OpenAI Releases GPT-6.1 Sol: Cost-Effectiveness Reshapes Agent Development

Overview

OpenAI has officially launched GPT-6.1 Sol, an upgraded version of GPT-6 Sol, which excels particularly in agent programming, computer use, and professional work domains. Its core selling point is delivering frontier performance close to that of GPT-6 Astra at approximately one-fifth of the price, aiming to create more room for developers to build higher-capability agents within the same budget.

GPT-6.1 Sol significantly enhances the ability to handle complex tasks while maintaining high cost-effectiveness, making it one of the most cost-efficient models available at its current performance level.

Performance Breakthroughs and Evaluation Results

Compared to its predecessor, GPT-6.1 Sol achieves dual optimization in both performance and cost across a range of high-difficulty benchmark tests. In the programming domain, DeepSWE v1.1 evaluations show that the model reaches parity with GPT-6 Astra at roughly one-fifth of the cost, surpassing the highest score of GPT-6 Sol by 6.4 percentage points.

In professional work applications, the model's ability to answer questions using complex PDF documents surpasses that of Opus 5.5 with fallback mechanisms, at less than half the cost per task. Meanwhile, on AutomationBench—which measures multi-step business workflows—its score at medium reasoning intensity exceeds that of Opus 5.5 by 2.2 percentage points, while costing only one-third as much.

Progress in computer use and scientific research is equally significant. In the OSWorld 2.0 offline set evaluation, the gap in scores between GPT-6.1 Sol and Astra narrows to within 2.1 percentage points at maximum reasoning intensity, with a cost per task of approximately one-seventh of that of Astra. In Terminal-Bench Science 0.1, its average cost per task is only $5.47, representing a cost reduction of more than 75% compared to both Opus 5.5 and Astra.

Safety and Factuality

Regarding safety alignment, GPT-6.1 Sol shows marked improvement in alignment evaluations compared to GPT-6 Sol, drawing closer to the level of GPT-6 Astra. The model is more reliable in respecting user intent and truthfully reporting tool failures in high-difficulty scenarios rather than providing speculative answers. Data indicates that it failed to inform users of search tool failures in only 2.8% of cases, far lower than the 4.9% observed for GPT-6 Sol.

Additionally, the model's factual error rate under high-difficulty prompts drops to 4.1%, lower than the 4.5% seen with GPT-6 Sol, while its cost per task is approximately 83% lower than that of Astra, further lowering the barrier to use while ensuring output accuracy.

Pricing and Availability

Starting today, all Plus, Pro, Business, Enterprise, and Edu users can access GPT-6.1 Sol in ChatGPT Work and Codex. Developers can invoke the model via the API using gpt-6.1-sol, with standard API pricing set at $2 per million input tokens, $0.10 per million cached input tokens, and $10 per million output tokens. Cached input pricing is 95% cheaper than standard input, making it ideal for agent workloads that require context reuse.

Meanwhile, OpenAI has also launched GPT-6 Astra Ultrafast and GPT-6.1 Sol Ultrafast. The latter allows developers to pay approximately the same API fee as using GPT-6 Astra to achieve intelligence levels close to Astra's, with speeds up to 8 times faster. These ultrafast models are now available via the API and open to users on the new Pro tier ($500 per month) in ChatGPT Work and Codex.