GPT-6 Luna, Sol, or Astra in ChatGPT Work: Where to Start Your Next Task

Last updated: September 27, 2026

GPT-6 Luna, Sol and Astra are OpenAI’s GPT-6 models. In ChatGPT you choose Sol and Luna in Work, the agent for multi-step tasks and finished files, or in Codex; regular Chat does not offer them. Astra is also in Work and Codex, and Chat offers it on eligible plans as “GPT-6 Pro.” OpenAI introduced Astra on September 3, 2026, and Sol and Luna on September 22. What you see depends on your plan, workspace and rollout.

Best for, not for, and when to wait

At a glance

01 What you’ll see in Work, and what it isn’t

Open Work, not Chat: OpenAI calls Work “an agent designed for longer, multi-step work and finished deliverables.” OpenAI’s steps: on the web or mobile, select Work; in the desktop app, choose ChatGPT from the top-left menu, then Work from the toggle at the top. On the web and desktop, OpenAI places the model and reasoning control beneath the composer, the box where you type.

The picker offers Default, one slider that combines models and reasoning levels, or a specific model with its own levels. OpenAI’s Enterprise and Edu page says the slider “shows only models and reasoning options available to you,” and its Models page says choosing a model does not grant access to it. If Astra is missing in the desktop app, OpenAI says to use Menu > Check for Updates, even if you updated recently.

Free and Go need a separate check. Work on the web and mobile began rolling out “to paid plans except Free and Go,” and OpenAI’s help page still limits it to eligible paid plans. The Pricing page names Free and Go among plans that include Work and Codex and lists GPT-6 Luna “in the desktop app, subject to rollout” for them, but on cards that describe Codex use. FSR found no page that settles whether Free or Go get Work in the desktop app, so check the app before you plan around it.

In Enterprise and Edu workspaces, all three GPT-6 models are off at launch. An admin must enable Sol and Luna. Workspace owners can enable Astra for the workspace or for specific roles, and access “will not automatically turn on two weeks after launch.” Admins can also set a starting model, reasoning level and speed for Work and Codex, which OpenAI calls “defaults, not permissions.”

02 Start from the default your account shows

OpenAI’s instruction is short: “Start with the default Power setting available to your account.” From there, move toward Smarter for deeper reasoning or Faster for quicker, lower-cost work, or open Advanced to pick a model, effort or speed. OpenAI’s Models page shows six presets: Luna High, Sol Light (the starting one), Sol Medium, Astra Light, Astra Medium and Astra Extra High; some paid plans omit the last.

OpenAI’s pages call this slider both Default and Power, and the Models page says options and defaults “vary by plan, client, workspace settings, and rollout.” Treat Sol Light as OpenAI’s example, not a promise about your picker; in Enterprise and Edu, your admins may have chosen the default.

If your default is still GPT-5.5, it works until it retires on October 14, 2026, but OpenAI asks for it to be replaced in saved settings and workspace defaults before then. Pick another setting your picker shows, or ask your admin.

Each setting pairs a model with a reasoning effort. Higher effort can improve results on complex tasks, OpenAI says, but takes longer and uses more tokens. If you pick a model yourself, OpenAI gives each a starting effort:

03 Use OpenAI’s guide only for a clear fit

OpenAI’s Model selection page has an interactive guide: pick a type of work and a task (together, a category), then a priority on a slider from More usage to Higher quality outputs, and it recommends a model and effort. OpenAI says to “treat the guidance on this page as a starting point,” and that if usage and time don’t matter to you, “you can default to Astra.” The guide asks nothing about your plan, remaining usage or admin settings, and its slider starts at Balanced.

Take an ordinary job: turning a week of meeting notes into a one-page decision memo. Filed under Research & analysis › Organize information and prioritize work, the guide suggests Luna Medium, or Sol Medium at Higher quality outputs. Filed under Creating deliverables › Turn source material into a final output, it suggests Astra Medium at every position. Both descriptions fit, and the guide does not say which to pick.

