Sol, Terra, Luna: ChatGPT's Model Lineup, Explained

Your ChatGPT model picker changed twice this month, and the cheapest model in it now beats the flagship everyone was building on in April.

Here is what actually matters, in five minutes.

OpenAI did something useful this year: it stopped naming models after version numbers alone. The GPT-5.6 family, generally available since 9 July 2026, splits into three named tiers, Sol, Terra, and Luna. The number tells you the generation. The name tells you the capability tier, and OpenAI says those tiers are durable and advance on their own cadence.

So the next flagship will still be called Sol. Only the number in front of it moves. That is a far saner mental model than the GPT-4o / o3 / 4.1 / 4.5 alphabet soup we spent 2025 untangling.

The three tiers

Two notes on that pricing.

OpenAI cut Luna by 80% and Terra by 20% on 30 July, then cut Sol's API and credit pricing by more than 20% on 21 August. That last one is a three-month promotional window, not a permanent rate. If you are budgeting an API project around Sol, budget for it expiring.

There is also a Sol Pro option for Pro and Enterprise users, plus effort settings that change behaviour meaningfully. max gives the model more time to reason and revise. ultra coordinates four agents in parallel by default, trading tokens for quality on genuinely hard tasks, Pro and Enterprise only in ChatGPT Work, Plus and above in Codex.

What changed inside ChatGPT

The API lineup and the ChatGPT model picker are not the same thing. This is where most people get confused.

Free and Go accounts: since 6 August, Luna is the default, replacing GPT-5.5 Instant. Text chats became unlimited, though limits still apply to file uploads, images, and other tools. A Think button appeared for harder questions.

Plus and Pro: now default to Sol. And here is the part people keep missing: the old Instant and Thinking entries have been folded into one model with a reasoning effort slider. If you were switching manually to Thinking for hard prompts, that habit is now the slider.

Retired: GPT-5.1 went in March, GPT-4.5 in June, and o3 was retired from ChatGPT on 26 August 2026. Any tutorial telling you to switch to o3 for reasoning is describing a picker that no longer exists.

What the benchmarks show

The headline claim for GPT-5.6 is not raw capability. It is intelligence per token, OpenAI is competing on efficiency as much as on scores.

Selected results from OpenAI's published tables, with the previous flagship for reference:

The floor moved, not just the ceiling. Luna costs $1 per million input tokens and beats the model most teams were building on three months ago.

That is the actual story of this release. Terra sits between the two and beats GPT-5.5 on effectively everything.

Two honest caveats.

These are OpenAI's own numbers, on OpenAI's harnesses. Competing labs publish tables where their models win. Treat any first-party benchmark as directional, not settled.

And the lead is not uniform. On SWE-Bench Pro, Sol posts 64.6% against Claude Fable 5's 80%. On the Artificial Analysis Intelligence Index, Sol sits at 58.9 against Fable 5's 59.9. OpenAI's answer in those cases is cost and latency, not the score, a legitimate argument, but a different one.

Long context is another place to check before committing. Luna drops sharply on multi-needle retrieval past 256K, 41.3% against Sol's 91.5%. If your workflow stuffs whole document sets into context, the cheap tier is a false economy.

So which one should you use?

On Plus? You are on Sol already. The highest-leverage thing you can do is learn the reasoning effort slider. Low effort for drafting, summarising, reformatting. High or max when the task has a right answer you cannot easily verify yourself, such as financial modelling, code review, or technical analysis.

On Free or Go? Luna plus the Think button covers far more than people assume, and text chats are unlimited. For teaching contexts and student use, this tier is genuinely sufficient.

Building on the API? Start on Terra, not Sol. Terra beats the old flagship at half of Sol's price, and OpenAI's own partner quotes describe agents that ran on GPT-5.5 performing just as well on Terra at half the cost. Move up only where you can measure what the extra spend buys.

Running agents or long tool chains? Look at Programmatic Tool Calling in the Responses API before you look at a bigger model. Letting the model write and run small programs that coordinate tools, instead of round-tripping every tool response through the model, is where the reported savings come from. Several partners reported 20–60% fewer tokens on identical workflows.

What's coming

OpenAI publicly named its next frontier model, Astra, on 1 August 2026. It has not shipped. In August, OpenAI added monitoring over Astra's chain of thought designed to halt the model if it moves toward high-risk activity, and later revised its safety rules following a security incident elsewhere in the ecosystem.

There is a signal in the cadence, too. GPT-5.5 to GPT-5.6 took roughly eleven weeks, a noticeable slowdown from the three-week gaps of late 2025. Part of that is regulatory: GPT-5.6 slipped from a planned June launch to a limited preview for trusted partners after a US government request for early access to frontier models. That is now a real variable in when any of us get new models.

The one-line version

The flagship got better. The cheap models got much better. For most people reading this, the right upgrade is not a bigger model, it is using the tier you already have more deliberately.

Benchmark figures are from OpenAI's own GPT-5.6 release materials and help-centre release notes, July–August 2026, and should be read as directional.

If this was useful, subscribe, I write about LLM releases and practical AI workflows here every week.