OpenAI on Tuesday, 17 November, released two new models, GPT-6.2 Astra and GPT-6.2 Sol, pitching the first at demanding professional work and the second at faster, cheaper everyday tasks. Both went live the same day in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and through the company's API, according to an OpenAI developer forum post and reports by The Decoder and 9to5Mac.

Who gets what

A moderator on OpenAI's developer forum said the models were not being added to regular ChatGPT chat, and that access depends on plan, workspace permissions and rollout progress. Developers can call them as gpt-6.2-astra and gpt-6.2-sol.

The launch also brings an "Ultrafast" mode, which OpenAI says can produce tokens up to 4.5 times faster than standard but uses up more of a user's allowance, 9to5Mac reported. Subscribers on the $500-a-month Pro plan get it on both models, while those on the $200 Pro plan get it for the first time, starting with Sol. Enterprise and Edu workspaces can switch it on through an admin.

GitHub said on the same day that both models were generally available in Copilot for Pro+, Max, Business and Enterprise plans, with a gradual rollout across editors including Visual Studio Code, JetBrains IDEs and Xcode.

Prices

According to OpenAI's forum post, GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens, while Sol costs $2 and $10. The Decoder said Astra's prices are 20% below those of the earlier GPT-6 Astra, and Sol keeps the pricing of GPT-6.1 Sol.

What OpenAI claims, and what testers found

OpenAI describes Astra as "a frontier model built for your hardest projects". The company says it scored 5.2 points higher than GPT-6 Astra on a benchmark called Agents' Last Exam while costing far less per task. It also says Sol came within half a point of Astra on Cognition's FrontierCode 2.0 coding test at about a fifth of the cost. These are the company's own figures.

Independent results cited by The Decoder were more mixed. On the Artificial Analysis Intelligence Index, Astra scored 62, second behind Anthropic's Claude Fable 5.2 at 65, while Sol scored 57. The same firm found no gain in coding on its agent index, and measured "hallucination rates" of 59% for Astra and 55% for Sol at maximum effort on a set of deliberately hard factual questions. The Decoder noted that this test measures how often a model guesses instead of admitting uncertainty, and does not reflect everyday error rates.

Epoch AI, which tested Astra before launch, placed it level with Claude Fable 5.2 on its capabilities index, but said it trailed Fable 5.2 on real-world research tasks, The Decoder reported.

Cyber-security safeguards

OpenAI says both models crossed the "Critical" cyber-security threshold under its Preparedness Framework, as GPT-6 Astra had, The Decoder reported. The models launch with tighter safeguards on higher-risk security help, and OpenAI said in its forum post that it is working to bring them to its Daybreak programme so approved defenders face fewer refusals. The Decoder added that OpenAI rates Astra as the first externally released model to reach the "High" level for AI self-improvement, meaning it can meaningfully speed up AI research.

What happens next

The rollout is gradual, so some users may not see the models immediately. A forum moderator advised users to update the ChatGPT desktop app and the Codex command-line tool, and said GPT-6.1 Terra remains the recommended model for prose and fiction.

PS · Unnerving duck, CC BY-SA 4.0