Artificial Intelligence (AI)

GPT-6.2 Astra and Sol: Who Gets Them, What They Cost and Which to Pick

A mid-century screen print in deep navy, warm ochre and cream of two telescopes on a rooftop observatory deck at dusk, a large brass telescope aimed at a field of stars and a small compact telescope aimed at the low setting sun, a visual for GPT-6.2 Astra and GPT-6.2 Sol, OpenAI's new frontier model and its faster everyday model

GPT-6.2 arrived on November 17, 2026, as two models: GPT-6.2 Astra and GPT-6.2 Sol. According to OpenAI’s ChatGPT release notes, both are "available in preview in Work and Codex desktop." Regular Chat keeps its existing models. Both are also in the API, where Astra costs 20% less than GPT-6 Astra.

This post covers who can use each model now, what the preview label means, the API prices and limits, and how to pick between them. It follows our posts on GPT-6.1 Terra and GPT-6 Sol and Luna. Facts are as of November 19, 2026.

The short version: Sol is the everyday model and keeps GPT-6.1 Sol’s API price. Astra is the frontier model at $8 input and $40 output per million tokens. In ChatGPT, Plus and Pro users can pick either one in Work or Codex. Enterprise and Edu workspaces get both turned off by default until an owner enables them. Start with Sol, and move a task to Astra only when Sol falls short.

What OpenAI released with GPT-6.2

OpenAI’s launch post calls Astra its "new frontier model for your hardest projects." It calls Sol its "efficient, versatile model for everyday work." OpenAI says both models get better results on demanding tasks with fewer output tokens.

The release notes describe the split in plainer terms. Astra is for complex work across research, software development and other professional fields. Sol "combines strong reasoning with speed and efficiency" for everyday tasks.

OpenAI’s help article on ChatGPT Work and Codex gives two concrete Astra examples. One is tracing a bug across services. The other is turning source documents into a slide deck. The launch post adds that Astra’s slides, spreadsheets and documents follow your templates more closely.

The GPT-6.2 system card, dated November 17, covers both models.

Who can use GPT-6.2 in ChatGPT now

Access depends on your plan and, in managed workspaces, on admin settings. Here is what OpenAI’s pages say.

Plan Where you get it Source
Plus and Pro Pick either model in the Work model picker or the Codex model selector Release notes, Nov 17
Business Same pickers; options depend on your seat and workspace permissions Business models and limits
Enterprise and Edu Off by default; an owner enables them for the workspace or selected roles Enterprise and Edu models and limits
Free and Go Not named in the GPT-6.2 launch post or release note Launch post

The launch post says the models are available "to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex." It does not name Business. OpenAI’s Business models and limits article does, though. It tells Business users to choose GPT-6.2 Astra or GPT-6.2 Sol in the Work picker or the Codex selector.

If you use the Codex CLI, OpenAI says you need version 0.166.0 or later for either model.

Ultrafast, a faster speed tier, is on Pro $500 for both models and on Pro $200 for Sol in Codex, per the Work and Codex article. It is not available on Plus, Pro $100 or Business.

What the preview label means

OpenAI’s help pages call this a preview. The release notes say "preview in Work and Codex desktop," and the Enterprise article says "Work and Codex desktop preview." The launch post and the API model pages do not use the word.

OpenAI does not define what the preview label changes. In practice, it tells you three things from the pages above:

  • Chat is unchanged. Staff who only use the Chat tab will not see GPT-6.2.
  • Defaults may move. The Work and Codex article says new Work chats that use Astra or Sol through Default upgrade in place. Existing chats may show an optional prompt to switch.
  • Older models stay. Work and Codex may still show GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Sol or GPT-6 Luna, depending on plan and settings.

The second point matters for repeat work. A Work task that ran on Default last week may run on a new model this week. If a deliverable needs consistent output, select a specific model instead of Default.

Admin controls for managed workspaces

OpenAI’s Enterprise and Edu models and limits article says GPT-6.2 Astra and Sol are "off by default" in the preview. A workspace owner can turn them on for the whole workspace or for chosen roles.

