Table of Contents
- 1. What GPT-6 Sol and Luna are — release date, specs and positioning
- 2. Where you can use them — Work, Codex and the API, but not regular Chat
- 3. Pricing — "1/5 and 1/100 of Astra" only holds at the same token count
- 4. The official guidance — Sol at Medium, Luna at High, Astra from Light
- 5. OpenAI's numbers — the eval where Sol xhigh beat Astra low
- 6. Third-party numbers — the Artificial Analysis comparison
- 7. Our small-scale test — short tasks where cheaper settings were enough
- 8. How to choose for your own work — where the decision turns
- FAQ
On September 22, 2026 (US time; the early hours of September 23 in Japan), OpenAI released GPT-6 Sol and GPT-6 Luna (OpenAI's announcement). They are faster, cheaper models built with the same line of training as GPT-6 Astra, the top model that came out on September 3.
This article is written especially for people who currently use Astra at low (shown as "Light" in the app) in Codex or ChatGPT Work. The short answer: standardizing on Astra low is not a mistake. But OpenAI's own guidance by use case is Sol for everyday work and complex coding, Luna for well-defined repetitive work, and Astra for the hardest work, and "Astra from Light" means "if you use Astra, start at Light". It is not advice to run every job on Astra.
We report the numbers separately for each publisher: OpenAI's evals, the third-party evals from Artificial Analysis, and the small test we ran ourselves use completely different methods and tasks.
GPT-6 Sol / GPT-6 Luna
Released September 22, 2026 (US time) / API, ChatGPT Work, Codex
Sources: OpenAI API documentation, ChatGPT Models
1. What GPT-6 Sol and Luna are — release date, specs and positioning
As OpenAI describes them, Sol is aimed at complex coding and agentic workflows, and Luna is its most efficient model, for focused, high-volume work. The announcement frames the release as bringing Astra's strengths (professional work, factuality, coding and computer use) to faster, more affordable models, while also saying that the most demanding, high-stakes projects still call for Astra's depth.
| Item | GPT-6 Sol | GPT-6 Luna | For reference: GPT-6 Astra |
|---|---|---|---|
| Model ID | gpt-6-sol | gpt-6-luna | gpt-6-astra |
| Context | 1,050,000 | 1,050,000 | 1,050,000 |
| Max input / max output | 922,000 / 128,000 | 922,000 / 128,000 | 922,000 / 128,000 |
| Knowledge cutoff | April 20, 2026 | May 18, 2026 | April 30, 2026 |
| Reasoning effort (API) | none to max (default medium) | none to max (default medium) | low to max (no none) |
| Input / output | Text, images / text | Text, images / text | Text, images / text |
Source: the model pages in the OpenAI API documentation (Sol, Luna, Astra). Checked September 24, 2026.
The surprise is the knowledge cutoff: Luna's (May 18) is more recent than Astra's (April 30). If you use the models with web search, though, the difference hardly matters (knowledge cutoff dates of major AI models).
One note for developers. Sol and Luna accept none (no reasoning) in the API, but function calling in Chat Completions requires none, and OpenAI points you to the Responses API for built-in tools and function calling.
2. Where you can use them — Work, Codex and the API, but not regular Chat
Putting the announcement and the Models and Pricing pages together gives the picture below. The way in is ChatGPT Work and Codex; the models do not appear in the regular chat interface (Chat).
Both Sol and Luna in ChatGPT Work and Codex. The rollout started on release day and runs gradually over the course of the day; if you don't see them yet, OpenAI suggests checking again later. On Enterprise and Edu, an admin may need to enable Luna.
Luna only, in the desktop app (standard speed, subject to rollout). Sol is not included.
Available as gpt-6-sol and gpt-6-luna, billed pay-as-you-go for what you use.
Not available. The announcement says the models are not yet offered in Chat, and the Models page says they are available in Work and Codex but not in Chat.
The older models won't disappear right away. The Models page says GPT-5.6 Sol, Terra and Luna remain available during the transition period. On the other hand, GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on October 14, 2026 (the API is not affected), and the page names Sol as its replacement on paid plans (Plus, Pro, Business, Enterprise, Edu) and Luna in the desktop app on Free and Go. If you specify gpt-5.5 in a config file or a scheduled task, you need to change it before then (ChatGPT Models).
