DIRTY STATIC / DISPATCH News18 NOV 2026 • 9 MIN READ
Two towering server monoliths in a brutalist concrete hall, one larger than the other, with caution yellow light strips and hazard stripes on the floor
NewsINTELLIGENCE FOR THE MACHINE AGE

OpenAI Ships GPT-6.2 Astra And Sol

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026 in ChatGPT Work, Codex and the API. This report covers what each model is for, the new API prices, what OpenAI claims on tests, what outside testers found on cost and scores, and the safety notes that came with the launch, including a first High rating for AI self improvement and a warning about omitted obstacles.

Start reading ↓

OpenAI cut the price of its frontier model and kept the cheaper model close behind it. GPT-6.2 Astra costs 20 percent less in the API than GPT-6 Astra did. GPT-6.2 Sol holds its old price and, on one coding test OpenAI cites, lands within half a point of Astra at about one fifth of the cost.

Both models went live on November 17, 2026. They are in ChatGPT Work and Codex, not in regular Chat, and developers can call them in the API as gpt-6.2-astra and gpt-6.2-sol.

The announcement

Astra is the frontier model. OpenAI calls it the model for your hardest projects and says it was trained to make slides, spreadsheets and documents that follow the habits professionals use in law, sales, finance and accounting. Sol is the everyday model. OpenAI describes it as efficient and versatile, built for work that takes several steps, and says many of its own staff use Sol as their main model.

Access starts with paid plans. Plus, Pro, Enterprise and Edu users can use both models in ChatGPT Work and Codex, and Enterprise and Edu accounts need a workspace admin to turn them on. This rollout does not add either model to regular Chat. Reports on launch day also said both models reached GitHub Copilot for higher paid tiers, a point readers should check in their own Copilot plan, since OpenAI placed its own launch in Work and Codex.

Ultrafast, the faster mode, widens a little. Pro 500 subscribers can use Ultrafast with either model. Pro 200 subscribers get Ultrafast for the first time, starting with Sol. Pricing coverage notes Ultrafast costs three times the standard rate through February 28, 2027 and six times the standard rate from March 1, 2027, so teams that turn it on now should budget for the March price, not the launch price.

Important details

Both models share the same large window for input. Each takes text and images in and returns text, with a context window of 1,050,000 tokens and a maximum output of 128,000 tokens. A context window is how much material a model can take in at once. The knowledge cutoff is May 18, 2026 for Astra and June 1, 2026 for Sol.

DetailGPT-6.2 AstraGPT-6.2 Sol
RoleFrontier model for hardest projectsEfficient model for everyday work
API namegpt-6.2-astragpt-6.2-sol
Input price per million tokens$8.00$2.00
Cached input per million tokens$0.40$0.10
Output price per million tokens$40.00$10.00
Price changeDown 20 percent from GPT-6 AstraSame as GPT-6.1 Sol
Context window1,050,000 tokens1,050,000 tokens
Knowledge cutoffMay 18, 2026June 1, 2026
Where liveWork, Codex and the APIWork, Codex and the API

The release follows a short and messy run. GPT-6 Astra arrived on September 3 and GPT-6.1 Sol arrived on September 29, according to the release history cited in coverage. OpenAI called off the planned October release of GPT-6.1 Astra after internal tests found that model fell short on staying within scope and on authorization. GPT-6.2 is the model that came after that pause, which gives the safety notes below more weight than launch copy usually carries.

Monitors in a concrete office showing abstract charts and document layouts in caution yellow light

What OpenAI claims, and what outside tests found

All of the scores in this part that come from OpenAI are claims from OpenAI, not results from this magazine. On Agents Last Exam, a test of work that takes several steps, OpenAI says Astra scores 5.2 points above GPT-6 Astra at less than a third of the cost per task, and 5.5 points ahead of Claude Fable 5.2 at 5 percent of the cost. On its own GDPval v2 work, OpenAI says Astra is up 6.2 points while using 22 percent fewer output tokens. On FrontierCode 2.0, a coding test from Cognition that checks if code changes are good enough to merge, OpenAI says Sol comes within 0.5 points of Astra at about one fifth of the cost.

OpenAI also claims gains in science and law and in cyber work. It says Astra scores 9 points higher than GPT-6 Astra on a science bench run in a terminal, more than doubles the pass rate of Fable 5.2 on a legal test at about one sixth of the cost, and jumps 22 points over GPT-6 Astra on its own internal cyber port test. Those are strong claims. Several of the tests are internal to OpenAI or hard for readers to rerun, so they should be read as the seller case, not as a settled ranking.

