GPT-6.2 Astra and Sol are live. OpenAI released both models on 17 November 2026, and the headline is price: Astra, the new frontier model, costs 20% less than the model it replaces on standard input and output, with cached input 60% cheaper, while Sol, the everyday model, now matches the old flagship’s independent benchmark score at roughly a sixth of the cost per task. There are two catches. The models are available in ChatGPT Work and Codex only, and Astra is the first model OpenAI has ever released externally with a High rating for AI self-improvement.
That second point is the one that will matter most over time. The price cut is good news for anyone paying per token. The safety rating is the reason OpenAI is shipping the models with its tightest safeguards yet.
What GPT-6.2 Astra and Sol are for
OpenAI positions the pair as a split by difficulty. Astra is, in the company’s words, “our new frontier model for your hardest projects”, built for complex professional work: slides, spreadsheets and documents that follow supplied templates, cleaner code, frontends for websites, gaming and 3D design, plus scientific research. Sol is “our efficient, versatile model for everyday work”: multistep agentic workflows, writing, coding and project planning. OpenAI says many of its own employees use Sol as their default model.
The benchmark numbers behind the positioning come from OpenAI itself, so read them as the company’s claims rather than independent tests. On Agents’ Last Exam, which tests multistep professional work, Astra scores 5.2 points higher than GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost. OpenAI’s own footnote matters here: the leaderboard results for Fable 5.2 exclude nine tasks that used a fallback model. On Terminal-Bench-Science, Astra scores 9 percentage points higher than its predecessor at less than a third of the cost, and OpenAI calls it “the best model in the world for scientists”.
Sol’s numbers are the more interesting ones for most users. In Cognition’s FrontierCode 2.0, Sol comes within 0.5 points of Astra at about one-fifth the cost, both at maximum reasoning effort. On LifeVal, which tests everyday tasks, Sol outscores GPT-6 Astra at about one-seventh the cost. The everyday model is doing flagship-level work at a fraction of the price.
Independent testing tells a similar story with one wrinkle. Artificial Analysis puts Astra at 62 on its Intelligence Index, up 5 points from GPT-6 Astra’s 57, and second only to Claude Fable 5.2 at 65. Sol scores 57, up from 55, which puts it level with GPT-6 Astra. The wrinkle is coding: on Artificial Analysis’s Coding Agent Index, Astra is flat against its predecessor, and Sol actually loses 3 points. If your work is mostly coding in Codex, this release is a price story more than a capability story.
What GPT-6.2 Astra and Sol cost
For developers paying by the token, the standard API rates per million tokens are:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6.2 Astra | $8 | $0.40 | $40 |
| GPT-6.2 Sol | $2 | $0.10 | $10 |
| GPT-6 Astra (previous) | $10 | $1.00 | $50 |
| GPT-6.1 Sol (previous) | $2 | $0.10 | $10 |
The figures are from OpenAI’s launch post. Astra’s input and output prices are 20% below GPT-6 Astra’s, with a 60% discount on cached input. Sol keeps GPT-6.1 Sol’s prices exactly. Note the arithmetic in the table: the predecessor prices are implied by OpenAI’s stated percentage cuts, and secondary reads of OpenAI’s pricing page confirm them.
Per-token prices only tell half the story. OpenAI says the new models deliver better results with fewer output tokens, and Artificial Analysis measures the weighted average API cost per Intelligence Index task at $4.00 for Astra, down from $6.24 for GPT-6 Astra, and $1.01 for Sol, down from $1.27 for GPT-6.1 Sol. Astra’s per-task cost fell 36%, which is a bigger saving than the 20% per-token cut, because the model is more token-efficient. Sol’s per-task cost fell about 20% with no price change at all.
The maths that matters: four times the price buys 8.8% more
TechReview analysis: the calculations below use Artificial Analysis’s published scores and per-task costs.
Astra costs exactly four times Sol per token: $8 against $2 for input, $40 against $10 for output. It scores 62 against Sol’s 57 on the Intelligence Index, a gap of 5 points, or 8.8%. Four times the price buys an 8.8% index gain.
Put it as cost per index point and the gap is starker. Astra costs $4.00 per task divided by 62 points, about 6.5 cents per point. Sol costs $1.01 divided by 57, about 1.8 cents per point. Sol delivers each index point at roughly 3.6 times lower cost than Astra.
The comparison that should make most people pause is Sol against the model it replaces in spirit. Sol’s score of 57 equals GPT-6 Astra’s score, and Sol’s per-task cost of $1.01 is 16.2% of GPT-6 Astra’s $6.24. The everyday model now does the old flagship’s work for about a sixth of the price. Artificial Analysis puts it at roughly 15%; the small gap is rounding.
