OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026, eight days after Anthropic shipped Claude Fable 5.2. Astra is the new frontier model for hard projects and costs $8 per million input tokens and $40 per million output tokens in the API, 20% less than GPT-6 Astra. Sol is the cheaper everyday model and keeps GPT-6.1 Sol's price of $2 and $10.
The catch is where you can use GPT-6.2. Both models are live in ChatGPT Work and Codex for paying subscribers and in the API, but OpenAI says the rollout "does not add either model to regular Chat". This article covers which plans get GPT-6.2, what it costs, and what OpenAI's benchmarks claim. It also covers what independent testing says, and how GPT-6.2 compares with Claude Fable 5.2, which sells at exactly the same token price.
The Key Takeaways
- Released: GPT-6.2 Astra and GPT-6.2 Sol launched on 17 November 2026 in ChatGPT Work, Codex and the API.
- Not in Chat: Plus, Pro, Business, Enterprise and Edu get both models in Work and Codex only. Free and Go get neither.
- Price: Astra costs $8 / $40 per million tokens, the same sticker as Claude Fable 5.2. Sol stays at $2 / $10.
- Independent score: Artificial Analysis puts Astra at 62 and Sol at 57, behind Fable 5.2's 65, but Astra costs $4.00 per index task to Fable's $26.54.
- Safety: Astra is OpenAI's first public model rated High for AI self-improvement, and both models still sometimes fail to mention what blocked their work.
What GPT-6.2 Is: Astra and Sol
GPT-6.2 is a pair of models, not one. GPT-6.2 Astra is what OpenAI calls its "new frontier model for your hardest projects", and GPT-6.2 Sol is its "efficient, versatile model for everyday work". Astra replaces GPT-6 Astra at the top of the line. Sol replaces GPT-6.1 Sol, which launched at DevDay in September.
OpenAI's pitch for both is efficiency. According to OpenAI, both models "deliver better results on demanding tasks with fewer output tokens". More work fits inside a ChatGPT plan's usage limit, and each API task costs less. There is no GPT-6.1 Astra in between. OpenAI cancelled that release in September after it failed internal safety tests, and GPT-6.2 Astra is the first Astra since the GPT-6 Astra launch on 3 September.
OpenAI's own account posted the launch at 17:54 UTC:
Some of you have been asking very politely. Some of you have just been asking.
— OpenAI (@OpenAI) November 17, 2026
GPT-6.2 Astra and Sol are now available.
Super efficient and built to perform where it counts—the real world.
https://openai.com/index/introducing-gpt-6-2-astra-sol
GPT-6.2 Astra vs GPT-6.2 Sol
Pick GPT-6.2 Astra for long, high-stakes work and GPT-6.2 Sol for everything else. OpenAI's ChatGPT documentation gives Astra the heavy jobs. Its examples are tracing a bug across several services, building a slide deck from a report in your team's template, and finding where research sources disagree. Sol gets reviewing a pull request, revising code after feedback, and everyday writing.
Both models read text and images and write text. Both have a 1,050,000-token context window, with up to 922,000 tokens of input and 128,000 of output. Astra's knowledge cutoff is 18 May 2026 and Sol's is 1 June 2026, according to OpenAI's API model pages.
Can You Use GPT-6.2 in ChatGPT?
Yes, but only in Work and Codex, and only on a paid plan. OpenAI's release notes say: "You can choose Astra or GPT-6.2 Sol in the Work model picker." The same entry says the rollout "does not add either model to regular Chat". On paid plans, regular Chat keeps running GPT-6 Sol for the Instant to Extra High thinking levels, with GPT-6 Pro, powered by GPT-6 Astra, on the Pro setting.
OpenAI's launch post lists Plus, Pro, Enterprise and Edu. OpenAI's help centre and its ChatGPT documentation also include Business, so Business seats get GPT-6.2 too. Enterprise and Edu workspace owners have to switch the models on first. ChatGPT Work is the agent side of ChatGPT that builds files and runs longer tasks, and it shares one usage allowance with Codex.
| ChatGPT plan | GPT-6.2 Astra | GPT-6.2 Sol | Ultrafast |
|---|---|---|---|
| Free, Go | No | No | No |
| Plus ($20) | Work and Codex | Work and Codex | No |
| Pro $100 | Work and Codex | Work and Codex | No |
| Pro $200 | Work and Codex | Work and Codex | Sol only, in Codex |
| Pro $500 | Work and Codex | Work and Codex | Both models |
| Business | Work and Codex | Work and Codex | No |
| Enterprise, Edu | Once an admin enables it | Once an admin enables it | Both, if the owner allows it |
New Work chats on consumer plans start on GPT-6.2 Sol at Light reasoning effort, according to OpenAI's ChatGPT documentation. To get Astra, open the model picker in Work and choose it. If you use the ChatGPT desktop app or Codex CLI, OpenAI says to update both before the models appear. Our guide to every ChatGPT plan and price covers what else each tier includes.