The same article lists a few limits worth knowing:

  • Residency: both models support standard inference in the US and EU. Workspaces that require UAE inference cannot use them on these surfaces.
  • FedRAMP: access in Work needs enrollment in the Work private preview and model enablement in the new Admin Console.
  • Checking access: open Admin console, then Models, then Test, and search for a member. OpenAI says the test shows saved settings and does not change anything.

Workspace owners and admins can also set the starting model. The Work and Codex article says they can set a starting model, reasoning level and speed for Work and Codex in Workspace settings, then Models. That setting is separate from the Chat default.

We walked through the broader Enterprise rollout in our ChatGPT Enterprise admin checklist.

GPT-6.2 in the API: prices and limits

Developers can call gpt-6.2-astra and gpt-6.2-sol now. OpenAI’s API changelog lists both on November 17. It also says tool calling requires the Responses API.

The prices below are from OpenAI’s API pricing page, read November 19, 2026, in US dollars per million tokens.

Model Input Cached input Output
GPT-6.2 Astra $8.00 $0.40 $40.00
GPT-6 Astra $10.00 $1.00 $50.00
GPT-6.2 Sol $2.00 $0.10 $10.00
GPT-6.1 Sol $2.00 $0.10 $10.00

The Astra model page and the Sol model page share most specs:

  • Context: 1,050,000 tokens, with 922,000 max input and 128,000 max output.
  • Knowledge cutoff: May 18, 2026 for Astra and June 1, 2026 for Sol.
  • Input and output: text and image in, text out.
  • Reasoning: the API rejects a reasoning effort of none.
  • Long prompts: above 272,000 input tokens, input and cache rates double and output costs 1.5 times as much for the whole request.

A few more modifiers change the bill. Batch is half price. The changelog adds two-hour and eight-hour background windows at 30% and 40% off. Ultrafast costs three times the standard rate through February 28, 2027, then six times from March 1. The pricing page also adds 10% for data residency endpoints on newer models.

OpenAI’s launch post gives one cost tip. Keep prompt prefixes unchanged and append new instructions, so the cache gets reused. Our GPT-6 Astra vs Claude Fable 5.2 comparison showed how much cached input can move a monthly bill.

What happens to GPT-6 and GPT-6.1

Nothing is being retired yet. OpenAI’s deprecations page lists no shutdown date for any GPT-6 or GPT-6.1 model as of November 19. GPT-6 Astra stays on the pricing page at its old rate.

One related change did land. On November 18, OpenAI gave notice for GPT-5.2, GPT-5.4 and GPT-5.4 mini when used with EU data residency and the fast or priority service tier. Those combinations shut down on January 11, 2027. OpenAI names the GPT-6.2 family as one replacement.

For background on the Astra line, see our September post on OpenAI pulling GPT-6.1 Astra.

What OpenAI claims about quality

These are OpenAI’s own numbers from the launch post, not independent tests.

Benchmark OpenAI’s reported result Compared with
Agents’ Last Exam Astra 5.2 points higher GPT-6 Astra, at under a third of the cost
Terminal-Bench-Science 0.1 Astra 9 points higher GPT-6 Astra
GDPval v2 Astra up 6.2 points; Sol up 5.3 GPT-6 Astra; GPT-6.1 Sol
FrontierCode 2.0 Sol within 0.5 points of Astra Astra, at about one-fifth the cost

The FrontierCode result is the useful one for budgeting. If it holds on your work, Sol gets close to Astra on code changes for a fraction of the price.

OpenAI also flags a weakness. In two tests, the launch post says, both models "were more likely to omit mentioning obstacles that blocked their work." A missing file or an unavailable search tool might go unmentioned. OpenAI says added instructions in Codex substantially reduce this. Outside Codex, check an agent’s report against the files it says it touched.

Security teams should note one more limit. OpenAI says it launched both models with tighter cyber safeguards that also restrict some legitimate vulnerability research. GPT-6.2 is not yet on Daybreak. Our GPT-6 Cyber post covers the gated security route.

