What changes with the two new models
OpenAI has announced the release of two new language models, GPT-6.2 Astra and GPT-6.2 Sol, according to The Decoder. The first is designed for complex professional activities, the second for rapid and versatile everyday use. Both are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and are accessible to developers via API. Pro-500 subscribers can use Ultrafast mode with both; for the first time, Pro-200 users also gain access, starting with Sol.
On pricing, OpenAI reports a 20 percent cut for Astra compared with its predecessor: $8 per million input tokens, $0.40 for cached input and $40 per million output. Sol instead maintains GPT-6.1 Sol's rate, at $2 input, $0.10 for cache and $10 output. The company stresses the price-performance ratio, arguing that Astra beats GPT-6 Astra by 5.2 points on Agents' Last Exam at less than one-third of the cost, and that on Harvey LAB-AA it more than doubles Anthropic's Claude Fable 5.2 pass rate at about one-sixth of the cost.
Performance: claims versus independent verification
OpenAI claims progress in several areas: from the ability to handle professional conventions in legal, financial and accounting contexts to computer use, with demos in which the model builds 3D models, writes physics solvers in Python and produces animated assembly guides. On AA-3D, a benchmark in which human judges compare interactive experiences blind, Astra leads with 1,318 Elo points ahead of Claude Fable 5.2 (1,271) and Sol (1,259). Sol, for its part, trails Astra by just 0.5 points on FrontierCode 2.0, at one-fifth of the cost.
Independent evaluations partly confirm the picture. On the Artificial Analysis Intelligence Index, Astra ranks second behind Claude Fable 5.2, gaining five points over GPT-6 Astra and surpassing Google's Gemini 4 Carbon (60 points) and Anthropic's Claude Opus 5.5 (59). Sol rises two points to 57, matching its predecessor at just 15 percent of the cost per task. According to Artificial Analysis, OpenAI holds the cost-performance Pareto frontier, with Anthropic and Meta the only other names present.
The data, however, are not uniformly positive. Also according to Artificial Analysis, coding is flat: in the specific index, calculated with the Codex system at maximum compute, Astra does not improve over GPT-6 Astra, and Sol loses three points compared with the previous Sol, with a 14-point drop on SWE-Atlas-QnA-AA due to the omission of key facts in answers. A high hallucination rate is also reported: at maximum effort Astra hallucinates in 59 percent of cases and Sol in 55 percent, albeit on particularly difficult questions, and therefore not representative of everyday use.
Cybersecurity and the limits of research
Both models, according to OpenAI, improve in cybersecurity: they find and exploit vulnerabilities more effectively, write more secure code and provide better fixes. They reach the "Critical" threshold of the Preparedness Framework, as the predecessor already did. The company announces stricter safeguards for potentially risky requests, which will make abuse harder but will also limit legitimate vulnerability research. The new models will be available through OpenAI Daybreak, the program that already allows accredited researchers to access with fewer filters; the published numbers on cybersecurity benchmarks come from internal tests with safeguards and classifiers disabled.
One relevant element concerns safety and transparency: according to The Decoder, Astra is the first released model to reach the high self-improvement threshold on OpenAI's internal scale, which the company itself classifies as a risk. Epoch AI also notes that Astra performs complex reasoning within a single output token, without a visible chain of thought, with 12.3 serial steps versus the predecessor's 10.5: models of this type are harder to monitor. On Epoch AI benchmarks, Astra ties Claude Fable 5.2 on the Epoch Capabilities Index (174 points) but lags on real research tasks, where it skipped key estimates and remained too cautious. It remains to be verified, as always with companies' internal evaluations, whether these results actually transfer to concrete use.
