The release of GPT-6 Sol and this update
OpenAI released the previous model, GPT-6 Sol, on Sept. 22, 2026. Sol is a model designed for developing code and carrying out work using multiple tools.
GPT-6.1 Sol is an updated version that arrived a week later. OpenAI positions it as a model that offers performance close to the higher-end GPT-6 Astra at a lower price than Astra. It is intended for repeatedly delegating work involving code, apps and documents.
Improvements in code fixes and document reading
OpenAI reported that scores improved for investigating causes and making changes when fixing code, and for carrying out multiple steps while looking at the screen when operating apps. It also said reading of documents containing tables and figures in PDFs has improved.
The published evaluation results show the following differences from GPT-6 Sol. The reasoning setting used in the comparisons specifies how much the model thinks, and the conditions differ by evaluation.
| Task evaluated | Evaluation name | Difference from GPT-6 Sol | Comparison conditions |
|---|---|---|---|
| Development work on real code | DeepSWE 1.1 | 6.4 points higher | Compared with the previous model's best score. 6.1 Sol's reasoning setting is lower |
| Business procedures using multiple tools | AutomationBench 1.0.6 | 4.8 points higher | Both models at medium |
| Operating app screens | OSWorld 2.0 | 7 points higher | Both models at max. Offline evaluation that also scores partial completion |
Medium and max are names of reasoning settings. OSWorld used the Aug. 8, 2026, version of the evaluation set. These are scores measured in OpenAI's research environment and API, and results in Work and Codex also depend on the instructions, available apps and settings.
Eligible plans and where to use it
Among individual plans, Plus and Pro are eligible. For company and school contracts, administrator settings and user permissions also matter in addition to the plan.
| ChatGPT plan | Can use GPT-6.1 Sol | Where to use it |
|---|---|---|
| Plus | Yes | ChatGPT Work, Codex |
| Pro | Yes | ChatGPT Work, Codex |
| Business | Yes. Check company contract and user permissions | ChatGPT Work, Codex with permission |
| Enterprise, Edu | Yes once an administrator enables the model | ChatGPT Work, Codex with permission |
| Free, Go | Not eligible at launch | GPT-6.1 Sol not offered |
Work is a feature for delegating document creation and work using apps, and Codex is a development agent for delegating the writing and fixing of code. Users who normally converse in Chat open Work or Codex to use it.
According to the launch guidance, Work is available on the web and mobile, and Codex in the desktop app and the CLI, which is operated with commands. When the model appears as an option depends on the rollout and app support. It is also offered to developers through the OpenAI API, with contracts and billing separate from ChatGPT.
Usage guidelines and Fast consumption
The amount available also differs by plan and contract type. The counts given in the official guide are guidelines, and how much of the allowance is consumed varies with document length, task complexity and reasoning settings.
| Plan and contract type | Usage guideline |
|---|---|
| Plus | 15 to 160 local messages per 5 hours |
| Pro | Currently no 5-hour limit. Weekly limits may apply |
| Business Standard seat | 15 to 160 local messages per 5 hours |
| Business usage-based | Follows the credit system and administrator settings |
| Enterprise, Edu | Varies by contract. Options include a credit system or applying the same limits as Plus for most features |
The number of local messages is a guideline for messages used in work such as handling files and code on a PC. Work in the cloud consumes the same allowance, and weekly limits may also apply on Plus and Standard Business.
At launch, the speed settings are Standard and Fast. Choosing Fast consumes the allowance included in the plan at 2.5 times the rate of Standard for the same model. For purchased credits and Enterprise usage-based billing, the rate is 2 times. These are consumption multipliers, and actual wait times vary with the task. Ultrafast for GPT-6.1 Sol is planned for the future.
How to choose GPT-6.1 Sol
In the desktop app and the web version of Work, open the model settings below the input box and choose GPT-6.1 Sol under Advanced. Advanced is a detailed setting for individually choosing the model, reasoning setting and speed.
In Codex CLI, users can switch with the "/model" command. The available models and settings vary with the contract plan, the app used and administrator settings.