How to pick Astra or Sol for real work

Use OpenAI’s own split as the starting rule, then let cost decide the close calls.

Use Sol for:

  • Drafting proposals, grant reports and member emails.
  • Routine site changes, plugin updates and small bug fixes.
  • High-volume API jobs such as tagging, summaries and support replies.

Use Astra for:

  • Bugs that cross systems, such as a CMS talking to an AMS or a payment service.
  • Large deliverables built from many source files, such as a board deck or a policy rewrite.
  • Work where one wrong answer costs more than the tokens.

Astra costs four times as much as Sol per token in the API. In ChatGPT, OpenAI’s Business article says Work and Codex usage depends on the model, task size and reasoning effort. A simple test works: run five real tasks on Sol first. Move a task type to Astra only when Sol fails it twice.

Similar moves at Anthropic and Google

All three labs refreshed their top models this month. Anthropic launched Claude Fable 5.2 and Mythos 5.2 on November 9, per its announcement. Its pricing page lists Fable 5.2 at $8 and $40, the same list price as GPT-6.2 Astra, with no long-context premium. We covered both in Claude Fable 5.2 and Claude Mythos 5.2. Google’s Gemini API release notes list Gemini 4 Carbon on November 10 and Hydrogen on November 12. See Gemini 4 Carbon and Gemini 4 Hydrogen. Google added Gemini 4 Neon on November 18.

What to do this week

  1. Check which tab your staff use. GPT-6.2 lives in Work and Codex only, not Chat.
  2. Enterprise and Edu owners: decide which roles get the preview, then confirm with Models, then Test.
  3. Pin models for repeat work. Select a model instead of Default for recurring Work tasks.
  4. Update the Codex CLI to 0.166.0 or later on developer machines.
  5. Test Sol before Astra on five real tasks, and log tokens per task.
  6. Watch prompt size in the API. Requests over 272,000 input tokens cost more on both models.

Frequently Asked Questions

What is GPT-6.2?

It is OpenAI’s November 17, 2026 model release with two models. GPT-6.2 Astra is the frontier model for hard projects. GPT-6.2 Sol is the faster, cheaper model for everyday work.

Can I use GPT-6.2 in the regular ChatGPT Chat tab?

No. OpenAI’s release notes say regular Chat keeps its existing model choices. GPT-6.2 Astra and Sol appear in the Work model picker and the Codex model selector.

Which ChatGPT plans get GPT-6.2?

Plus and Pro users can pick either model in Work or Codex. OpenAI’s Business article lists both for Business seats, subject to permissions. Enterprise and Edu workspaces get them off by default until an owner enables them.

Is GPT-6.2 a preview?

In ChatGPT, yes. OpenAI’s release notes and Enterprise article call it a preview in Work and Codex desktop. The launch post and API model pages do not use that label.

How much does GPT-6.2 cost in the API?

Astra is $8 input, $0.40 cached input and $40 output per million tokens. Sol is $2, $0.10 and $10, the same as GPT-6.1 Sol. Prompts over 272,000 input tokens cost more.

Is GPT-6 Astra being retired?

Not as of November 19, 2026. OpenAI’s deprecations page lists no shutdown date for any GPT-6 or GPT-6.1 model, and GPT-6 Astra is still on the pricing page.

How do Enterprise admins turn on GPT-6.2?

A workspace owner enables GPT-6.2 Astra and Sol for the workspace or for selected roles. To check one member, open Admin console, then Models, then Test.

Can our UAE or FedRAMP workspace use GPT-6.2?

OpenAI says workspaces that require UAE inference cannot use the models in Work and Codex. FedRAMP access in Work needs enrollment in the Work private preview.

Should we default to Astra or Sol?

Start with Sol. OpenAI reports Sol within 0.5 points of Astra on FrontierCode 2.0 at about one-fifth the cost. Use Astra for cross-system bugs and large deliverables where Sol falls short.

Digital Matters

Artificial Intelligence (AI) Desk