OpenAI introduces GPT-6.1 Sol at a lower cost than Astra

OpenAI has introduced GPT-6.1 Sol, a model designed to approach the performance of GPT-6 Astra in programming, computer use and professional tasks, but at a much lower cost. According to the company, it can deliver similar results on several tests for approximately one-fifth the price of Astra.

The difference matters most for people building AI agents: systems that can read information, make decisions and complete multiple steps on their own. The more these agents interact with documents, applications or code, the more tokens they use. Tokens are the units of text a model processes to understand a request and generate a response.

More capacity for complex tasks

GPT-6.1 Sol is an update to GPT-6 Sol and improves performance on work that goes beyond answering simple questions. OpenAI highlights advances in:

• Writing and debugging code.
• Understanding documents with tables, charts and fine print.
• Multi-step business workflows.
• Using computer applications.
• Data analysis, simulations and mathematical demonstrations.

In DeepSWE v1.1, a test that measures the ability to solve software engineering problems in real codebases, Sol matched Astra at an approximate five-times-lower cost. It also outperformed GPT-6 Sol by 6.4 percentage points, using a lower reasoning level and at a lower cost.

In GDP.pdf, an evaluation of professional questions based on complex documents, the model scored higher than Opus 5.5 with backup systems, at less than half the cost per task. Its result was close to Astra's, but at around one-fifth the price.

The difference also appears in business tasks. In AutomationBench, which checks whether an agent can complete multi-step processes, Sol outperformed Opus 5.5 by 2.2 percentage points at a medium reasoning level, for approximately one-third of the cost. Compared with GPT-6 Sol, it improved by 4.8 points under the same conditions.

Better computer use and fewer errors

In OSWorld 2.0, a test of demanding tasks inside computer applications, GPT-6.1 Sol outperformed GPT-6 Sol by 7 percentage points at the highest reasoning level. It also came within 2.1 points of Astra, at an approximate seven-times-lower cost per task.

In Terminal-Bench Science 0.1, which focuses on scientific workflows such as data analysis, simulations and mathematical demonstrations, Sol more than doubled the score of GPT-6 Sol. Each task cost an average of $5.47, compared with $23.21 for Opus 5.5 and $23.80 for Astra. Astra did retain the highest score among the models evaluated, at 68.1%, and remains the recommended option for the most difficult scientific research.

The new model also reduced factual errors in an evaluation made up of especially challenging conversations in which users had flagged previous mistakes. At a low reasoning level, the share of responses containing at least one error fell from 11.4% to 7.7%, an approximate reduction of 32%.

OpenAI clarifies that these questions were designed to be difficult and do not represent typical use. Across the different configurations tested, the error rate was no more than 1.9 points away from Astra's, at less than one-fifth the cost per task.

Price and availability

The cost reduction is one of Sol's main attractions. In the API, the model costs:

• $2 per million input tokens.
• $0.10 per million reused input tokens.
• $10 per million output tokens.

The price of reused input, known as cached input, is 95% lower than standard input and 50% lower than GPT-6 Sol input. This allows an agent to retain context between requests without paying the full price again for the same information.

GPT-6.1 Sol is already available to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex. It is not yet offered in ChatGPT. Developers can access it through the API under the name gpt-6.1-sol.

OpenAI has also announced a GPT-6.1 Sol Ultrafast version, expected in the coming days, which will be able to generate tokens up to eight times faster in Codex. For users, the most important change is not that Astra is going away, but that advanced tasks such as programming, reviewing documents or running complete processes may become viable at a much lower price. The next test will be whether that cost advantage holds up in daily use, not just in controlled evaluations.