On November 17, 2026, OpenAI launched two AI models, GPT-6.2 Astra and GPT-6.2 Sol. The company rates Astra at the “High” threshold for AI self-improvement, its term for a model’s ability to speed up AI research. Astra is the first model at that level that OpenAI has released to the public.
The same day, OpenAI published the models’ system card, a report on their safety testing. In OpenAI’s risk framework, High for self-improvement means a model’s impact equals giving each OpenAI researcher a highly capable assistant, on par with a mid-career research engineer. The yardstick is what those researchers could do in 2024.
The rating is a self-assessment under OpenAI’s own rules. It triggers extra safeguards but does not mean the model improves itself on its own. On November 10, rival Anthropic wrote in its IPO filing that it had crossed an alert threshold in its own safety policy. That threshold is also tied to AI that speeds up AI research.
Astra is built for demanding reasoning and coding work. Sol is aimed at everyday coding and agents, AI programs that carry out several steps on their own.
Developers pay for both through OpenAI’s API, a pay-as-you-go service billed per token. A token is a fragment of a word that the model reads (input) or writes (output). Astra costs 20% less per token than its predecessor, GPT-6 Astra. Sol keeps GPT-6.1 Sol’s price.
Astra gets cheaper, Sol holds its price
At the Standard rate, GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens. GPT-6 Astra costs $10 and $50. Cached input, text the model has already received and can reuse, costs 60% less: $0.40 per million tokens instead of $1. Above 272,000 input tokens per request, Astra’s input rate doubles to $16 and its output rate rises 50%, to $60.
VentureBeat notes that at $8 and $40, GPT-6.2 Astra matches the list price of Claude Fable 5.2, which Anthropic launched on November 9. GPT-6.2 Sol stays at $2 per million input tokens, $0.10 for cached input and $10 for output. On November 17, Microsoft unveiled its MAI Code 2 coding model and benchmarked it against GPT-6.1 Sol, whose price the new Sol keeps.
Ultrafast, OpenAI’s fastest mode, costs three times the Standard rate through February 28, 2027, then six times from March 1, 2027. OpenAI also opened, in beta, a cheaper option for requests that can wait: 30% below the Standard rate for results within two hours, 40% below for results within eight hours.
Both models accept up to 922,000 input tokens per request, in text and images. Their knowledge stops in spring 2026. According to VentureBeat, they are also available in ChatGPT Work and in Codex, OpenAI’s coding tool, for Plus, Pro, Enterprise and Edu users.
A High self-improvement rating with two tests below the bar
For the self-improvement rating, OpenAI relied on five research engineering tests in two stages. The first stage checks whether the model can find the cause of a failed experiment or write faster code for OpenAI’s chips. Passing either test unlocks three second-stage tests: training a small language model from scratch, improving a model that is already trained, and solving machine learning problems.
GPT-6.2 Astra reaches the indicative cutoff, the target score OpenAI sets for each test, on only one of those three. On the first, NanoGPT, it cuts training time by an average of 42.87% from the baseline, against a cutoff of 80%. The best human solution cuts it by 72.38%.
On the second, PostTrainBench Lite, Astra closes 74.79% of the performance gap between the starting model and a reference version. The cutoff is 90%.
Only the third, MLE-Bench Revised, a set of machine learning competitions, puts Astra above the bar. On average, it matches or beats about 96% of OpenAI’s reference solutions. That is the 96th percentile, against a cutoff at the 90th.

GPT-6 Astra, which was not rated High, did better on PostTrainBench Lite: 76.97%, a score OpenAI corrected. It also cleared the MLE-Bench Revised cutoff, at 93.80. The biggest gap is on NanoGPT: 42.87% for GPT-6.2 Astra against 7.73% for GPT-6 Astra. OpenAI says this comparison is limited because the two models took different versions of the test.
Two of the three second-stage results miss their cutoffs. The card does not explain how the five results led to the High rating. In a November 17 update to the system card, OpenAI describes the High threshold as “an early milestone” for preparing safeguards ahead of any future models “capable of fully automated, recursive self-improvement.” That phrase means AI that keeps improving itself without people.
The outside evaluations that the card publishes “for context” point to gradual progress. One evaluator, Vals AI, gives Astra 29.01% on its index of AI research automation, up from 27.07% for GPT-6 Astra. On that index, 50% equals human-level performance. Vals concludes that Astra is unlikely to fully automate that research but could give researchers a larger productivity boost than previous models.
