OpenAI announced GPT-6.1 Sol in Codex and ChatGPT Work on September 29, 2026. The company positions it as a model for complex coding, computer use and professional work, promising performance close to GPT-6 Astra at a lower cost than Astra. Its launch announcement highlights repeated, long-running tasks across code, applications and documents. [1]
For users, the central question is practical: can a more affordable model deliver a finished result with fewer corrections? That is where this release deserves attention—and where OpenAI’s positioning still needs to be tested against real work.
Where GPT-6.1 Sol fits in OpenAI’s lineup
OpenAI’s current guidance places Astra at the top for the most demanding work, GPT-6.1 Sol in the middle for a balance of capability and cost, and Luna at the efficient end for focused, high-volume tasks. The company recommends comparing Sol with Astra on your own workloads. [2]
This suggests a useful way to approach model selection: match the model to the difficulty of the task and the cost of an error. A short extraction job and a software change spanning several files do not need the same level of reasoning.
“Near-Astra” should therefore be read as OpenAI’s product positioning, not a guarantee of equivalent results on every task. The documentation cited here does not establish a universal percentage improvement over GPT-6 Sol or an independently verified win over competing models.
What changes from GPT-6 Sol?
The newer model retains GPT-6 Sol’s standard API prices: $2 per million input tokens and $10 per million output tokens. Both list a 1,050,000-token context window and a maximum output of 128,000 tokens. These specifications indicate that the headline change is the claimed capability improvement, rather than a larger context window or a price cut from the previous Sol. [3] [4]
For developers, one compatibility change matters: GPT-6.1 Sol does not support the none reasoning setting. Its supported API settings run from low to max, with medium as the default. Tool calling requires the Responses API; Chat Completions is supported without tools. Existing integrations should check these differences before switching. [2]
What can it do in practice?
Potential use cases include investigating a software bug, preparing a document from several source files, or completing a workflow across connected applications. These are examples of how its stated focus could translate into daily work, rather than measured results from a DemystIA test.
The distinction matters because a successful workflow requires more than a persuasive answer. A coding assistant must understand the project, make the appropriate change and check the result. A document assistant must preserve source facts, follow the requested format and produce something usable.
For either task, a sensible evaluation asks: did the model finish the job, how much supervision did it need, and what remained to fix?
API pricing: affordable tokens do not tell the whole story
The model page lists the following standard text-token prices:
| Token type | Price per million tokens |
|---|---|
| Input | $2.00 |
| Cached input | $0.10 |
| Cache writes | $2.50 |
| Output | $10.00 |
Requests exceeding 272,000 input tokens carry higher rates for the full request. Tool fees and processing options can also affect the bill. These are API prices, not ChatGPT subscription prices. [3]
As a simple calculation, 100,000 uncached input tokens and 10,000 billed output tokens would cost $0.30 at those base rates, before any additional charges. This illustrates the tariff; it is not a typical task-cost estimate.
The better business metric is cost per accepted result. Retries, corrections and human review can outweigh a low token price. DemystIA explores this distinction in Claude Sonnet 5.5: Is It Really Cheaper Per Task?
Who can access GPT-6.1 Sol?
OpenAI lists Plus, Pro, Business, Enterprise and Edu in the launch rollout for Codex in the desktop app and CLI, and ChatGPT Work on the web and mobile. Enterprise and Edu administrators must enable access. Free and Go are not included at launch.
In ChatGPT, Sol is available through Work and Codex rather than Chat. Actual access depends on the client, rollout and workspace settings. Standard and Fast modes are available at launch, with Ultrafast support planned for later. [5]
What should users check before switching?
A useful comparison starts with a small set of familiar tasks: one coding change, one document task and one workflow requiring several steps. Give each model the same sources, instructions and acceptance criteria.
Record completion quality, elapsed time, corrections and total usage. Avoid changing the prompt after every disappointing answer: that makes it harder to distinguish model performance from prompt improvements.
For teams, the potential benefit is a more economical default for recurring complex work, while reserving a more capable model for difficult cases. That is an interpretation of the release’s positioning, not a demonstrated saving for every organisation.
A release to judge by completed work
GPT-6.1 Sol’s proposition is clear: bring more of Astra’s capability into a lower-cost model while retaining the previous Sol’s base API tariff.
Its value will depend on whether that promise translates into reliable finished work. For developers and professional users, the most revealing test is a task they already understand well—with a result they can independently check.