OpenAI announced GPT-6.1 Sol, an upgraded version of GPT-6 Sol targeting tasks that combine high performance with low operating costs. The company says the model approaches GPT-6 Astra in coding, computer use, document analysis, and professional multi-step workflows, while its standard prices are approximately one-fifth of Astra’s prices.
Pricing Designed for Broad-Scale Use
GPT-6.1 Sol API costs $2 per million input tokens and $10 per million output tokens. Cached tokens cost $0.10 per million tokens, representing a 95% discount compared with the standard input price, according to OpenAI. This makes the model particularly suitable for agentic applications that reuse the same context across repeated requests.
Practical Test Results
In the DeepSWE v1.1 test, which measures long software-engineering tasks on real codebases, GPT-6.1 Sol achieved the same result as GPT-6 Astra, but at approximately one-fifth the cost. In the GDP.pdf test for professional document analysis, its performance approached Astra’s on tasks involving tables, charts, and PDF files in fields such as finance, healthcare, and law.
In the AutomationBench test for multi-step workflows, the model outperformed Opus 5.5 by 2.2 points and GPT-6 Sol by 4.8 points at the medium reasoning level. In the OSWorld 2.0 computer-use test, however, it scored 2.1 points below Astra, with a per-task cost of approximately one-seventh of Astra’s cost and less than half the cost of GPT-6 Sol in the stated configuration.
In Terminal-Bench Science 0.1, which measures data analysis, simulation, and theorem proving, the average cost per task was $5.47, compared with $23.21 for Opus 5.5 and $23.80 for GPT-6 Astra. Nevertheless, Astra maintained the highest performance in the test, with a score of 68.1%.
Accuracy and Security Limitations
GPT-6.1 Sol reduced the factual-error rate to 4.1%, compared with 4.5% for GPT-6 Sol and 4% for GPT-6 Astra at the highest reasoning level. OpenAI cautions that this evaluation is based on difficult prompts with a high likelihood of producing errors and does not represent typical ChatGPT use.
The company also says the model has improved at informing users about nonfunctional tools, avoiding disallowed outcomes, and following explicit constraints. It also stated that, like GPT-6 Astra and GPT-6 Sol, it did not attempt to bypass the automated security control unit. The source provides no independent details about the methodology used to verify these claims.
What Changes in Practice?
OpenAI presents GPT-6.1 Sol as an option for high-volume applications, where reducing token costs may matter more than achieving the highest possible score on every test. This matters to developers of software agents, document-analysis services, and automation, but the performance difference remains tied to the type of task and the reasoning level, so the lower cost does not mean it is a comprehensive replacement for Astra.
The model is available to Plus, Pro, Business, Enterprise, and Edu subscribers through ChatGPT Work and Codex, and developers can use it in the OpenAI API under the name gpt-6.1-sol. It has not yet been released in the standard ChatGPT chat interface. OpenAI also announced GPT-6.1 Sol Ultrafast and GPT-6 Astra Ultrafast, claiming speeds of up to eight times Astra’s, while it has not launched GPT-6.1 Astra, which was expected to be released this week, amid reports of security concerns in internal testing.