On 9 July 2026, OpenAI released GPT-5.6, which it calls its most capable system yet, to the public. It came in three tiers: Sol, the top model; Terra, a mid-range option; and Luna, a fast, cheap tier. It also arrived with something that may matter more for everyday work, an agent named ChatGPT Work.

The names are a deliberate change from the old “mini” and “nano” labels. OpenAI explains that “the number identifies a model’s generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence.” That means 5.6 tells you the generation, and the name tells you which job the model is built for.

Sol, Terra and Luna: one family, three jobs

The three tiers are split by what they can do and what they cost, not by size. Sol is the flagship, built for hard problems and long tasks it runs on its own. Terra is a balanced model that matches the older GPT-5.5. Luna is the smallest and cheapest.

The prices show how far apart the tiers sit. OpenAI lists Sol at $5 input and $30 output per million tokens, Terra at $2.50 and $15, and Luna at $1 and $6. OpenAI says Terra runs at “competitive performance to GPT‑5.5 while being 2x cheaper.” That halved cost, rather than any jump in capability, may be the bigger deal for anyone running large volumes of routine work.

The performance claims are strong, but keep in mind that they come from OpenAI. CEO Sam Altman told CNBC that Sol “is 54% more token efficient on agentic coding” . One detail businesses should note: OpenAI’s own safety report rates all three tiers at its “High” risk level for both cyber and biological or chemical misuse. The cheaper Terra and Luna are not lightweight in that sense.

ChatGPT Work: from answering to finishing

The agent is where OpenAI tries to change what ChatGPT is for. ChatGPT Work blends the company’s Codex coding tool with ChatGPT to handle longer tasks. OpenAI says it “can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work.”

The promised outputs are finished items: spreadsheets, slide decks, documents, even small web apps.

That shift, from a chatbot that answers questions to an agent that goes off and returns completed work, is OpenAI’s own pitch, not a tested result — and OpenAI’s own advice is cautious. It suggests that “the best way to learn how to use ChatGPT Work is to give it a task you already know well,” a quiet way of saying you will need to check its work against something you understand.

The launch also rearranged the desktop experience. Codex is moving into a new ChatGPT desktop app for macOS and Windows, sitting next to Chat and Work instead of running on its own.

The two weeks in a holding pattern

Before the public got GPT-5.6, a small group of vetted organisations had it. Early access went to about 20 government-vetted partners at the request of US officials, instead of the broad release OpenAI says it first wanted. The setup followed an executive order that created a voluntary review of the most capable models before release.

However, a White House official told CNBC that the administration did not give OpenAI a “green light, approval or clearance” to release its models, and that such decisions “rest entirely with the companies.” 

Altman’s case for working with reviewers rests on a trade, which he framed this way: “If you want broad access, which we do, and you have powerful models, you really want to be able to be confident in your safety claims.” His read on the market is blunter: “Every enterprise now is thinking about spend and the value they’re getting in exchange for AI,” Altman said, which points to the reasoning behind Terra’s halved price as much as any technical claim.

What to watch next

One date is already set. GPT-5.4 is due to be retired on 23 July. The rollout is staged, so if Sol does not yet appear in your model picker, that is not a contradiction of the announcement; access depends on your plan, region, and timing.

The bigger open question is not about benchmarks. OpenAI has spent years making models better at answering; its bet with ChatGPT Work is that letting an agent complete whole jobs will land the same way. Giving an agent access to your apps and files for hours could return finished work, or it could return finished mistakes at a scale that is harder to catch. A coding leaderboard won’t settle that — the first few weeks of people handing it tasks they know well enough to spot the errors will.