FSR’s rule: start on the guide’s recommended setting only if your task clearly fits just one category and your picker offers that setting. If it fits two, as this memo does, stay on your default. How you will use the output still matters, but in FSR’s reading it sets the slider, not the category. OpenAI lists “How will you use the output?” among four questions to start with, and notes that a draft for your review may need less polish than something you’ll share externally. Whichever setting you start on, the check in section 04 decides.

OpenAI’s pages also draw the lines differently: the Models page puts ambiguous tasks and polished documents under Sol, while the Get started page suggests Astra for a polished final file and the model-selection page’s Astra card lists “Ambiguous problems.” If the guide points to Astra and you are on Plus or Business Standard, check your usage first; those seats include limited Astra usage (section 05).

04 Run one check, then change what failed

Before real material goes in, confirm you may use it. OpenAI says Work inherits applicable workspace protections, but residency, retention, logging and features vary by plan, region, surface and connected system; “messages and task context may be stored in the cloud” even when work runs locally. On personal plans, whether your content helps improve OpenAI’s models depends on Settings > Data controls; OpenAI says it does not use Business, Enterprise or Edu content for that by default.

You can choose where to start from this article alone, but check the first result before you rely on it. FSR suggests this check, built on OpenAI’s advice to give a task “any files, constraints, and review criteria” and to compare settings on the same input.

1. Use your next task, or a similar one you repeat, with material you are allowed to use: public, made-up or cleared.
2. Write your pass criteria first and mark the must-haves. For the memo: every decision traces to a line in the notes; no invented names, dates or numbers; no missing decisions or owners; one page.
3. Run it once on your starting setting. To compare, run one other setting on the same input.
4. Score it against your criteria, not how finished it looks. If it meets every must-have, keep that setting; to save time or usage, try a lighter one on the same input and keep it only if it also passes.
5. If a must-have fails, don’t use that output. Note what failed, change one thing using the table below, and run again.

05 How usage works, and what the estimates leave out

OpenAI says Work uses the same pricing, credits and usage limits as Codex, and that its Codex pricing examples are based on coding tasks, “so Work usage will vary by task.” The table counts local messages, meaning chats that run on your computer: in the desktop app, the CLI or the IDE extension.

OpenAI says these figures “are not fixed message limits”; your usage dashboard shows current limits and reset times. Similar tasks can use different amounts, since model choice, context, reasoning, tool use, retrieval and caching all count. Enterprise and Edu plans without flexible pricing get Plus-level per-seat limits for most features; flexible pricing has no fixed rate limits.

Work on the web and mobile runs in the cloud, which the table may not cover. The Pricing page says cloud chats on ChatGPT plans “use GPT-5.6 Sol and may use more of your allowance than local messages,” but OpenAI’s glossary defines a cloud chat as a Codex chat run in the cloud. FSR found no estimate specific to Work on the web.

Astra draws on the same allowance. Pro $100, Pro $200 and Business Premium seats can use all of it for Astra; Plus and Business Standard seats include limited Astra usage, with optional credits after that. Astra “can use your allowance faster than GPT-5.6 Sol,” OpenAI says.

If you reach a limit mid-turn, OpenAI says the agent can keep working on that turn, subject to fair use limits. After that, Plus and Pro users can buy credits, and Business, Enterprise and Edu workspaces on flexible pricing can buy workspace credits; this article does not recommend paying to run a first check. If you want more Astra usage, note that OpenAI has temporarily paused new sign-ups and upgrades to Pro $200 since September 10, 2026, though some eligible former subscribers can return once; Pro $100 is not affected.

06 When to stop, wait, or step back

FSR suggests stopping or waiting when:

• You cannot confirm you may put the material into Work, or cannot check the output against a source.
• A must-have still fails after you fix the cause the check pointed to. Narrow the task or do it by hand.
• An Astra task paused and you cannot tell whether continuing is safe. OpenAI says to leave it stopped.
• GPT-6 models are not in your picker yet, or your Enterprise or Edu admins have not enabled them. Wait, or keep your current setup.