What Ultrafast Mode Does
Ultrafast is a speed setting that makes GPT-6.2 answer up to 4.5 times faster than Standard on the same model, according to OpenAI's ChatGPT documentation. You pay for the speed in usage. On Pro $500, Ultrafast draws down the included Work and Codex allowance at 6 times the Standard rate. After the allowance runs out, purchased credits are charged at a promotional 3 times Standard until 28 February 2027.
The bigger change is Pro $200. Until today, Pro $500 was the only Pro plan with Ultrafast. Now Pro $200 subscribers can use it with GPT-6.2 Sol in Codex, at the same 6x drawdown. Plus, Pro $100 and Business still get no Ultrafast at all, not even through credits.
GPT-6.2 API Pricing
GPT-6.2 Astra costs $8 per million input tokens, $0.40 for cached input and $40 for output, and GPT-6.2 Sol costs $2, $0.10 and $10. Astra's input and output prices are 20% below GPT-6 Astra's $10 and $50, and its cached-input price is 60% lower. Sol carries over GPT-6.1 Sol's rates unchanged. The model IDs are gpt-6.2-astra and gpt-6.2-sol.
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.2 Astra | $8 | $0.40 | $40 |
| GPT-6 Astra | $10 | $1 | $50 |
| GPT-6.2 Sol | $2 | $0.10 | $10 |
| GPT-6.1 Sol | $2 | $0.10 | $10 |
All prices are per million tokens at Standard speed. Several price modifiers apply to both GPT-6.2 models, according to OpenAI's model pages. A request with more than 272,000 input tokens pays double for input and cached input and 1.5 times for output on the whole request. Fast mode costs twice the Standard rate. Batch requests cost half.
Ultrafast costs 3 times Standard through 28 February 2027 and 6 times Standard from 1 March 2027. Anyone building on Ultrafast should budget for that price doubling in March.
GPT-6.2 Benchmarks: What OpenAI Claims
OpenAI says GPT-6.2 Astra sets new highs on professional and scientific work while costing far less per task than either GPT-6 Astra or Claude Fable 5.2. Every number in this section comes from OpenAI's launch post. OpenAI ran or chose these tests, and several of them are new, so read them as the vendor's case rather than a neutral ranking.
GPT-6.2 Astra Claims
Agents' Last Exam tests multistep professional work. On it, OpenAI says Astra scores 5.2 points higher than GPT-6 Astra at less than a third of the cost. It also puts Astra 5.5 points ahead of Claude Fable 5.2 "at just 5% of the cost". OpenAI adds a footnote: the leaderboard result for Fable 5.2 excludes nine tasks that used a fallback model.
On Terminal-Bench-Science 0.1, Astra scores 9 percentage points higher than GPT-6 Astra, and OpenAI calls it "the best model in the world for scientists". On Harvey's legal benchmark LAB-AA v1.1, OpenAI says Astra more than doubles Fable 5.2's pass rate at about one-sixth of the cost. On OpenAI's internal GDPval v2, Astra gains 6.2 points over GPT-6 Astra while using 22% fewer output tokens.
OpenAI also points to design work. It says Astra can match a slide template "down to the exact font size", even from a screenshot. On Artificial Analysis's blind 3D test, AA-3D, OpenAI reports Astra first at 1,318 Elo, ahead of Fable 5.2 at 1,271 and GPT-6.2 Sol at 1,259.
GPT-6.2 Astra—a frontier model built for your hardest projects. It sets new highs on evals across professional work and scientific research at less than a third of the cost per task of GPT‑6 Astra.
— OpenAI (@OpenAI) November 17, 2026
You’ll see the difference in things like slideshows, websites, and 3D designs.
GPT-6.2 Sol Claims
OpenAI's headline for Sol is that it nearly matches Astra for a fraction of the price. Cognition's FrontierCode 2.0 checks whether code changes would pass a maintainer's review. On it, Sol comes within 0.5 points of Astra at about one-fifth of the cost, with both at maximum reasoning. On LifeVal, OpenAI's test of everyday tasks like shortlisting kindergartens after a move, Sol beats GPT-6 Astra at about one-seventh of the cost. On GDPval v2, Sol gains 5.3 points over GPT-6.1 Sol.
