The three GPT-6 models at a glance
| Model | OpenAI's positioning | Starting effort | API (per 1M, in / out) | Codex credits (per 1M, in / cached / out) |
|---|---|---|---|---|
| GPT-6 Astra | The hardest end-to-end work: complete workflows across code, apps and research that need sustained reasoning and judgment | Light (low in the CLI) | $10 / $50 | 250 / 25 / 1,250 |
| GPT-6 Sol | Everyday and complex work: complex code changes, deep research, polished documents, agentic workflows | Medium | $2 / $10 | 50 / 5 / 250 |
| GPT-6 Luna | Clear, repeatable, high-volume tasks: extraction, classification, transformation, structured summaries, focused coding | High | $0.10 / $0.50 | 2.5 / 0.25 / 12.5 |
OpenAI says it trained Sol and Luna with methods similar to Astra's, bringing much of Astra's strength into faster, cheaper models, and cut API prices by half versus GPT-5.6 promotional pricing (Sol from $4 / $20 to $2 / $10, Luna from $0.20 / $1.2 to $0.10 / $0.50). GPT-6 Astra remains the strongest model in the family.
Where you can use them, and who can
- Surfaces: you pick Sol and Luna by name in ChatGPT Work and Codex — the desktop app, Work on the web, Codex CLI, the IDE extension. The 2026-10-07 release changed nothing there: same models, allowances and credit rates. In Chat you cannot pick them by name: since 2026-10-07 the GPT-6 option in the model menu is powered by GPT-6 Sol on Plus, Pro, Business and Enterprise, but it is not labelled Sol, and you choose a thinking level instead (Medium or High on Plus; Extra High and Pro too on Pro, with the Pro level still on GPT-6 Astra).
- Plans: in Work and Codex, Plus, Pro, Business, Enterprise and Edu get Sol and Luna; Free and Go get Luna in the desktop app (rolling out by account). In Chat, OpenAI's announcement and release notes say Free and Go get GPT-6 powered by GPT-6 Luna from 2026-10-08, while the Help Center model page still listed GPT-5.6 Luna for them on 2026-10-10; the rollout is gradual, so check your own model menu. Free and Go have none of the Medium, High, Extra High or Pro levels in Chat; they have Think.
- Cloud tasks: they share one allowance with local messages and may use more of it (OpenAI's pricing page, checked 2026-10-07; the late-September version said cloud chats use GPT-5.6 Sol, and that line has been removed). How to use them: Codex Cloud, step by step.
- Transition: GPT-5.6 Sol, Terra and Luna stay selectable during the rollout; picking a new model doesn't change workspace permissions.
- API:
gpt-6-solandgpt-6-luna, 1,050,000-token context (922,000 max input, 128,000 max output), knowledge cutoffs 2026-04-20 (Sol) and 2026-05-18 (Luna).
How much Plus gets, and how the Pro tiers work
For Plus, OpenAI publishes ranges of local messages per five hours — estimates, not hard caps; task size, context and effort move them, and a weekly limit also applies. Local messages and cloud tasks share one allowance. As of 2026-09-30 OpenAI's pricing doc says the Pro tiers (Pro 100 / 200 / 500) have no 5-hour limit in Codex, only a weekly allowance, so the table lists Plus only:
| Model | Plus (per 5 hours) |
|---|---|
| GPT-6 Astra | 5–45 |
| GPT-6 Sol | 15–150 |
| GPT-6 Luna | 350–3,000 |
| GPT-5.6 Sol | 10–100 |
Two things stand out: GPT-6 Sol gets more messages than GPT-5.6 Sol on the same plan, so switching is free allowance; and Luna's range is an order of magnitude higher than Sol's, so routing bulk work to Luna saves your Sol allowance.