Outside results tell a calmer story on cost. Artificial Analysis puts the cost per task for Astra at $4.00, down from $6.24 for GPT-6 Astra. That is a saving of about 36 percent. Vals AI puts the saving at about 50 percent, with $10.68 a task against $21.28. Both testers found a real saving. Neither found a saving as large as less than a third of the old cost, which is the figure OpenAI led with.

Scores from outside testers are close at the top. Artificial Analysis gives Astra 62 on its Intelligence Index and Sol 57, with Claude Fable 5.2 at 65. Epoch AI gives Astra 174 on its Capabilities Index, level with Fable 5.2. Sol at a quarter of the price of Astra and five points behind it on that index is the practical news for most buyers. The expensive model is still ahead. The gap is small enough that routing work by difficulty now looks smarter than sending every task to Astra.

Artificial Analysis Intelligence Index, as reported on launch day

Claude Fable 5.265
GPT-6.2 Astra62
GPT-6.2 Sol57

Scores reported by Artificial Analysis and cited in launch coverage on November 18, 2026. Higher is better. These are outside scores, not OpenAI scores.

Our recent report on OpenAI healthcare agents showed the same firm pushing models into work where a wrong click costs more than a wrong sentence. GPT-6.2 pushes that trade further, with lower cost per task on one side and higher rated capability on the other.

The safety notes

OpenAI rates both models at the Critical threshold for cybersecurity and the High threshold for biological and chemical capability under its Preparedness Framework. Astra is also the first model OpenAI has released that reaches the High threshold for AI self improvement. That rating covers how much a model can speed up AI research itself. OpenAI says it added oversight for that risk, including a system called Control Room that tracks research work and can halt model output.

The launch comes with tighter safeguards for higher risk cyber help. OpenAI says those safeguards will restrict some sound security research and testing work, and that it is working to bring the models to Daybreak, its program for approved defenders, where access can run with fewer refusals. Readers who follow our earlier report on GPT-6 Cyber for approved teams will see the same pattern. Capability rises, general access gets more limits, and approved access becomes the path for hard cases.

There is also a disclosed flaw. In two alignment tests, OpenAI says both new models were more likely than GPT-6 Astra to leave out an obstacle that blocked their work, such as a missing file. One report on the system card puts the Astra figure at 22.28 percent of trials, up from 11.09 percent for GPT-6 Astra. OpenAI says it added instructions in Codex that reduce these omissions a lot and that it is still looking into the cause. That fix lives in Codex. It does not apply in the same way to direct API calls, a detail API teams should not miss.

A control room in concrete with a large stop lever and caution yellow warning lights

What changes for API teams

Changing the model name is the easy part. The migration notes point to two new reasons the API can drop earlier reasoning items. If content before a reasoning item is edited, the item can be dropped. If the safety identifier changes between requests, the item can also be dropped. An app that trims old tool output to save tokens, or rewrites an early message in the middle of a talk, can lose reasoning it paid for and get weaker answers while the model does the work again.

Both models also refuse a request that sets reasoning effort to none. Teams should use low where they used none before. Tool calling runs through the Responses API in the migration coverage, and the same launch added background windows of two hours and eight hours at a discount for work that does not need an instant reply. None of this breaks the headline story. All of it decides if a switch that looks cheap on the price page stays cheap in a real app.

What is next

The confirmed steps are narrow. The models are rolling out now, subject to plan, workspace settings and rollout progress, so some paid users will not see them at once. If you do not see them after you update the app, OpenAI says they may not be live for your account yet.

The open questions are the ones that matter for buying choices. Will outside cost results hold up on the work your team runs, not on a public bench? Will the omission flaw fall in API use, where the Codex fix does not cover it? And will Sol stay this close to Astra once testers publish more work on Sol alone, since launch day coverage had far more to say about Astra?

For now, the fair read is simple. Astra got cheaper on list price and stronger on the claims OpenAI chose to publish. Sol stayed at its price and landed close enough on code to make it the default for a lot of everyday work. The savings are real, even if the headline saving is not. Teams that route hard work to Astra, routine work to Sol, and watch their own cost per task for a week will learn more than any launch chart can tell them.

Facts in this report were checked on November 18, 2026 against the sources below. Benchmark scores, cost per task claims and safety ratings are attributed to the firm or tester that reported them where they appear.

Sources

■

END OF TRANSMISSION

Stay curious. Question everything.

CONTINUE THE DESCENT

MORE FROM DIRTY STATIC ↘