OpenAI Launches GPT-6.1 Sol: High-Performance Agentic Programming at One-Fifth the Price of Astra

Overview

OpenAI has officially launched GPT-6.1 Sol, an upgraded version of GPT-6 Sol, specifically optimized for agent programming, computer use, and professional workflows. At one-fifth the price of GPT-6 Astra, this model delivers intelligence levels approaching its frontier counterpart, aiming to provide developers with a better cost-effectiveness balance when building powerful AI agents.

GPT-6.1 Sol brings Sol’s performance on high-difficulty tasks closer to that of GPT-6 Astra, while the standard input and output token prices are only one-fifth of Astra’s.

Performance Breakthrough: Closing in with State-of-the-Art

In multiple benchmark tests, GPT-6.1 Sol demonstrated significant performance leaps, particularly in coding, complex document processing, and multi-step business automation.

• Programming capability: In the DeepSWE v1.1 evaluation, the model’s ability to complete complex software engineering tasks in real codebases was greatly enhanced. At approximately one-fifth the cost, it reached parity with GPT-6 Astra, outperforming the previous generation Sol’s highest score by 6.4 percentage points.
• Professional work: In the GDP.pdf test, when faced with complex documents containing tables, charts, and fine print details, it scored higher than Opus 5.5 with fallback mechanisms, at less than half the cost per task.
• Automation and research: AutomationBench showed its score at medium reasoning intensity was 2.2 percentage points higher than Opus 5.5. In Terminal-Bench Science 0.1, its score exceeded that of GPT-6 Sol by more than double, with a per-task cost lower than both Opus 5.5 and Astra by over 75%.

Although Astra still ranks first in scientific research tasks with a 68.1% score, GPT-6.1 Sol offers a more cost-effective choice for most routine high-precision needs.

Safety and Cost Control

Regarding safety, GPT-6.1 Sol showed significant improvement over GPT-6 Sol in alignment evaluations, being more inclined to candidly acknowledge its own limitations and strictly adhere to safety constraints. In high-difficulty scenario tests, its error rate was lower than GPT-6 Sol’s, with no observed attempts to bypass automated safety reviews, performing consistently with GPT-6 Astra.

In terms of cost control, the model supports cached input optimization: cached inputs cost only $0.10 per million tokens, which is 95% cheaper than the standard input price, significantly reducing the cost of agent workloads that need to reuse context across multiple requests. Additionally, at extremely high reasoning intensity, its factual error rate was 4.1%, close to Astra’s 4.0%, at only about 17% of the latter’s cost.

Pricing and Availability

Effective immediately, Plus, Pro, Business, Enterprise, and Edu users can use GPT-6.1 Sol in ChatGPT Work and Codex. Developers can access it via the API with the following standard pricing:

• Input Token: $2 per million
• Cached Input Token: $0.10 per million
• Output Token: $10 per million

Additionally, OpenAI has launched GPT-6 Astra Ultrafast and GPT-6.1 Sol Ultrafast. The latter achieves speeds up to 8 times that of GPT-6 Astra, allowing developers to obtain nearly the same high-performance intelligence and ultimate response speed for costs similar to using GPT-6 Astra. This model is now open via the API and available in ChatGPT for users on the new Pro tier ($500 per month).