What Independent Tests Say About GPT-6.2
Independent testing confirms that GPT-6.2 is cheaper to run, but not that it is the smartest model available. The Artificial Analysis Intelligence Index v5 scores GPT-6.2 Astra at 62 and GPT-6.2 Sol at 57, both at maximum reasoning. Claude Fable 5.2 still leads at 65. The scores below were read from Artificial Analysis's model pages on 17 November 2026 and can move as the board updates.
| Model | AA Intelligence Index v5 | Cost per index task |
|---|---|---|
| Claude Fable 5.2 | 65 | $26.54 |
| GPT-6.2 Astra | 62 | $4.00 |
| Gemini 4 Carbon | 60 | $5.48 |
| GPT-6.2 Sol | 57 | $1.01 |
| GPT-6 Astra | 57 | $6.24 |
| GPT-6.1 Sol | 55 | $1.27 |
Two things stand out.
GPT-6.2 Sol now matches GPT-6 Astra's score of 57 for about a sixth of the cost per task, which makes the old Astra hard to justify. And GPT-6.2 Astra lands 3 points behind Fable 5.2 while costing about 15% as much per task. That second gap is the real story of this launch. OpenAI did not overtake Anthropic on raw intelligence, but it made near-frontier work much cheaper.
For context on what these indexes measure and why their versions move scores, see our guide to AI benchmarks.
GPT-6.2 Astra vs Claude Fable 5.2
GPT-6.2 Astra and Claude Fable 5.2 cost exactly the same per token: $8 per million input tokens and $40 per million output tokens. Anthropic cut Fable to that price on 9 November, and OpenAI matched it eight days later. The difference is how many tokens each model spends on a job.
Fable 5.2 scores higher on Artificial Analysis's index, 65 to 62. GPT-6.2 Astra finishes the same index for $4.00 a task against Fable's $26.54, because it uses far fewer tokens to get there. OpenAI's own Agents' Last Exam claim of "5% of the cost" points the same way. That figure comes from a test OpenAI chose, though, and from a Fable result missing nine tasks.
So the choice is quality ceiling against running cost. If you pay per token through the API and run many long agent tasks, GPT-6.2 Astra is the cheaper way to get close to the top. If you need the single best answer and the bill is secondary, Fable 5.2 still leads the independent board. You may not want to pick a side every time one lab leapfrogs the other. Fello AI lets you ask GPT and Claude the same question in one Mac app, without a second subscription.
GPT-6.2 Safety: What OpenAI Disclosed
GPT-6.2 ships with what OpenAI calls "our strongest safeguards to date", and its safety section is unusually frank about what still goes wrong. That matters because the last Astra update never shipped at all.
Why There Was No GPT-6.1 Astra
OpenAI cancelled GPT-6.1 Astra in late September after internal testing. Saachi Jain is OpenAI's head of safety systems. Jain told Al Jazeera that GPT-6.1 Astra did not meet the bar for "scope and authorization, and how it communicates back to the user about the type of work it's done". In plain terms, it went beyond what it was asked to do and did not report its own work back clearly. Our GPT-6.1 Sol article covers that episode in full.
The Reporting Problem Is Smaller, Not Gone
GPT-6.2 still sometimes fails to report what blocked its work, a milder form of the GPT-6.1 Astra problem. OpenAI says both models performed similarly to GPT-6 Astra on most alignment tests and took fewer improper actions in simulated workplace tasks. But in two tests, they were "more likely to omit mentioning obstacles that blocked their work, such as a missing file or unavailable search tool". OpenAI says it added instructions in Codex that "substantially reduce" these omissions and is still investigating.
When GPT-6.2 says a task in Work or Codex is done, check that the file it was meant to edit actually changed.
Cyber and Self-Improvement Ratings
Both GPT-6.2 models meet the Critical threshold for cybersecurity and the High threshold for biological and chemical capability under OpenAI's Preparedness Framework, the same ratings as GPT-6 Astra. Because of that, OpenAI launched them with tighter cyber safeguards, and it says those safeguards "also restrict some legitimate vulnerability research and security testing workflows". Security teams that need fewer refusals can still use GPT-6 Astra through OpenAI's Daybreak program, and OpenAI says it will work to bring the GPT-6.2 models there.
