In This Article
Key Takeaways
- GPT-5.6 reached general availability on July 9, 2026, across ChatGPT, Codex, and the OpenAI API, in three tiers: Luna, Terra, and Sol.
- Luna is the fast, cheap tier ($1 / $6 per million tokens); Terra is the balanced everyday tier ($2.50 / $15) that OpenAI positions as competitive with the older GPT-5.5; Sol is the flagship ($5 / $30) built for hard, long-running work.
- Sol is tuned toward frontier work in cybersecurity, life sciences, chemistry, and quantitative biology, and it powers the "ultra" reasoning mode in Codex.
- All three share a 1M-token context window, 128K max output, and a February 16, 2026 knowledge cutoff.
OpenAI's naming finally makes the pricing decision visible. On July 9, 2026, the company brought GPT-5.6 to general availability not as a single model but as three: Luna, Terra, and Sol, rolling out across ChatGPT, Codex, and the API. Independent developer Simon Willison's write-up confirmed the same three-tier structure on launch day. This is our analysis of what the split actually buys you — because the interesting decision with GPT-5.6 is no longer "should I use it," but "which one, and for what."
What OpenAI shipped on July 9
Three models under one version number, sized from smallest to largest: Luna, Terra, Sol. They share the same plumbing — a one-million-token context window, up to 128,000 output tokens, and a February 16, 2026 knowledge cutoff — and differ in capability and price. Alongside the models, OpenAI shipped an agent it calls ChatGPT Work, and the family also landed in GitHub Copilot the same day. The headline is not a single new best model; it is a menu.
That menu is the point. For the last few years the model choice was mostly binary — the flagship or the cut-down cheap one — and teams routinely overpaid by sending trivial work to an expensive model out of habit. Three named tiers make the trade-off explicit at the moment you pick a model, which is exactly when you should be thinking about it.
The three tiers, and what each is for
Luna is the fast, inexpensive tier — the one you reach for when volume and latency matter more than raw depth: classification, extraction, routing, simple drafting, high-throughput pipelines. Terra is the balanced everyday model. OpenAI positions it as competitive with the older GPT-5.5 at a lower price point, which makes it the natural default for most general work — the tier that quietly handles the bulk of real traffic. Sol is the flagship, built for hard problems and long tasks it runs largely on its own.
Sol is also where the specialization lives. Reporting on the release describes Sol delivering its biggest gains in cybersecurity, life sciences, chemistry, and quantitative biology, per Tech Wire Asia's launch coverage. On long-horizon evaluations it posts real numbers: benchmark write-ups report Sol scoring 53.6 on Agents' Last Exam, an evaluation of extended professional workflows across dozens of fields. Treat vendor benchmarks as a reason to test on your own workload, not as a verdict — but the pattern is clear: Sol is aimed at the deep end.
Pricing side by side
The clearest way to see the strategy is to line up the API prices. All figures are per million tokens, as reported at launch by Simon Willison.
GPT-5.6 tiers at a glance
| Tier | Input / output (per 1M) | Best for |
|---|---|---|
| Luna | $1 / $6 | Fast, high-volume, cost-sensitive work |
| Terra | $2.50 / $15 | Balanced everyday tasks; GPT-5.5-class quality |
| Sol | $5 / $30 | Hard problems, long autonomous tasks, science & security |
Sol costs five times Luna on input and output alike. That spread is the whole argument for reading this before you wire an app to the flagship by reflex. OpenAI's own framing leans on efficiency: it says GPT-5.6 at medium reasoning beats the prior generation of rivals at a fraction of the cost, with the smaller tiers doing it more cheaply still. Whether that holds for your task is an empirical question — run the numbers on our LLM price calculator, and if you are weighing GPT-5.6 against Claude or Gemini, our AI model picker lays the options next to each other.
Sol, Codex, and the "ultra" mode
The piece developers keep asking about is how Sol shows up in Codex. GPT-5.6 introduces a reasoning setting OpenAI calls "ultra," which it describes as going beyond a single agent by coordinating subagents across parallel work to finish complex tasks faster. In Codex and ChatGPT Work you can choose Sol, Terra, or Luna with effort controls, and launch coverage notes that ultra is available to Plus and higher plans in Codex. So "Sol Ultra powers Codex" is roughly right, stated precisely: Codex lets you run the flagship at its highest reasoning effort, and that combination is gated to paid tiers.
"Ultra" is a setting, not a fourth model
It is easy to read "Sol Ultra" as a separate product. It is not. Ultra is the top reasoning effort within GPT-5.6, and it burns more tokens because it fans work out to subagents. That makes it powerful for genuinely hard, multi-step jobs — and needlessly expensive for anything a lower effort would have finished. Reach for it deliberately, not by default.
How to pick a tier without overpaying
Our practical rule is the same one that has worked for every tiered model line: start at the bottom and move up only when the output fails. Draft your prompt against Luna. If the answers are shallow or the tool use wobbles, move to Terra, which will carry most real work. Escalate to Sol only for the tasks that actually justify it — long autonomous runs, deep coding sessions, and the science and security domains it was tuned for. Reserve ultra effort for the hardest of those.
If your work is mostly coding, the tier choice interacts with your editor. We compared the current crop of coding assistants in our 2026 guide to AI coding assistants, and if you live in an agentic editor, our Cursor course covers how model selection changes cost and behavior. The through-line: the model is now a dial, not a switch, and the savings come from turning it no higher than the job needs.
Zoom out and this release is less about a new capability ceiling than about pricing legibility. When OpenAI first opened this family up it paired it with a job-doing agent, which tells you where the company thinks the value is: not in a smarter chatbot, but in cheaper tiers you can point at long, real work. We traced that shift when the prior generation shipped, in GPT-5.5: agents for everyone. GPT-5.6 continues it, and hands you three price points to do it with.
Match the tier to the job
Use our free model picker and price calculator to compare Luna, Terra, and Sol against Claude and Gemini on your own token volumes — before you commit an app to the flagship out of habit.
Open the model pickerSources: OpenAI — GPT-5.6; Simon Willison; GitHub Changelog; Tech Wire Asia; CodeRabbit benchmark; TestingCatalog. Analysis and framing by Precision AI Academy.
Common questions
Which tier is the default in ChatGPT? ChatGPT routes general use to the mid and lower tiers, and lets paid users select Sol and higher effort in Codex and ChatGPT Work. For API calls, you choose the tier explicitly.
Is Terra really as good as GPT-5.5? OpenAI positions Terra as competitive with GPT-5.5 at a lower price. As always, "competitive" is a claim to validate on your own tasks, not a guarantee across the board.
Do all three share the same context window? Yes — reporting puts all three at a one-million-token context window, 128,000 max output tokens, and a February 16, 2026 knowledge cutoff.
What is ChatGPT Work? An agent OpenAI shipped alongside GPT-5.6, built to carry out whole jobs rather than answer single questions. It uses the same three tiers under the hood.