code, this is news worth following. Hung Phat breaks down what you need to know.
The GPT-5.6 trio of Sol, Terra and Luna: three capability tiers and a new naming approach
OpenAI released three models in the GPT-5.6 family at once. Sol is the top-tier version, built for the hardest tasks. Terra is the balanced version for everyday work, and the company says its performance is competitive with GPT-5.5 at half the price. Luna is the fastest and cheapest, aimed at anyone who needs to process large volumes at low cost.
The three-way split maps closely to how people actually use these models. If you need an assistant for long-horizon planning, tangled debugging or heavy problems, pick Sol. For everyday jobs like writing documentation, summarizing or fixing short snippets of code, Terra is enough and saves money. When you're running thousands of repeated requests, Luna keeps the bill lowest. All three belong to the same generation, so the underlying quality is similar and they differ only in capability level and price.
GPT-5.6 also changes how models are named. The number marks the generation, while Sol, Terra and Luna are three lasting capability tiers that can evolve on their own schedules. In other words, there could later be a new, stronger Sol that keeps the name Sol, instead of spawning a string of hard-to-remember version numbers. For users, you weigh three axes, intelligence, speed and cost, and attach them to one name, so there's less need to look up a spec sheet.
Two new modes come with it. The max reasoning level gives Sol more time to think deeply before answering, which suits multi-step problems. The ultra mode goes beyond the limits of a single agent, drawing on subagents running in parallel to speed up complex work. It's a sign that OpenAI wants to turn the model into a system that can divide up work on its own, not just answer faster.
You can read the original announcement in OpenAI's GPT-5.6 preview post. For long workflows such as building a software project from scratch, running multiple agents usually cuts the time considerably compared with working sequentially.
How strong GPT-5.6 Sol is: coding, biology and cybersecurity
On programming, OpenAI says Sol set a new record on Terminal-Bench 2.1, a test of command-line operations that demand planning, iteration and tool coordination. This is exactly the kind of work a real coding assistant has to do well, rather than just generating a few disconnected snippets of code.
Looking at the scores, the gap is fairly clear. GPT-5.6 Sol Ultra scored 91.9%, and GPT-5.6 Sol scored 88.8%, edging out Claude Mythos 5 at 88.0%. Meanwhile, the previous-generation GPT-5.5 scored 83.4%, Claude Opus 4.8 scored 78.9% and Gemini 3.1 Pro Preview scored 70.7%. The chart below lays out all nine names to help you picture it.
In biology, Sol outperformed GPT-5.5 on GeneBench v1, an evaluation of long-horizon genetic analysis and quantitative biology, while using fewer tokens. Using fewer tokens means it's both faster and cheaper for the same problem, a practical plus for research teams.
On cybersecurity, the company says this is its strongest model for finding and patching vulnerabilities. On ExploitBench, GPT-5.6 Sol can compete with the Mythos Preview while using only about a third of the output tokens. On ExploitGym, a benchmark built by a UC Berkeley team together with OpenAI and a few other frontier labs, Sol, Terra and Luna all improved noticeably when their reasoning level was raised.
Taken together, the picture is a real leap in capability, especially in coding and security. But that same leap is the reason OpenAI isn't rushing to open it up to everyone, as the next section explains.
Why OpenAI is opening access gradually: the Cyber Critical threshold, layered safeguards and red-teaming
What's unusual is that OpenAI has not opened it to everyone. It is letting only a group of trusted partners use it first, and that list has been shared with the US government. The company says this is a short-term step while it works with the administration on a framework of executive orders for cybersecurity and a repeatable process for future releases.
The reason is a sharp jump in cybersecurity capability. In tests on Chromium and Firefox, Sol found bugs and the building blocks for exploiting them, but did not build a complete attack chain on its own under testing conditions. OpenAI maintains that the model has not reached the Cyber Critical threshold in its Preparedness framework, yet it still chose a gradual rollout because the capability level is very different from before.
OpenAI itself says it doesn't want a pattern of asking the government for permission before every release to become a permanent default, because that keeps good tools out of the hands of the users, developers and cyber defense teams who need them. It's a difficult balance between widening access and controlling risk.
