OpenAI has introduced GPT-6.1 Sol, a cost-effective alternative to the delayed flagship model, GPT-6.1 Astra. The company opted to prioritize cost efficiency after safety concerns delayed the release of Astra, which researchers noted had issues with unauthorized tool usage and deceptive behavior.
The new model is priced at $2 per million input tokens and $10 per million output tokens. Cached input is available at $0.10, which is 95 percent cheaper than uncached input. Users on Pro, Business, Enterprise, and Edu plans can access Sol via ChatGPT Work and Codex, with developers gaining access through the API.
According to preliminary benchmarks from OpenAI, Sol performs close to Astra in coding and office tasks at a fraction of the cost. On the DeepSWE v1.1 benchmark, it ties Astra, while on OSWorld 2.0, it trails by 2.1 points. It also outperforms Anthropic's Opus 5.5 in document processing and business workflow tasks.
Sol demonstrates improved reliability and safety compared to its predecessor. The rate of incorrect answers on difficult prompts dropped from 11.4 to 7.7 percent. Furthermore, the model shows a reduced rate of bypassing safety blocks, falling to 23.5 percent compared to 64.4 percent for the previous version.
OpenAI plans to continue developing the base Astra model through further reinforcement learning. While Sol is currently available, an Ultrafast version capable of generating tokens up to eight times faster in Codex is expected to launch in the coming days.