GPT-6.2 Astra is also OpenAI's first publicly released model rated High for AI self-improvement, meaning OpenAI judges it able to meaningfully speed up AI research. OpenAI says it responded with tighter containment and Control Room, a system that tracks research workloads and can halt model output when intervention is needed.
Early Reactions to GPT-6.2: Developers Push Back
The first reactions from developers who tested GPT-6.2 before launch were harsh. Two well-known AI developers said within an hour of the release that they would not recommend GPT-6.2 Astra. Both tested the models before launch, and each is one tester, so read their posts as warnings rather than verdicts.
Matt Shumer Calls GPT-6.2 Unusable
Matt Shumer, the former CEO of HyperWrite, quoted OpenAI's announcement eight minutes after it went up. He said he would not write a full review at all:
I will not be writing a full review of GPT-6.2 Astra / Sol.
— Matt Shumer (@mattshumer_) November 17, 2026
I had a terrible experience testing them. They were essentially unusable.
I do not recommend these models.
Theo on Long Tasks and Unreadable Output
Theo, the CEO of T3 Chat and T3 Code and a developer YouTuber at t3.gg, went further. He published the letter he had sent OpenAI during early testing, titled "Why I hate GPT-6.2 Astra":
https://x.com/i/article/2122759503888437248
— Theo - t3.gg (@theo) November 17, 2026
Theo's letter is not all negative. He calls GPT-6.2 Astra a "great reviewer" that "catches things other models miss", "world class at 3d & game stuff", and says its code quality "has improved meaningfully". That matches what OpenAI claims for 3D and coding.
His complaints are about long, unattended work, which is exactly the job OpenAI built Astra for. Theo says Astra stops two hours into a task to ask questions about internal details no user could answer. He says its explanations are so dense that he switches to Claude Opus 5.5 at the end of a run "just to get a readable summary". On his hardest benchmark project, he says Astra ran for seven days and landed no improvements. His conclusion: "I would not choose this over Opus 5.5 for my work as a developer."
Theo writes that he is "likely overreacting", and he says OpenAI fixed "a small handful" of his complaints before launch. But his point about readable output matters for anyone using ChatGPT Work, where the summary at the end is often all you read.
Should You Switch to GPT-6.2?
If you already use ChatGPT Work or Codex on a paid plan, try GPT-6.2 now, but keep your old model in reach. It costs you nothing extra, and new Work chats already default to GPT-6.2 Sol. Use Astra for decks, 3D work, code review and focused coding tasks. For long, unattended runs, check its progress and its summary before you trust either, because that is where the early testers say it breaks down. If you only use regular Chat, nothing changes today: GPT-6.2 is not there, and Chat still runs GPT-6 Sol.
For API developers, GPT-6.2 Sol is the easy upgrade, with the same price as GPT-6.1 Sol and a higher independent score. GPT-6.2 Astra is the better buy than GPT-6 Astra on every count we could check. Against Claude Fable 5.2, the call depends on whether you are paying for the best answer or for the cheapest near-best answer. For the full ranking across every lab, see the best AI models right now.
Frequently Asked Questions
When was GPT-6.2 released?
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026. Both became available the same day in ChatGPT Work, Codex and the OpenAI API, with the ChatGPT rollout reaching accounts gradually.
Is GPT-6.2 free in ChatGPT?
No. GPT-6.2 is not available on the Free or Go plans. It comes with Plus ($20 a month), Pro, Business, Enterprise and Edu, and only inside ChatGPT Work and Codex, not regular Chat.
What is the difference between GPT-6.2 Astra and GPT-6.2 Sol?
GPT-6.2 Astra is OpenAI's top model for hard, long projects and costs $8 / $40 per million tokens. GPT-6.2 Sol is the everyday model at $2 / $10. Artificial Analysis scores Astra 62 and Sol 57, and OpenAI says Sol comes within half a point of Astra on its FrontierCode 2.0 coding test.
Is there a GPT-6.2 Luna?
No. OpenAI released only GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026 and has not announced a GPT-6.2 Luna.
Is GPT-6.2 better than Claude Fable 5.2?
Not on raw intelligence. Artificial Analysis scores Claude Fable 5.2 at 65 and GPT-6.2 Astra at 62. GPT-6.2 Astra is much cheaper to run, though. Both cost $8 / $40 per million tokens, but Astra cost about 15% as much per Artificial Analysis index task: $4.00 against $26.54.
L’image principale de cet article a été créée avec l’aide de l’IA.