OpenAI unveiled two new models on November 17, 2026: GPT-6.2 Astra for demanding professional tasks and GPT-6.2 Sol for everyday work. Both are available immediately in ChatGPT Work, in Codex, and in the API.
For Astra, OpenAI is cutting API prices per token by 20 percent, and by 60 percent for reused, cached inputs. Sol costs as much as its predecessor GPT-6.1 Sol. OpenAI does not state a euro price; all amounts are in US dollars.
Who gets the models?
Everyone on a Plus, Pro, Enterprise, or Edu plan has access. In Enterprise and Edu, admins must first enable the models. Ultrafast mode is available with both models on the Pro 500 plan. It is available on the Pro 200 plan for the first time, but initially only with Sol. OpenAI makes no statements about Germany or the EU.
By our own calculation, a run with one million input tokens and 100,000 output tokens costs 12 with Astra and 3 with Sol without caching, in each case in US dollars.
How good are Astra and Sol?
OpenAI puts Astra's lead on Agents' Last Exam, a test for multi-step professional tasks, at 5.2 points ahead of GPT-6 Astra, at less than a third of the cost. Compared with Anthropic's Claude Fable 5.2, the lead is 5.5 points at 5 percent of the cost. For Fable 5.2, OpenAI excludes nine tasks in which a substitute model stepped in.
Intelligence Index im Vergleich
Messung von Artificial Analysis auf der höchsten Stufe, abgerufen am 18. November 2026
1Claude Fable 5.265 Punkte
2GPT-6.2 Astra62 Punkte
3GPT-6.2 Sol57 Punkte
Die Punkte sagen nichts über die Kosten. Die Kosten je Aufgabe unterscheiden sich stark, Einzelheiten stehen im Text.
Quelle: Artificial Analysis
On FrontierCode 2.0, a test for changes to real codebases, OpenAI measures Sol as 0.5 points behind Astra, at about a fifth of the cost.
An independent cross-check is provided by Artificial Analysis. In the Intelligence Index, Astra reaches 62 points at the highest level, Sol 57, and Claude Fable 5.2 65. In terms of price per million tokens, Astra and Fable 5.2 are on par: 8 for input and 40 for output, in each case in US dollars. For the cost per task in the index, Astra comes in at 4.00, Sol at 1.01, and Fable 5.2 at 26.54, also in US dollars.
So Astra costs around 15 percent of Fable 5.2 per task, more than the 5 percent from OpenAI's test, but still only a fraction. With the same price per token, consumption explains the gap: Fable 5.2 generated 275 million output tokens in the index, while the median of the compared models is 65.1 million.
OpenAI classifies both models as "Critical" for cybersecurity and concedes that the stricter safeguards also restrict some legitimate security testing. In two evaluations, both models more often left obstacles such as a missing file unmentioned than GPT-6 Astra did. OpenAI has added instructions in Codex that significantly reduce such omissions.
What this means for you
For everyday use, it's best to try Sol first. At Artificial Analysis, Sol is five points behind Astra and needs only a quarter of the cost per task. Astra is worthwhile for tasks where Sol falls short. At the end of October, OpenAI had cut API prices by at least 50 percent with GPT-6.1 Terra.
In the API, it's best to leave the beginning of your prompt unchanged and append new instructions to the history. OpenAI recommends this because it allows the cache to be reused better and costs to fall.
In ChatGPT Work, it's best to quickly check whether both models are available for selection on your plan.
How to phrase tasks for ChatGPT cleanly and check the results is shown in the ChatGPT Masterclass on TutKit.
OpenAI presented the models on November 17, 2026. We retrieved the values from Artificial Analysis on November 18, 2026.