3. Pricing — "1/5 and 1/100 of Astra" only holds at the same token count
| Model | API input | Cached input | API output | Codex credits (input / cached / output) |
|---|---|---|---|---|
| GPT-6 Astra | $10 | $1 | $50 | 250 / 25 / 1,250 |
| GPT-6 Sol | $2 | $0.20 | $10 | 50 / 5 / 250 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 | 2.5 / 0.25 / 12.5 |
Per 1M tokens, standard speed. Sources: the model pages in the OpenAI API documentation and ChatGPT / Codex Pricing. Checked September 24, 2026.
On every line, Sol is exactly 1/5 of Astra and Luna is 1/100, and Codex credits follow the same ratios. OpenAI describes the cut from the previous generation as 50% below GPT-5.6's promotional pricing, but to be precise, Sol is half price on both input and output, while Luna is half price on input and $1.20 → $0.50 on output (about 58% lower).
The ratio of unit prices is not the ratio of what a job costs. Your bill is unit price × the tokens actually used, and cheaper models sometimes think longer and produce more output. In the third-party measurements from Artificial Analysis (section 6), Luna high used about 5 times as many output tokens per task as Astra low, and its cost per task came to about 1/27, not 1/100. Sol medium was likewise about 1/3, not 1/5.
The ratios only apply to the arithmetic where token counts are the same. For example, with 100,000 input tokens (80,000 of them cached) and 10,000 output tokens, the API cost comes to Astra $0.78, Sol $0.156, Luna $0.0078. Other conditions apply: input over 272,000 tokens doubles the input and cached-input rates and multiplies output by 1.5; API cache writes cost 1.25 times the input rate (Codex credits carry no separate charge for writes); and Fast mode costs 2 times as much in the API and 2.5 times in Codex credits.
Unit prices don't tell you how much your subscription covers
When you use the models through a ChatGPT plan, you can't say that Sol gets you 5 times as much use as Astra, or Luna 100 times. Codex's own pricing page says that credit rates alone do not determine how much usage a subscription includes.
The same page gives rough ranges for messages per 5 hours. On Plus, they are Astra 5–45, Sol 15–150 and Luna 350–3,000. Comparing low end to low end and high end to high end, Sol gets about 3 times Astra and Luna about 67–70 times, which doesn't match the 5× and 100× from unit prices. On top of that, OpenAI notes that these are not fixed limits, and that similar tasks can consume different amounts depending on the model, context, reasoning, tool use and caching. The only way to know how many you get for your own work is to check the usage screen.
4. The official guidance — Sol at Medium, Luna at High, Astra from Light
On which model to use, OpenAI's Models page gives clear guidance. In summary:
Long-running work across code, apps and research that has to keep its judgment the whole way. Starting point: Light (setting value low).
Ambiguous, difficult, high-value work (complex code changes, deep research, documents that need polish). Starting point: Medium. For narrow tasks, hand it a clear definition of done.
High-volume work where you know what a good result looks like (extraction, classification, transformation, summaries in a fixed format). Starting point: High.
Three things are easy to mix up here.
- Starting points by use case: the "Sol at Medium, Luna at High, Astra from Light" above.
- The app's default presets: the Power presets in the desktop app and on the web are six — Luna High, Sol Light, Sol Medium, Astra Light, Astra Medium and Astra Extra High — and Sol Light is selected by default (some paid plans don't have Astra Extra High).
- The API default:
mediumfor both Sol and Luna.
All three differ, but they don't contradict each other. The default preset is "fast and light for now"; the use-case guidance is "where to start to get results on that kind of work". OpenAI also says that reasoning effort levels don't map one-to-one across generations, and suggests trying familiar tasks at a lower setting and adjusting from there. You can't pick a setting for the new Sol with the same feel you had for "GPT-5.6 Sol at high". For a comparison with the previous generation, see our measured comparison of Astra low and GPT-5.6 Sol high (a different model from this GPT-6 Sol).
How to read "Astra from Light". It is the starting point once you have chosen Astra. It doesn't mean every job should go to Astra. Standardizing on Astra low is a reasonable choice if you want to skip the switching and put capability first. But whether it is optimal once usage limits and waiting time are counted is a separate question, and if you care about that, splitting work with Sol and Luna is worth considering.
