OpenAI has officially launched the GPT-6.1 Sol model, positioning it as the most cost-effective solution in its performance class. The model delivers capabilities comparable to the flagship GPT-6 Astra while reducing operational costs to one-fifth of the price for standard input and output tokens.
In technical benchmarks, GPT-6.1 Sol demonstrates significant efficiency. For instance, in OSWorld 2.0 tasks, the model achieves results within 2.1 percentage points of Astra at one-seventh of the cost. In scientific terminal tasks, the average cost per task is $5.47, compared to $23.21 for Opus 5.5 and $23.80 for Astra.
The new model also shows improvements in reliability and alignment. The factual error rate at low inference intensity has decreased from 11.4% in GPT-6 Sol to 7.7%. Furthermore, the model exhibits better adherence to safety constraints and user instructions, with error rates consistently within 1.9% of the Astra flagship across all tested settings.
Access to GPT-6.1 Sol is available starting September 29 for ChatGPT Work and Codex users across Plus, Pro, Business, Enterprise, and Edu tiers. Developers can integrate the model into their applications via the API using the gpt-6.1-sol identifier.
