OpenAI released GPT-6.1 Sol on its website, a cheaper model that the company says trails its top-tier GPT-6 Astra only slightly when agents write code, operate software, or handle office work. Its pitch is price: Sol costs one-fifth of what Astra does per token, according to OpenAI’s announcement. Every performance figure below is OpenAI’s own reporting.

Through the API, where the model goes by gpt-6.1-sol, developers pay $2 per million tokens sent in and $10 per million tokens generated. Cached input, meaning context an agent reuses between requests, runs $0.10 per million tokens. OpenAI says that is 95 percent below the standard input rate and half what the previous GPT-6 Sol charged for cached input. Agents that re-read the same long context over and over are where that discount bites hardest.

On coding, OpenAI says Sol matches Astra on DeepSWE v1.1, a test built on real software projects, at roughly a fifth of the cost. The company adds that Sol outscores the earlier GPT-6 Sol’s top result by 6.4 percentage points while using less reasoning effort. On OSWorld 2.0, which tests long computer-use tasks, OpenAI reports that Sol finishes 2.1 percentage points behind Astra while spending about a seventh as much on each task.

The rival comparisons are more pointed. On GDP.pdf, a test of questions about dense professional documents in fields such as finance, healthcare, and law, OpenAI says Sol beats Anthropic’s Opus 5.5 at under half the cost per task. On AutomationBench, which grades multi-step business workflows, OpenAI says Sol outscores Opus 5.5 by 2.2 points in the medium reasoning setting, for roughly a third of the price. OpenAI also flags that the figure for Claude Fable 5.1 understates that model’s true cost, because it leaves out fallback runs that happened on about 40 percent of tasks.

Science work shows the widest gap. OpenAI says Sol costs an average of $5.47 per task on Terminal-Bench Science 0.1 at maximum effort, against $23.21 for Opus 5.5 and $23.80 for Astra. The company concedes Astra still posts the highest score there, at 68.1 percent, and says Astra remains the model for the hardest research problems. That is a notable admission from a vendor selling the cheaper option.

On accuracy, OpenAI says Sol trimmed the share of answers with a factual error from 11.4 percent to 7.7 percent at low reasoning effort, compared with GPT-6 Sol. The test uses conversations where users had flagged an earlier model’s mistake, and OpenAI itself says those prompts do not reflect typical use. The company also describes better behavior in its own alignment tests, including honesty about broken search tools, and says its testers saw Sol make zero tries at getting around an automated safety reviewer. Those are self-reported results, with details in a system card addendum OpenAI published.

Availability is broad but not complete. Sol is live for Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and in Codex, though it has not reached the standard Chat experience. An Ultrafast option, with up to 8x faster generation in Codex, is promised in the coming days, with no price given in the announcement.

The market angle is what matters. Frontier labs have long sold their best model at a steep premium and left buyers to guess how much of it they needed. OpenAI is now pricing a near-flagship model so low that Astra has to justify itself only on the hardest tasks, and it is doing so with head-to-head charts against Anthropic. No independent evaluator has yet confirmed any of these numbers.

Teams running agents on Astra or Opus 5.5 should replay a slice of real traffic through Sol this month, since a fivefold price gap only matters if quality holds on your own workload.

Reported by OpenAI in its announcement of GPT-6.1 Sol, which carries no publication date.