5. OpenAI's numbers — the eval where Sol xhigh beat Astra low
OpenAI's announcement contains one figure that puts Sol and Astra side by side: AutomationBench 1.0.6, an eval of workflows that span business apps.
| Model (reasoning effort) | Score | Cost per task |
|---|---|---|
| GPT-6 Sol (xhigh) | 33.2% | $0.27 |
| GPT-6 Astra (low) | 30.3% | 3.9× Sol |
| Claude Opus 5 (max) | 26.9% | 11.1× Sol |
Source: the table in OpenAI's announcement (the vendor's own evaluation).
On this eval alone, Sol pushed to think hard scores above Astra low at about a quarter of the cost. Note, though, that the comparison uses Sol at xhigh, not medium or low. The other figures OpenAI gives are:
- DeepSWE v1.1 (software development in real codebases): Sol max 68.8%, Luna max 66.6%.
- Agents' Last Exam (long professional tasks): Sol max 56.4%.
- OSWorld 2.0 offline (operating a computer): Sol xhigh 60.5%. OpenAI itself says Astra remains the best for computer use.
- Factuality: in an internal eval built from past conversations where users flagged errors, Sol made about half as many errors as the previous generation (GPT-5.6 Sol).
Two caveats when reading these. The factuality eval is a collection of conversations where errors are likely, and OpenAI itself notes that it is not representative of typical use. OpenAI also says the evals were run in a research environment or through the API, and that ChatGPT's production product uses different system instructions and tools, so outputs may differ slightly. You can't take "half the errors" and apply it directly to your own day-to-day work.
6. Third-party numbers — the Artificial Analysis comparison
The independent evaluator Artificial Analysis (AA below) has published comparison pages that put Sol low, Sol medium and Luna high each against Astra low. The table below takes the values from those three pages; Astra low's values were identical on all three.
| Setting | Intelligence Index | Terminal-Bench 4.0 | AutomationBench-AA | Cost per task | Time per task |
|---|---|---|---|---|---|
| Astra low | 46 | 42% | 59% | $0.82 | 97.18 s |
| Sol medium | 40 | 19% | 58% | $0.25 | 59.99 s |
| Sol low | 34 | 9% | 54% | $0.13 | 34.98 s |
| Luna high | 32 | 5% | 48% | $0.03 | 146.71 s |
Source: Artificial Analysis comparison pages (Sol medium vs Astra low, Sol low vs Astra low, Luna high vs Astra low). Intelligence Index v4.3.2 (10 evals). Checked September 24, 2026. The values in the table may be updated.
Three things stand out in this table.
On Terminal-Bench 4.0, Astra low scores 42%, against 19% for Sol medium and 5% for Luna high. This gives no grounds for replacing Astra across the board in hard, terminal-driven development.
On AutomationBench-AA, Sol medium scores 58% against Astra low's 59%, at about a third of the cost per task. The size of the gap varies enormously by eval. Don't convert the composite index gap into "X% smarter".
Luna high costs about 1/27 of Astra low, but takes 146.71 seconds per task, longer than Astra low's 97.18 seconds, because it spends more tokens thinking.
AA's time per task is generation time calculated from output speed; it doesn't include the wait for the first response or any surrounding processing. It is not the same as how long a job takes in your own environment. On AA's knowledge-accuracy metric (AA-Omniscience), Astra low scores 41, Sol low and medium both 27, and Luna high −6, a wide spread that shows choosing Luna on price alone is risky for answering factual questions without supplied sources.
7. Our small-scale test — short tasks where cheaper settings were enough
At AI Arte we also ran a small test with the Codex CLI. It is not a test for ranking the models. Here are the conditions first.
Tasks: (1) write a function that handles choosing among duplicate IDs, tie-breaking on identical timestamps, missing rows and so on (graded by running 24 cases); (2) classify 8 claims from source material, compute error rates with different denominators, and write a recommendation of up to 160 characters (graded on 12 items). No AI grading was used.