Astra’s safety monitoring in Work and Codex runs asynchronously and can pause a task, sometimes after the activity that triggered it; OpenAI says it “doesn’t replace sandboxing, permissions, or review of the result.” A pause does not necessarily mean the agent made a mistake. OpenAI’s steps: read the findings, compare them with what you intended and what the agent did, and resume only if the task can safely continue. The pages FSR checked describe this for Astra only.

Stepping back to GPT-5.6 is possible during the rollout on plans that list it in Work, with a trade-off: OpenAI’s estimates give GPT-5.6 Sol and Luna fewer local messages than their GPT-6 namesakes in every column in section 05. Efforts don’t map exactly between generations, so compare the two on the same input. GPT-5.5 works only until it retires on October 14, 2026, so don’t build a routine on it.

07 FAQ

08 Methodology and limits

Starter Guide, Tier C. FSR did not examine a signed-in Work screen, run a Work task, test any GPT-6 model or buy anything, so nothing here measures quality, speed or usage on real work. The section 03 rule, the check in section 04 and the table after it are FSR’s proposals.

FSR read the 18 OpenAI pages linked in this article, from ChatGPT Learn and the OpenAI Help Center, on September 26 and 27, 2026 (Asia/Tokyo), and last checked all of them on September 27. They were read in English from Japan; FSR did not check whether they differ by region. FSR operated the model-selection guide on OpenAI’s public docs page, and FSR’s editor also viewed ten of the pages in a browser. Where this article says a page does not settle a question, that covers only these pages.

FSR researched and drafted this article with AI assistants, including Claude and ChatGPT, and placed sources beside the claims they support; agreement between AI tools is not independent verification. This page has no affiliate or paid links.

09 Verdict

The case rests on OpenAI’s own advice to start from the default and to treat its guide as a starting point; FSR has not measured any GPT-6 model. The table below maps common situations to a next step.

10 Corrections and contact

FSR BRIEFING PACKET - GPT-6 Luna, Sol, or Astra in ChatGPT Work: Where to Start Your Next Task
Publisher: Future Stack Reviews. Source: https://future-stack-reviews.com/gpt-6-chatgpt-work-starter-tierc/
Format: Starter Guide. Evidence depth: Tier C. OpenAI's public documentation only. FSR did not examine a signed-in Work screen or test any GPT-6 model.
Read: 26 and 27 SEP 2026 Asia/Tokyo. Section numbers refer to that article.
Disclosure: no affiliate or paid links. FSR used AI assistants, including Claude and ChatGPT, to research and draft the article.

WHAT IT IS
OpenAI's GPT-6 Luna, Sol and Astra. In ChatGPT, Sol and Luna are in Work and Codex only, not regular Chat. Astra is in Work and Codex, and Chat offers it on eligible plans as GPT-6 Pro. OpenAI introduced Astra on Sep 3, 2026, and Sol and Luna on Sep 22. Access depends on plan, workspace settings and rollout.

FROM OPENAI'S PAGES (not FSR tests)
- "Start with the default Power setting available to your account." Documented presets: Luna High, Sol Light (starting preset), Sol Medium, Astra Light, Astra Medium, Astra Extra High. Options and defaults vary by plan, client, workspace settings and rollout.
- OpenAI's positioning: Luna for clear, repeatable, high-volume tasks; Sol for everyday and complex work; Astra for the hardest end-to-end work. If you pick a model yourself, starting efforts: Sol Medium, Luna High, Astra Light.
- The model-selection guide is to be treated as a starting point.
- Local messages per five hours, meaning chats that run on your computer (estimates, not limits; Work usage varies by task), Plus and Business Standard / Pro 5x ($100) / Pro 20x ($200): Astra 5-45 / 25-225 / 100-900. Sol 15-150 / 70-700 / 300-3,000. Luna 350-3,000 / 1,750-14,000 / 7,000-56,000. Business Premium seats get 5x Standard's included usage and no 5-hour limit, so there is no Premium figure. Work on the web and mobile runs in the cloud; FSR found no estimate for it. Astra is limited on Plus and Business Standard and can use the allowance faster than GPT-5.6 Sol.
- GPT-5.6 Sol, Terra and Luna remain during the rollout. GPT-5.5 works until it retires from ChatGPT, Work and Codex on Oct 14, 2026; OpenAI asks for it to be replaced in saved settings and workspace defaults before then.

