OpenAI releases GPT-6.2 Astra: second place in the rankings and a 20% lower price
OpenAI has released GPT-6.2 Astra and Sol, and the independent Artificial Analysis ranking puts Astra in second place, behind only Anthropic's Claude Fable 5.2. Astra gained 5 points and got 20% cheaper: $8 per million input tokens and $40 per million output tokens. Sol's price is unchanged ($2/$10), but its score matches the previous flagship GPT-6 Astra at ~15% of that model's cost per task. Coding agents see no improvement, so test the model on your own tasks first.
AI-processed from Artificial Analysis; edited by Hamidun News
OpenAI has released two updated models: the flagship GPT-6.2 Astra and the cheaper GPT-6.2 Sol.
The independent lab Artificial Analysis ran them through its Intelligence Index, a composite score across a set of hard tests, and compared what it costs to solve a single task. The result: Astra took second place, behind only Anthropic's Claude Fable 5.2, and got 20% cheaper at the same time.
Imagine you run a support chatbot on GPT-6 Astra in production. If you move to 6.2, it should answer better, and the bill for the same tokens will shrink.
But if you build a coding agent that edits a repository on its own, the ranking gain did not carry over, and it is too early to switch models just because of the version number.
What changed in price and ranking position
GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens, 20% less than GPT-6 Astra. Cached input, meaning the repeated part of a request such as a system prompt or a long document, dropped from $1.00 to $0.40 per million tokens. For bots and agents with a large, constant context, that is a noticeable saving.
GPT-6.2 Sol stays at the previous $2 and $10 per million tokens. It scored 57 points, 2 more than GPT-6.1 Sol. That matches the level of the previous flagship GPT-6 Astra, at roughly 15% of its cost per task.
Astra gained 5 points over GPT-6 Astra and overtook Gemini 4 Carbon (60) and Claude Opus 5.5 (59) at maximum reasoning settings. OpenAI's models also occupy most of the Pareto frontier for cost per task. That term refers to the set of options where you cannot get better quality without paying more. Only Anthropic and Meta share this frontier with OpenAI. On output tokens per task, the frontier belongs almost entirely to OpenAI: the only other lab on it is Claude Fable 5.2.
Where the models are strong
The most notable result concerns science and knowledge. OpenAI leads Terminal-Bench-Science 0.1, which tests agentic workflows for scientific research. On CritPt v2 (the hard part, physics reasoning), it ranks second behind Claude Fable 5.2. On the factual knowledge test AA-Omniscience v1.1 and on HLE-Diamond, which tests reasoning and knowledge, Astra again trails only Anthropic's model.
There are two more results. On AA-3D, where human experts rate interactive 3D scenes created by the model, Astra scored 1318, Claude Fable 5.2 scored 1271, and Sol scored 1259. On the experimental AA-Robotics v0.1, where language models must understand an environment, plan, and carry out actions in the real world, Astra and Sol scored 79 and 78 against 75 for Claude Fable 5.2. The authors themselves caveat that this is a preliminary research version, and that OpenAI's and Anthropic's models already handle most of the simple tasks. The lab is adding harder ones.
The downsides: coding and made-up answers
On the Artificial Analysis Coding Agent Index (Codex harness at maximum effort), Astra stayed at the level of GPT-6 Astra, while Sol lost 3 points relative to GPT-6.1 Sol. The cause is a 14-point drop on the SWE-Atlas-QnA-AA test, where the model answers questions about a codebase: it misses important facts more often than its predecessor. Even so, on cost per task, several effort modes of Astra and Sol still land on the Pareto frontier.
The second caveat concerns fabrication. The AA-Omniscience Hallucination Rate metric shows how often a model guesses when it does not know the correct answer to a factual question. For Astra this figure is 59%, and for Sol it is 55% at maximum effort. If you need verifiable facts, you cannot do without source search and checking the output.
What to do in practice
If your service runs on GPT-6 Astra, moving to 6.2 gives you higher quality and a 20% discount, and savings will be larger if you have heavy cacheable context. If GPT-6 Astra's level is enough for you, try Sol: its score is comparable, and it costs about 15% of the cost per task. For science, 3D, and robotics scenarios, the new Astra is worth testing first.
There is no need to rush with coding agents. The index gain did not carry over, and Sol shows a regression in codebase answers. Run the model on your own tasks before swapping it in production. All figures come from an independent index, but it averages results, and your data may show a different picture.
- Nov 18, 2026Three "token-saving" utilities raised the AI agent bill by 18–46%
- Nov 18, 2026Meituan open-sources 1.6-trillion-parameter agentic model trained without Nvidia
- Nov 12, 2026Jensen Huang steps down as Nvidia CEO: company loses $1.4 trillion in two weeks
- Nov 12, 2026Oakland Union Wins Retraining Instead of AI Layoffs: 94.1% Vote for Strike
Key Questions Answered
What changed in price and ranking position?
GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens, 20% less than GPT-6 Astra. Cached input, meaning the repeated part of a request such as a system prompt or a long document, dropped from $1.00 to $0.40 per million tokens. For bots and agents with a large, constant context, that is a noticeable saving. GPT-6.2 Sol stays at the previous $2 and $10 per million tokens. It scored 57 points, 2 more than GPT-6.1 Sol.
Where the models are strong?
The most notable result concerns science and knowledge. OpenAI leads Terminal-Bench-Science 0.1, which tests agentic workflows for scientific research. On CritPt v2 (the hard part, physics reasoning), it ranks second behind Claude Fable 5.2. On the factual knowledge test AA-Omniscience v1.1 and on HLE-Diamond, which tests reasoning and knowledge, Astra again trails only Anthropic's model. There are two more results.
What are the key details on The downsides: coding and made-up answers?
On the Artificial Analysis Coding Agent Index (Codex harness at maximum effort), Astra stayed at the level of GPT-6 Astra, while Sol lost 3 points relative to GPT-6.1 Sol. The cause is a 14-point drop on the SWE-Atlas-QnA-AA test, where the model answers questions about a codebase: it misses important facts more often than its predecessor. Even so, on cost per task, several effort modes of Astra and Sol still land on the Pareto frontier. The second caveat concerns fabrication.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.