Limitations: the cache was not cleared, and most of the input (about 11,000–12,000 tokens) was read from cache. The grading code was written after the answers came in; this was not a pre-registered test.
| Setting | Passed | Median time | Avg. output tokens (incl. thinking) | API-equivalent cost per run (avg.) |
|---|---|---|---|---|
| Astra low | 4/4 | 10.33 s | 184.75 | $0.0447 |
| Sol low | 4/4 | 7.96 s | 175.50 | $0.0086 |
| Sol medium | 4/4 | 10.28 s | 266.25 | $0.0095 |
| Luna high | 4/4 | 11.52 s | 371.25 | $0.0006 |
Time is measured from CLI launch to exit. The API-equivalent cost is not what we were billed (the runs were on a subscription; it is a reference figure applying the official API rates to the token counts the CLI reported).
From 5:47 to 5:49 the same day, we ran one more, separate task in Japanese — spotting date-conversion and counting errors across several sources — with 12 grading items fixed in advance, twice per setting. All 8 runs passed, with median times of Sol low 9.65 s, Sol medium 13.29 s, Astra low 13.48 s and Luna high 17.32 s (the task and time window differ from the table above, so we did not mix the averages).
All this shows is that for short tasks with checkable answers, there are cases where Sol low or Luna high was enough. Every setting passed because the tasks were easy, and that is no evidence the models are equally capable. There were 3 independent tasks, with 4 and 2 runs per setting. We did not measure fixes to an existing repository, design, long-form writing or long autonomous work. And because the remaining-usage indicator shows whole percentages and is shared with other work, this test can't tell you how much of the usage limit each model consumed.
Two things were still observable. Sol low finished faster than Astra low at about 1/5 of the API-equivalent cost. Luna high, meanwhile, had the lowest API-equivalent cost but took longer than Astra low — the same pattern as in AA's table.
8. How to choose for your own work — where the decision turns
From what we've covered (the official use-case guidance, AA's evals and our small test), the following split is a practical one. Read it as a proposed starting point, not a proven optimum.
The official starting point. On AA's business-workflow eval it is 1 point behind Astra low at about a third of the cost. If you're unhappy with the results, step up to Astra.
Narrow requirements where a test or format check tells you pass or fail right away. Luna high for extraction, classification and transformation. Luna isn't fast, though, so if waiting bothers you, use Sol low.
Overall architecture, bugs with unknown causes, reconciling multiple sources, pre-release reviews, terminal-centric development. Start at low, and if you see misses or rework, compare against medium and above.
The deciding question is whether you can check the result of the job yourself right away. If passing tests or a correct format settles it, a failure on a cheap model is noticed immediately and redoing it costs little. Conversely, for work where mistakes surface late (design decisions, factual writing without sources, visual quality), rework from a cheap model ends up costing more than the price difference.
If you've been sticking with Astra low: there's no need to stop. If you want to try, a safe approach is to switch only your new everyday tasks to Sol medium for a week and compare how often you have to redo work and how fast your usage limit goes down. Keep tasks that are midway through important work on Astra, so a failed comparison doesn't affect real work. Note that changing the default in the config file (model and model_reasoning_effort in config.toml) only affects newly created tasks. For tasks you've already created, check the model and effort under each one's input box.
The most reliable way to decide is to compare on your own representative work. From the same starting state, did the first output pass, how many times did you fix it, and how many minutes did your own final check take? Count those three and you'll get a better answer for you than any table.
FAQ
Q1. Can I use GPT-6 Sol and Luna yet?
From September 22, 2026 (US time), they have been rolling out in ChatGPT Work and Codex on Plus, Pro, Business, Enterprise and Edu. Free and Go get Luna only, in the desktop app. In the API they are available as gpt-6-sol and gpt-6-luna. They are not offered in regular Chat (per the official pages as of September 24, 2026).
Q2. I use Astra low. Should I switch to Sol?
There's no reason you have to. OpenAI's guidance is Sol for everyday work and complex coding (starting at Medium) and Astra for the hardest work (starting at Light). In Artificial Analysis's evals, Sol medium was 1 point behind Astra low on business workflows, but far behind on terminal work, at 19% vs 42%. We recommend trying Sol medium on new everyday tasks only and comparing how often you have to redo work.
Q3. If Sol costs 1/5 of Astra, do I get 5 times as much use?