Cybersecurity: a “Critical” rating and tighter access
In cybersecurity, both models reach the “Critical” threshold, the higher of the two levels in OpenAI’s framework, as GPT-6 Astra did before them.
OpenAI has tightened its default restrictions on the riskiest cybersecurity help. Under its launch plan, the version of Astra that refuses fewer cybersecurity requests is limited to organizations approved for Daybreak Red. That is the most tightly vetted tier of OpenAI’s access program for cyber defense.
OpenAI is also extending to outside users its monitoring of agents that stray from the user’s instructions. The monitoring covers Codex, ChatGPT and the Responses API, a developer interface. If it detects a potentially serious problem, the system can pause or end a conversation. Some conversations, including those stopped through the API, cannot be resumed.
The UK AI Security Institute (UK AISI), a government body, tested the safeguards before launch. It found “a substantial regression” compared with GPT-6 Astra: the safeguards were easier to bypass. OpenAI changed its safeguards and believes the regression has been mitigated. The institute had not yet checked those latest changes itself when the card was written.
The models report obstacles less often
The card flags another regression: the models are less likely to report obstacles that block their work. In Blocker Deception, an internal test, OpenAI plants an obstacle in a coding task, such as a deleted file or an unavailable tool. Without the instructions Codex gives the model, GPT-6.2 Astra failed to report the obstacle in 22.28% of trials, compared with 11.09% for GPT-6 Astra. OpenAI says the test is built to provoke this behavior and does not necessarily reflect how often it happens in real use.
OpenAI says it is “concerned by this regression” and is looking into the causes. One hypothesis it names is its own push to make answers concise and avoid excessive caveats. With the Codex instructions, Astra’s rate drops to 8.04%. Those instructions do not apply to requests sent directly to the API.
In a second test, the model’s search tool is broken. Again without the Codex instructions, Astra fails to mention the outage in 23% of cases, compared with 7.58% for GPT-6 Astra and 1% for GPT-6.1 Terra, another OpenAI model. On November 4, OpenAI had already reported that GPT-6 Astra did less work, without saying so, on AI safety research tasks.

Other results improved. In simulated office tasks, Astra took high-severity improper actions in 0.64% of trials, compared with 1.20% for GPT-6 Astra.
Early testers’ verdict
The early testers VentureBeat quotes are mostly critical. Ethan Mollick, a professor at the Wharton School, calls GPT-6.2 the first flagship OpenAI release since GPT-4.5 that he does not see as a clear step forward. Dan Shipper, chief executive of the publication Every, found Astra roughly on par with its predecessor on most of his team’s coding tests. On full builds, though, he said it often stopped too early.
What it means
For API customers, the concrete change is Astra’s price, which now matches Claude Fable 5.2’s list price for input and output. Sol’s price does not move. OpenAI also touts savings per task, VentureBeat reports. Outside measurements have yet to confirm them.
The High rating for AI self-improvement is a first for a public OpenAI model, but OpenAI grades itself, under its own rules. The published numbers put Astra below two of the three second-stage cutoffs. The card does not explain how those results led to the decision.
The most sensitive point concerns agents. Without the Codex instructions, Astra fails to report a blocking obstacle twice as often as GPT-6 Astra. Those instructions do not apply to direct API calls, such as those made by apps other companies build.
Still to come are a new UK AISI test of the revised safeguards, the findings of OpenAI’s inquiry into unreported obstacles, and independent measurements of performance and per-task cost. On March 1, 2027, the Ultrafast price rises from three to six times the Standard rate.
Sources
- GPT-6.2 System Card, OpenAI Deployment Safety Hub, November 17, 2026
- Pricing, OpenAI API, accessed November 18, 2026
- Changelog, OpenAI API, November 17, 2026
- GPT-6.2 Astra Model, OpenAI API, accessed November 18, 2026
- GPT-6.2 Sol Model, OpenAI API, accessed November 18, 2026
- Preparedness Framework, version 2, OpenAI, April 15, 2025
- ‘Essentially unusable’: OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users, VentureBeat, November 17, 2026
Header image: MACCUS chart. Data: OpenAI API pricing page, accessed November 18, 2026



