What OpenAI has released
OpenAI has introduced two new language models, GPT-6.2 Astra and GPT-6.2 Sol. According to The Decoder, Astra is aimed at demanding professional work in science, law and finance, while Sol targets fast everyday tasks. Both are available now for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Pro-500 subscribers can use the Ultrafast mode with both models; for the first time, Pro-200 subscribers also get access to Ultrafast, initially with GPT-6.2 Sol.
Prices
According to OpenAI, API prices for GPT-6.2 Astra are below those of its predecessor: 8 dollars per million input tokens, 0.40 dollars for cached inputs and 40 dollars per million output tokens. GPT-6.2 Sol keeps the prices of GPT-6.1 Sol: 2 dollars for input, 0.10 dollars for cached inputs and 10 dollars for output per million tokens.
Performance claims
OpenAI itself reports that Astra scores 5.2 points higher than its predecessor GPT-6 Astra on the Agents' Last Exam, at less than a third of the cost. Compared with Claude Fable 5.2, OpenAI cites a 5.5-point lead at only five percent of the cost. On the AA-3D benchmark, Astra reaches 1,318 Elo points, ahead of Claude Fable 5.2 (1,271) and GPT-6.2 Sol (1,259). These figures come from OpenAI's announcement and from Artificial Analysis's AA-3D ranking respectively.
Independent assessment
In the Artificial Analysis Intelligence Index, GPT-6.2 Astra ranks second with 62 points, just behind Claude Fable 5.2 (65). According to Artificial Analysis, OpenAI models dominate the so-called Pareto frontier for cost-performance. In programming, however, both models stagnate: in the Coding Agent Index Astra stays at its predecessor's level, while Sol loses three points compared with GPT-6.1 Sol. Artificial Analysis also points to high hallucination rates: 59 percent for Astra and 55 percent for Sol on difficult factual questions, though according to Artificial Analysis these values are not representative of everyday use.
Cybersecurity
OpenAI states that both models reach the "Critical" cybersecurity threshold in its Preparedness Framework. Stricter safeguards for risky cybersecurity requests are therefore to apply, which according to OpenAI may also restrict legitimate security research. The published cyber benchmarks come from internal tests with safeguards disabled.
Assessment
For companies using AI models via API, Astra's lower prices could change cost calculations, provided performance holds for their use case. It remains open whether the coding stagnation and hallucination rates matter for factual and research tasks. Benchmark claims should be checked for your own use case before any decision.
Sources
Editorially owned by Ideal Syka. Sources and method: Newsroom & method. Tips and corrections: ai@i6eal.de.