You can't say that. API rates and credits are 1/5, but Codex's own pricing page says credit rates alone don't determine a subscription's usage. The same page's rough ranges per 5 hours on Plus are Astra 5–45 and Sol 15–150 messages, only about 3 times as many. They are also not fixed limits, and consumption varies with the work.
Q4. Luna is the cheapest. Can I use it for everything?
It suits repetitive work where you know what a good result looks like (extraction, classification, transformation, summaries in a fixed format). In Artificial Analysis's evals, Luna high costs about 1/27 of Astra low per task, but it takes longer, scores just 5% on terminal work, and drops sharply on the knowledge-accuracy metric. Avoid handing it factual questions without supplied sources, or long autonomous development, on price alone.
Q5. The app selects "Sol Light" by default, but OpenAI says "start Sol at Medium". Isn't that a contradiction?
No. Sol Light is the default of the Power presets in the desktop app and on the web, while "Sol at Medium" is the starting point for getting results on that kind of work. The API default is something else again: medium for both Sol and Luna. A simple way to think about it: keep the default for light work, and go to Medium or above for work that needs planning or analysis.
Q6. What happens to GPT-5.6 Sol and GPT-5.5?
GPT-5.6 Sol, Terra and Luna remain available during the transition period. GPT-5.5 retires from ChatGPT, ChatGPT Work and Codex on October 14, 2026 (the API is not affected), and OpenAI names GPT-6 Sol as the replacement on paid plans (Plus, Pro, Business, Enterprise, Edu) and GPT-6 Luna in the desktop app on Free and Go. If you specify gpt-5.5 in a config file or a scheduled task, change it before then.
Sources
The figures in this article were checked on the publishers' pages below on September 24, 2026, together with AI Arte's own measurement data.
- OpenAI — Introducing GPT-6 Sol and Luna
Positioning, price cuts from the previous generation (Sol $4→$2 and $20→$10, Luna $0.20→$0.10 and $1.20→$0.50), AutomationBench 1.0.6 (Sol xhigh 33.2% at $0.27, Astra low 30.3% at 3.9×, Opus 5 max 26.9% at 11.1×), DeepSWE v1.1 (Sol max 68.8%, Luna max 66.6%), Agents' Last Exam (Sol max 56.4%), OSWorld 2.0 (Sol xhigh 60.5%), the internal factuality eval and its caveats, availability - OpenAI Developer Community — GPT-6 Sol and Luna announcement
Release time (September 22, 2026, 18:16 UTC), availability in Work, Codex and the API, desktop Luna for Free and Go - OpenAI — GPT-6 Sol, GPT-6 Luna, GPT-6 Astra (API documentation)
Model IDs, context, max input and max output, knowledge cutoffs, reasoning effort and defaults, API pricing, the surcharge above 272K, 1.25× cache writes, 2× Fast mode - ChatGPT — Models
Use-case guidance (Sol at Medium, Luna at High, Astra from Light), the six Power presets with Sol Light as default, not available in Chat, GPT-5.6 available during the transition, GPT-5.5 retirement (October 14, 2026) - ChatGPT / Codex — Pricing
Availability by plan, credit rates, the note that rates alone don't determine usage, rough ranges per 5 hours (Plus: Astra 5–45, Sol 15–150, Luna 350–3,000), 2.5× Fast mode - Artificial Analysis — comparison pages (Sol medium vs Astra low, Sol low vs Astra low, Luna high vs Astra low)
Intelligence Index v4.3.2, Terminal-Bench 4.0, AutomationBench-AA, AA-Omniscience, cost, output tokens and time per task - AI Arte's measurements (September 24, 2026, Codex CLI 0.155.0-alpha.16)
4 settings × 2 tasks × 2 runs = 16 runs, plus an additional Japanese-language task run twice per setting = 8 runs. Passes, time, output tokens, API-equivalent cost
Related Articles
- GPT-6 Astra release: the complete guide — pricing, availability by plan and benchmarks for the top model
- Is GPT-6 Astra low really cheaper? — a measured comparison with the previous-generation GPT-5.6 Sol at high
- GPT-5.6 release: the complete guide — the previous generation's Luna / Terra / Sol
- Claude Opus 5.5 release — Anthropic's new model from the same week
- Knowledge cutoff dates of major AI models — how far each model's knowledge reaches
