OpenAI introduces GPT-6.1 Sol model

OpenAI has introduced GPT-6.1 Sol, an updated model in the GPT-6 family designed for complex professional tasks. The model approaches the capabilities of the flagship GPT-6 Astra, but costs 5 times less at API.

Capabilities of GPT-6.1 Sol

GPT-6.1 Sol has become a notable upgrade to GPT-6 Sol in programming, document processing, computer control, and executing multi-step business tasks. At the same time, OpenAI positions Astra as a more capable model for the most complex scenarios.

Factual accuracy has also improved. In OpenAI’s test on complex queries with a low level of reasoning, the share of responses with at least one factual error decreased from 11.4% for GPT-6 Sol to 7.7% for GPT-6.1 Sol, a reduction of approximately 32%. The company warns that the test specifically used queries designed to provoke errors, so these figures do not reflect the frequency of hallucinations during normal use.

OpenAI also reports that GPT-6.1 Sol adheres better to explicitly defined constraints, performs unauthorized actions less frequently in agentic scenarios, and provides more accurate feedback to users regarding issues with the tools being used.

Test results

In DeepSWE v1.1, which evaluates the performance of AI agents on real-world codebases, GPT-6.1 Sol performed on par with GPT-6 Astra at approximately one-fifth of the task execution cost. Compared to the best result from GPT-6 Sol, the advantage was 6.4 percentage points at a lower cost and with a lower level of reasoning.

In AutomationBench 1.0.6, the model outperformed Claude Opus 5.5 by 2.2 percentage points at a mid-level reasoning tier, while costing approximately three times less to execute the task. The new version pulled ahead of GPT-6 Sol by 4.8 percentage points under the same settings.

In GDP.pdf, where models answer questions about complex professional PDF documents containing tables, charts, and fine details, GPT-6.1 Sol outperformed Claude Opus 5.5 with a fallback mechanism at less than half the cost of completing the task. The model also approached GPT-6 Astra, while costing approximately five times less.

Computer control has also improved. In the OSWorld 2.0 offline suite, at the maximum reasoning level, GPT-6.1 Sol outperformed GPT-6 Sol by 7 percentage points and trailed GPT-6 Astra by 2.1 percentage points, while costing approximately seven times less to execute the task.

At the same time, Astra maintains an advantage in the most complex scenarios. For example, in Terminal-Bench Science 0.1, it showed the best result among the tested models at 68.1%.

Ultrafast

Along with the new tools, OpenAI introduced a premium Ultrafast mode for accelerated token output. It is currently available for GPT-6 Astra and provides up to an eightfold speed increase in Codex and up to a sixfold increase via API, depending on the scenario.

Support for GPT-6.1 Sol in Ultrafast has not yet been launched. OpenAI promises to add it in the coming days. For Sol, the company claims generation speeds up to eight times faster compared to the standard mode in Codex.

Availability and pricing

GPT-6.1 Sol is now available to developers via OpenAI API under the name gpt-6.1-sol. Users with Plus, Pro, Business, Enterprise, and Edu subscriptions have gained access to the model in ChatGPT Work and Codex. The model is not yet available in standard Chat mode.

GitHub has also begun a phased rollout of GPT-6.1 Sol in Copilot for Pro+, Max, Business, and Enterprise subscribers. The model can be used in Visual Studio Code, Visual Studio, Copilot CLI, the GitHub Copilot coding agent, the Copilot app, on github.com, in the GitHub mobile app, JetBrains IDE, Xcode, and Eclipse.

OpenAI’s standard rates API are:

• 1 million input tokens – $2;
• 1 million cached input tokens – $0,10;
• 1 million output tokens – $10.

Thus, standard input and output tokens for GPT-6.1 Sol are five times cheaper than those for GPT-6 Astra, and cached input data is 50% cheaper than for GPT-6 Sol.