Uni-Sine
AIUsing AI

OpenAI releases GPT-6.2 Astra and Sol at lower prices

The words Astra, GPT-6.2 and Sol in white on a dark blue sky between a gold star trail and a blue-white one
OpenAI's launch artwork for GPT-6.2 Astra and Sol. Image: OpenAI, from its announcement of 17 November 2026, resized.

On 17 November 2026, OpenAI released two new AI models: GPT-6.2 Astra, its new frontier model for the hardest projects, and GPT-6.2 Sol, a cheaper model for everyday work. Both went live that day for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and for developers through the OpenAI API.

OpenAI's pitch is efficiency: both models reach better results with fewer output tokens, and Astra's prices per token are 20% lower than those of GPT-6 Astra, the model it replaces. The independent tester Artificial Analysis ranks Astra second on its Intelligence Index, behind Anthropic's Claude Fable 5.2.

GPT-6.2 Astra sets a new quality bar for AI in complex professional work.
OpenAIIntroducing GPT-6.2 Astra and Sol, 17 November 2026source

What OpenAI released

  • GPT-6.2 Astra: the top model, for long professional jobs such as spreadsheets, slide decks, legal work, science and code. OpenAI says it follows templates more closely and writes cleaner code than GPT-6 Astra.
  • GPT-6.2 Sol: the efficient model, which OpenAI says many of its own staff use by default. It keeps the API prices of GPT-6.1 Sol.
  • Ultrafast: a faster mode, for Pro 500 subscribers on both models and, for the first time, for Pro 200 subscribers, starting with Sol. Both models have it in the API from launch.

The prices

API prices are set per million tokens, the small pieces of text a model reads and writes. Astra's input and output prices are 20% lower than GPT-6 Astra's, and cached input, text the model has already read in an earlier request, is 60% cheaper.

Model, US dollars per million tokensInputCached inputOutput
GPT-6.2 Astra$8$0.40$40
GPT-6 Astra (worked back from the cuts)$10$1.00$50
GPT-6.2 Sol, same as GPT-6.1 Sol$2$0.10$10

The benchmark scores

OpenAI published each model's score and cost per task at up to five reasoning settings, from low to max. The chart takes the best score each model reached on OpenAI's charts, and the cost per task at that setting on Agents' Last Exam. The numbers are the ones OpenAI published with the launch; the Intelligence Index is run by Artificial Analysis.

Agents' Last Exam v1.1

Long, multistep professional work. The best score each model reached across the reasoning settings OpenAI tested.

  1. GPT-6.2 AstraOpenAI78.3%
  2. Gemini 4 CarbonGoogle77.4%
  3. GPT-6 AstraOpenAI73.1%
  4. GPT-6.2 SolOpenAI73%
  5. Claude Fable 5.2Anthropic72.8%
  6. GPT-6.1 SolOpenAI72.5%
  7. Claude Opus 5.5Anthropic71.6%
Source: OpenAI, Introducing GPT-6.2 Astra and Sol, 17 November 2026 (Intelligence Index by Artificial Analysis)

On Agents' Last Exam, Astra scores 78.3%, 5.2 points above GPT-6 Astra at about a third of the cost per task ($2.89 against $8.99), and 5.5 points above Claude Fable 5.2, which cost $60.05 a task. Google's Gemini 4 Carbon is close behind on 77.4%. On Terminal-Bench-Science, Astra is 9 points above GPT-6 Astra.

The chart also shows where Astra does not lead. On FrontierCode 2.0, Claude Fable 5.2 and Claude Opus 5.5 both score higher, and Sol comes within half a point of Astra at about a fifth of the cost. On the Intelligence Index, Claude Fable 5.2 is ahead by 2.5 points at more than six times the cost per task.

Where the savings come from

The cost of a task is the number of tokens a model uses times the price of each one, so cutting both compounds. On OpenAI's internal GDPval v2 test of long knowledge-work tasks, Astra at its highest setting used 35,812 output tokens against GPT-6 Astra's 45,732, about 22% fewer, while scoring 50.4% against 44.2%. With the output price down from $50 to $40 per million tokens:

\frac{35812}{45732}\times\frac{40}{50}\approx0.627New output cost as a fraction of the old one

So the output part of the bill for the same set of tasks is about 37% lower, before the cheaper cached input is counted.

What independent testers found

Artificial Analysis, which runs the same tests on every model, put Astra 5 points up on GPT-6 Astra in its Intelligence Index and ahead of Gemini 4 Carbon and Claude Opus 5.5, so that OpenAI now trails only Claude Fable 5.2. Sol scores 57, matching GPT-6 Astra at about 15% of its cost per task.

It also found weak spots. Coding did not improve on its Coding Agent Index: Astra is level with GPT-6 Astra, and Sol is 3 points below GPT-6.1 Sol. On AA-Omniscience, its test of hard knowledge questions, it measured a hallucination rate of 59% for Astra and 55% for Sol at maximum effort.

Safety ratings

OpenAI rates both models at its Critical level for cybersecurity under its Preparedness Framework, as it did GPT-6 Astra, and at High for biological and chemical capability. It launched them with tighter limits on higher-risk security requests, which it says also restrict some legitimate security testing.

OpenAI says Astra is its first externally released model to reach High for AI self-improvement, which it describes as equal to giving every OpenAI researcher an extra, highly capable research engineer. In two alignment tests, both models were more likely than before to leave out an obstacle that blocked their work, such as a missing file. OpenAI says instructions it added in Codex substantially reduce this, and that it is still investigating.

Related articleWhat is a token in AI?How text becomes tokens, and why AI prices are counted per million of them.

Sources

Read next
AI · Using AIAI coding for beginners: how to start, step by stepHow to start coding with AI as a beginner: the kinds of AI coding tools, a step-by-step first project, how to fix errors, and how to learn as you go.