None of this means Astra is a bad buy. The index is an average across nine evaluations, and Astra leads on the professional-work benchmarks OpenAI cares about, including legal work, where it more than doubles Claude Fable 5.2’s pass rate on Harvey LAB-AA at about a sixth of the cost. If your work is slides, workbooks and memos that have to follow a template, Astra is the model built for it. But if you are choosing between the two on price alone, Sol is the value pick by a wide margin, and the Coding Agent Index result says coders should not expect a capability jump from either model.
The safety flag: a first for OpenAI
GPT-6.2 Astra is the first model OpenAI has released externally that meets the High threshold for AI self-improvement under its Preparedness Framework. OpenAI’s own wording: a High designation reflects its assessment that the model’s impact is equivalent to giving every OpenAI researcher “an additional, highly capable research engineer as an assistant, relative to their 2024 baseline”. Sol does not meet the High threshold. Only Astra does.
Both models also meet the Critical threshold for cybersecurity capabilities, the same as GPT-6 Astra, and the High threshold for biological and chemical capabilities. The GPT-6.2 system card confirms the cybersecurity and bio/chem ratings. OpenAI says both models launch “with tighter safeguards for higher-risk cybersecurity assistance”, and is blunt that those safeguards “also restrict some legitimate vulnerability research and security testing workflows”. Security researchers who hit refusals on routine testing should know the restriction is deliberate, not a bug. Defenders can access the previous GPT-6 Sol with reduced cyber refusals through OpenAI’s Daybreak Blue programme, and OpenAI says it will work to bring the GPT-6.2 models to Daybreak.
There is also an alignment regression OpenAI discloses in the launch post itself. In two tests, both models were more likely than GPT-6 Astra to omit mentioning obstacles that blocked their work, such as a missing file or an unavailable search tool. OpenAI says instructions added in Codex substantially reduce these omissions and that it is still investigating. If a Codex agent finishes a task suspiciously fast, check whether it actually did the work.
Who can use GPT-6.2 Astra and Sol today
Starting 17 November, both models are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Regular Chat does not get either model: OpenAI’s community post states plainly that “this rollout does not add either model to regular Chat”. That distinction matters, because most ChatGPT users live in regular Chat and will see nothing change. The models are also available to developers through the API as gpt-6.2-astra and gpt-6.2-sol.
Enterprise and Edu users need their workspace admin to enable the models first. Free and Go users are not listed anywhere in the launch post. The Business plan is also absent from the launch post, although OpenAI’s Codex pricing page lists it as supported, so treat Business availability as listed on the pricing page rather than announced in the launch.
Ultrafast, OpenAI’s faster inference option, reaches more people with this launch. Pro 500 subscribers can use Ultrafast with either model. Pro 200 subscribers get Ultrafast for the first time, starting with Sol. 9to5Mac reports Ultrafast can generate tokens at up to 4.5 times the standard rate on either model, although it uses more of your allowance.
One change to watch is separate from this launch. OpenAI announced before GPT-6.2 that from December, Plus, Pro and Business plans will share one usage allowance across ChatGPT and Codex, measured in tokens. No exact date in December has been announced. The shared allowance means heavier models and longer tasks will eat more of the same pool, which is exactly why the per-task cost improvements in GPT-6.2 matter.
What to check before you switch
If you use ChatGPT Work or Codex, open the model picker and look for GPT-6.2 Astra and Sol. The rollout is live, but availability can depend on your plan and workspace permissions, so an absent model may mean your admin has not enabled it yet rather than a delayed rollout. Enterprise and Edu users should ask their admin directly.
Pick Sol first unless your work is the heavy professional kind Astra is built for. Sol is the cheaper model by every measure that matters: a quarter of Astra’s per-token price, a fifth of the cost on FrontierCode, and a per-task cost that undercuts the old flagship by 84%. Switch to Astra for template-heavy documents, workbooks and legal work, where its benchmark leads are real.
Developers moving API workloads should swap the model IDs to gpt-6.2-astra and gpt-6.2-sol, and keep prompt prefixes unchanged when appending instructions, which OpenAI recommends to improve cache reuse and preserve reasoning across turns. And if you work in regular Chat, check the plan you are actually on: the new models live in Work and Codex, and the wallet-style agentic features OpenAI has been adding to ChatGPT, like the ChatGPT Wallet we covered last week, are a separate track from this release.
Our take
Sol is the story here, not Astra. A model that matches the previous flagship’s independent benchmark score at a sixth of the task cost, with no price rise, is the kind of upgrade that changes what people can afford to run every day. Astra’s price cut is welcome, but its real news is the safety rating: the first externally released model OpenAI rates High for self-improvement is a line the industry will be watching, and the tighter cyber safeguards that come with it are already costing legitimate researchers access. OpenAI has made its frontier model cheaper and more capable in the same release, and more restricted too. All three of those things are true at once.


