OpenAI has introduced two new frontier models, GPT-6.2 Astra and GPT-6.2 Sol, designed to handle complex professional workflows and everyday tasks respectively. The company reported that both models deliver improved results while using fewer output tokens to maximize user limits. Astra is positioned as a high-performance model for specialized fields such as science, law, and engineering, while Sol is intended as a versatile, efficient option for general productivity.
According to OpenAI, GPT-6.2 Astra demonstrates significant advancements in professional benchmarks. On the Agents’ Last Exam, which evaluates multistep professional work, Astra scored 5.2 points higher than GPT-6 Astra at less than one-third of the cost. The model also outperformed Claude Fable 5.2 by 5.5 points on the same test while operating at 5% of the competitor's cost. In legal applications, Astra more than doubled the pass rate of Claude Fable 5.2 on the Harvey LAB-AA v1.1 benchmark at approximately one-sixth of the cost.
The company stated that Astra is optimized for scientific and technical work, setting new highs on the Terminal-Bench-Science 0.1 and the Vals Science Index. The model's capabilities extend to 3D design and computer use, where it can simulate mechanical systems and animate results in Blender. On the Artificial Analysis AA-3D benchmark, which uses human judges to rank interactive browser experiences, Astra ranked first with an Elo of 1,318, surpassing Claude Fable 5.2 and GPT-6.2 Sol.
For developers and businesses, OpenAI announced a 20% reduction in input and output token prices for Astra compared to GPT-6 Astra. Cached input prices for Astra are set at $0.40 per million tokens, with standard input at $8 and output at $40 per million tokens. GPT-6.2 Sol maintains the same API token pricing as its predecessor, GPT-6.1 Sol, but the company noted it is smarter and more efficient, allowing users to complete more tasks on the same paid plans.
In performance testing, GPT-6.2 Sol reached within 0.5 points of Astra in Cognition’s FrontierCode 2.0 at one-fifth of the cost. The model also outperformed GPT-6 Astra on the LifeVal evaluation, which measures performance on challenging everyday tasks, at about one-seventh of the cost. OpenAI reported that many of its own employees now use Sol as their default model for both technical and non-technical tasks.













