OpenAI's new model costs a fifth as much—and tests put it one point from the flagship

A near-flagship with a lower API bill

OpenAI introduced GPT-6.1 Sol at its September 29 developer conference, a week after GPT-6 Sol. The company lists standard prices of $2 per million input tokens and $10 per million output tokens—one-fifth of GPT-6 Astra's rates—and $0.10 per million cached input tokens. openai +1 It is available through the API and to paid plans in ChatGPT Work and Codex, but not in ChatGPT's regular chat interface. openai +1

The release follows a reshuffle. OpenAI said it had decided not to release GPT-6.1 Astra after it failed the company's safety standards; the new Sol model is a separate release, not that shelved model. techcrunch +1

An outside score narrows the gap

OpenAI says Sol matches Astra on its DeepSWE v1.1 software-engineering benchmark at roughly one-fifth the cost, and approaches Astra on computer-use and professional-work tests. Those are company-reported comparisons, not a guarantee that the models behave alike in every application. openai

Artificial Analysis, which ran its own evaluations, placed Sol one point below Astra on its Intelligence Index. At maximum reasoning effort, it estimated $0.72 per task for Sol versus $3.26 for Astra on that index—a large cost difference on the evaluator's workload, not a universal price for completing software projects. artificialanalysis

Cheaper tokens do not settle the bill

Sol's cached-input rate is half the preceding GPT-6 Sol's, a discount relevant to agents that reuse long prompts. But Artificial Analysis also found that Sol used about 10% to 30% more output tokens than its predecessor across reasoning settings. Total cost therefore depends on how much a job makes the model generate, as well as the posted rates. openai +1

That distinction matters as OpenAI promotes software agents. Forbes' analysis of the company's Dots agents noted that Sol's lower API prices are not the price of running Dots, which launched powered by Astra. Model charges alone do not show what running an agent costs or how much human review its work needs. forbes

A narrower choice, not a replacement

OpenAI says Astra remains its highest-scoring model on the hardest scientific research tasks. It also says Sol's factuality evaluation used deliberately difficult prompts where users had previously flagged errors, and cautions that these do not represent typical use. The launch therefore presents Sol as a cheaper option for many demanding jobs, not proof that the flagship is redundant. openai

For developers, the practical next step is to test Sol on their own workloads: whether the lower token bill outweighs any extra reasoning, output or review needed to reach a reliable result remains task-specific. artificialanalysis +1