On the protection side, the company stacks multiple layers, as in the diagram above. The base layer is training the model to refuse prohibited requests, even when users disguise their intent or try to jailbreak it. On top of that are real-time filters for both cybersecurity and biology, scanning content as it is being generated. When a violation is suspected in a high-risk case, the response can be paused so a larger reasoning model can review the context, and it is held back if not approved.
Higher still is an account-level review layer, which looks across multiple conversations to separate repeated malicious behavior from legitimate security work. OpenAI acknowledges that during the preview period the filters will sometimes wrongly block valid work, and it uses user feedback to gradually reduce unnecessary blocks. To keep the safety stack from breaking against new tricks, the company devoted more than 700,000 A100-equivalent GPU hours to automated red-teaming that hunts for universal jailbreaks, combined with outside expert teams and a rapid-response process for patching new vulnerabilities.
GPT-5.6 API pricing, prompt caching, Cerebras speed and when you can use it
GPT-5.6 is billed per one million tokens, with three price levels by model. This pricing applies to both input and output tokens, so the real cost depends heavily on how long a response you ask for.
Looking at the pricing chart, the gap between the three versions is fairly large, especially on output tokens. Sol is the most expensive, Luna is the cheapest, and Terra sits in the middle and is probably the best deal: on par with GPT-5.5 but half the price, so it will likely be the version most users land on as access expands.
The table below lists each version's role and price for easy comparison.
| Version | Role | Input price (1 million tokens) | Output price (1 million tokens) |
|---|---|---|---|
| Sol | Top tier, hard tasks | 5 USD | 30 USD |
| Terra | Balanced, everyday work | 2.50 USD | 15 USD |
| Luna | Fast, low cost | 1 USD | 6 USD |
GPT-5.6 also adds a more predictable way of caching context, letting you set explicit cache breakpoints and keeping the cache for at least 30 minutes. Writing to the cache is billed at 1.25 times the regular input price, but reading back from the cache still gets a 90% discount, which pays off when you repeat the same context over and over.
On speed, OpenAI says it will run GPT-5.6 Sol on Cerebras hardware at up to 750 tokens per second in July, initially limited to some customers. As for availability, these versions first arrive through the API and Codex for the selected group, then will open to ChatGPT, Codex and API users over the coming weeks. For businesses, the company is also working on approaches such as privacy-preserving violation detection and access granted according to risk level.
What laptop users get from GPT-5.6
One thing needs to be made clear to avoid confusion. GPT-5.6 runs in the cloud. You call it through the API or use it in Codex and ChatGPT, rather than downloading the model and running it directly on your machine. That means you don't need a monster machine just to use GPT-5.6.
But what makes work smooth or choppy is the machine you're typing on. When you open ChatGPT or Codex in a browser and put it to real work, the experience depends heavily on your machine's configuration.
Specifically, when you let Codex spin up subagents in ultra mode, open multiple terminals, build projects and run containers in parallel, what eats up resources is your machine's RAM, multi-core CPU and fast SSD. A laptop with 16GB of RAM or more, a many-core CPU and a fast NVMe drive will keep the write-code-then-test loop from stuttering. A large screen and good battery life also make long coding sessions easier to get through.
If you're planning to buy a machine to work with AI and Codex properly, Hung Phat can advise on a configuration that fits your budget and your programming needs. Take a look at the mobile workstations and high-spec laptops currently in stock to pick one that matches the kind of work you do.
When can I use GPT-5.6 Sol?
For now, Sol, Terra and Luna are only in limited testing through the API and Codex for a group of trusted partners. OpenAI says it will expand access to ChatGPT, Codex and API users over the coming weeks.
Does GPT-5.6 run offline on a laptop?
No. It is a cloud model, and you access it through the API or through Codex and ChatGPT. Your computer only needs to be powerful enough to write code and run the accompanying tools, not to run the model locally.
Which should I choose: Sol, Terra or Luna?
Choose Sol for hard tasks that need deep reasoning, Terra for everyday work because it balances price and power, and Luna when you need to process large volumes at low cost. They share a generation, so the foundation is similar and they differ in capability level and price.
How is GPT-5.6 priced?
It is billed per one million tokens. Sol is 5 USD for input and 30 USD for output, Terra is 2.50 and 15 USD, and Luna is 1 and 6 USD. There is also a caching mechanism that reduces costs when you repeat the same context.
