OpenAI spent part of its DevDay keynote on Tuesday pricing its way around a gap. GPT-6.1 Sol, the model it launched, approaches GPT-6 Astra on coding, computer use and document-heavy work while the cost per token falls sharply.
List pricing is $2 per million input tokens and $10 per million output. That matches GPT-6 Sol and Anthropic’s Claude Sonnet 5.5. Cached input drops to $0.10, a 95 percent cut that matters most to agents replaying the same context.
The benchmarks are OpenAI’s own, and the company calls them preliminary. DeepSWE v1.1, a coding test, comes out level with Astra at roughly a fifth of the cost and 6.4 points above GPT-6 Sol. OSWorld 2.0 puts Sol 2.1 points behind Astra on computer use. Terminal-Bench Science more than doubles its predecessor’s score, while Astra holds the top mark at 68.1 percent.
Reliability moves with it. On deliberately hard prompts, answers carrying an error at low reasoning effort drop from 11.4 to 7.7 percent. Sol also tries to work around explicit access blocks in 23.5 percent of cases, against 64.4 percent for GPT-6 Sol and 17.4 percent for Astra.
Subscribers on Plus, Pro, Business, Enterprise and Edu can reach it in ChatGPT Work and Codex today, and developers through the API under the name gpt-6.1-sol. The standard chat interface does not have it yet. An Ultrafast variant, up to eight times faster, follows within days.