FSR OBSERVATION AND RULE (section 03)
- On OpenAI's public docs page, the guide asks only for work type, task type and priority, and its slider starts at Balanced.
- FSR example: a meeting-notes-to-memo task gets Luna Medium under Research & analysis > Organize information and prioritize work (Balanced), or Astra Medium under Creating deliverables > Turn source material into a final output (every position).
- FSR's rule: start on the guide's recommended setting only if a task clearly fits just one category (work type > task) and your picker offers that setting; if it fits two, stay on the default. In FSR's reading, how the output will be used sets the quality bar on the slider, not the category.

FSR'S SUGGESTED CHECK (section 04; not a test FSR ran)
- Write must-have criteria first and run once on the starting setting. If it passes, keep it; to save time or usage, try a lighter setting on the same input and keep it only if it also passes.
- If a must-have fails, do not use that output. Facts outside your material: fix the files and the brief. Shallow analysis: one step more effort, then a more capable model in your picker. If it still fails, narrow the task or do it by hand.

WHAT DEPENDS ON YOUR SETUP
- Free and Go: web and mobile Work launched for paid plans except Free and Go. The pages FSR checked do not settle whether Free and Go get Work in the desktop app.
- Enterprise and Edu: all three models off by default at launch; admins enable them and may set a starting default, which is not access.
- Astra tasks can pause as a safety precaution. On mobile, in Codex CLI, and with zero data retention, Modified Abuse Monitoring or non-US data storage residency, the task ends instead of resuming.
- Before real material: Work inherits workspace protections, but coverage varies; local chats may store messages and task context in the cloud; on personal plans, model-improvement use depends on Settings > Data controls.

OPEN QUESTIONS - check your own account or ask your admin
1. Which models and presets your Work picker shows, and what your default is.
2. On Enterprise or Edu: which GPT-6 models your admins enabled, and any starting default they set.
3. How much Astra usage your plan or seat includes, and what is left.
4. Whether you may put your material into Work.
5. On Free or Go: whether your desktop app offers Work at all.

FSR VERDICT
Start your next Work task on the default your account shows, unless your task clearly fits just one category in OpenAI's model-selection guide and your picker offers the setting it recommends; then start there. Judge the first result against criteria you wrote in advance, and change the input, effort or model based on what failed.

PRIMARY SOURCES (OpenAI; recheck before you act)
- Models: https://learn.chatgpt.com/docs/models
- Model selection: https://learn.chatgpt.com/docs/model-selection
- Pricing: https://learn.chatgpt.com/docs/pricing
- ChatGPT Work and Codex: https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex
- ChatGPT Business models and limits: https://help.openai.com/en/articles/12003714-chatgpt-business-models-and-limits
- Enterprise and Edu models and limits: https://help.openai.com/en/articles/11165333-chatgpt-enterprise-and-edu-models-limits
- Paused or ended as a precaution: https://help.openai.com/en/articles/20001509-why-was-my-chat-paused-or-ended-as-a-precaution

LIMITS OF THIS PACKET
Rollouts, defaults and estimates change; verify on OpenAI's pages before you act. FSR ran no tests, so nothing here measures quality, speed or usage on your work. Where this packet and the live pages disagree, the live pages win.

Use this briefing to help me decide which GPT-6 model and effort to start my next ChatGPT Work task with. Ask me about my plan, what my model picker shows, the task and how I would check the output before you answer.

Starter Guide. Read September 26 and 27, 2026 (Asia/Tokyo). Tier C: no signed-in Work screens, no tests, no purchase. No affiliate or paid links. Rollouts, defaults and estimates change; recheck OpenAI’s Models, Model selection and Pricing pages before relying on them.