In credits: what the same task costs
After your plan allowance, Codex bills credits per token. For a medium coding task (80,000 uncached input + 220,000 cached input + 20,000 output tokens — our illustrative example), OpenAI's rates give about 10 credits on GPT-6 Sol, 20 on GPT-5.6 Sol, 0.5 on GPT-6 Luna and 51 on GPT-6 Astra. In other words, GPT-6 Sol's credit rate is exactly half of GPT-5.6 Sol's. Fast mode on the GPT-6 models bills purchased credits at 2× (2.5× against the plan's included allowance). Full tables in How far do 500 Codex credits go?
OpenAI's published numbers (vendor-run)
All from OpenAI's launch post. They are OpenAI's own evaluations, run in its research environment, not independently reproduced — treat them as indicative:
- AutomationBench (business workflows across apps): GPT-6 Sol (xhigh) 33.2% at $0.27 per task; GPT-6 Astra (low) 30.3% at 3.9× the cost; Claude Opus 5 (max) 26.9% at 11.1× the cost.
- DeepSWE v1.1 (complex engineering in real codebases): GPT-6 Sol (max) 68.8%, GPT-6 Luna (max) 66.6%.
- OSWorld 2.0 offline (computer use): GPT-6 Sol (xhigh) 60.5%, on par with Claude Opus 5 (medium) at 60.3%, at roughly 80% lower cost per task per OpenAI.
- Factuality: on OpenAI's internal evaluation, GPT-6 Sol makes about half as many factual mistakes as its predecessor.
- Style: clearer, less jargon, fewer low-value details, slightly shorter answers.
Which to pick for which job
| Task | Model | Starting effort |
|---|---|---|
| Bulk extraction, classification, transformation, structured summaries — you know what a good result looks like | GPT-6 Luna | High |
| Everyday code changes, multi-file features, code review, polished documents | GPT-6 Sol | Medium |
| Ambiguous, difficult, high-value tasks (let it ask questions first) | GPT-6 Sol | High / Extra High |
| Complete long workflows across code, apps and research | GPT-6 Astra | Light to start |
| The single hardest problem, when depth beats speed | Current model on Max | Max |
| Big tasks that split into parallel parts | A model that supports Ultra (Luna goes up to Max, not Ultra) | Ultra |
The desktop Power slider has six presets — Luna High, Sol Light, Sol Medium, Astra Light, Astra Medium, Astra Extra High — starting at Sol Light; open Advanced to pick a specific model, effort or speed. OpenAI's caveat: efforts don't map one-to-one between model generations, so try a familiar task one notch lower and adjust.
How to switch: CLI, config, desktop app
# use GPT-6 Sol for this session
codex --model gpt-6-sol
# non-interactive run
codex exec -m gpt-6-sol "Review the current changes"
# long-term default: ~/.codex/config.toml
model = "gpt-6-sol"After GPT-5.5's retirement: on 2026-10-14 GPT-5.5 left ChatGPT, Work and Codex (the API is unaffected). If you sign in to Codex with ChatGPT, replace any remaining hard-coded gpt-5.5 in workspace defaults, saved settings, managed configs, custom agents, scheduled tasks and scripts: gpt-6-sol on Plus and above, gpt-6-luna in the desktop app on Free / Go. The already-retired gpt-5.4 and gpt-5.4-mini map to gpt-6-sol and gpt-6-luna.
Which plan for GPT-6 Sol
- Just curious: Free / Go get GPT-6 Luna in the desktop app at no cost.
- Everyday coding: ChatGPT Plus on your own account (¥185, the account must be on Free or an expired Plus) or a ready-made Plus account (¥199). Sol and Luna included, GPT-6 Astra limited.
- Hitting limits three or more times a week: Pro 100 (formerly Pro 5x) on your account (¥850, Free, Plus or Pro accounts) or a ready-made Pro 100 account (¥899): no 5-hour limit in Codex, plus GPT-6 Pro in chat; for more, see Pro 200 and Pro 500.
- The odd overrun: Codex credits ¥180 per 500, 1–20 packs per order, valid 12 months; not for Free accounts.
Pay with USDT on TRC20 or BEP20; top-ups usually land in 5–30 minutes; if anything goes wrong, contact the bottom-right live chat promptly.