If you open the usual ChatGPT chat model picker looking for GPT-6 Sol, you may be looking in the wrong place. OpenAI's September rollout puts Sol in ChatGPT Work and Codex, while regular Chat has a separate model selection. That distinction determines where you start a task and which picker you need to check.
The difference is practical. OpenAI's guide to Work suggests using Chat for a quick explanation, Work to produce a report or analysis you can review, and Codex to investigate a bug and prepare a code change. A person can have access to OpenAI's products and still not see Sol in an ordinary Chat conversation, because that is not where the September rollout places it. To try the model, open Work or Codex and check its model picker. If Sol is absent there too, rollout timing or your workspace's model settings may explain why.
GPT-6 Sol began rolling out on 22 September 2026. OpenAI recommends it for complex coding and agent workflows. This article separates the Work and Codex access route from the API, then shows how to test Sol on a task with a result you can verify.
Where is GPT-6 Sol available?
OpenAI's launch notice says GPT-6 Sol and the smaller GPT-6 Luna are rolling out to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex. Availability depends on the rollout and workspace settings. Enterprise administrators must enable the new models. Free and Go users can access Luna in the desktop app, but the notice does not say they receive Sol.
| Surface | GPT-6 Sol status at launch | What to check |
|---|---|---|
| ChatGPT Work | Rolling out to eligible plans | Look in Work's model picker; workspace settings can affect access. |
| Codex | Rolling out to eligible plans | Select it for a task when it appears; existing manually selected models remain selected. |
| Regular Chat | Not part of this rollout | Do not describe Sol as the new default Chat model. |
| OpenAI API | Available as gpt-6-sol |
API usage and billing are separate from plan access. |
The first three rows reflect OpenAI Docs' 22 September rollout note; the API model ID comes from the developer model page. Treat the rollout status as dated, because access can change faster than a launch article.
The distinction between Work, Codex and Chat is easy to miss. This is a launch in the work and coding surfaces, not an announcement that Sol has replaced the model used by an ordinary Chat conversation. A team writing an internal guide should name the surface it means and check what its own workspace exposes.
Developers can also use the API model ID gpt-6-sol, according to the OpenAI API model page. API availability and billing are a separate route from access through a ChatGPT plan.
What is GPT-6 Sol designed to do?
OpenAI's launch guidance describes Sol as built for complex coding and agentic workflows, and positions Luna for focused, high-volume tasks. The GPT-6 Sol model page gives developers the technical details. Together, these are a starting point for choosing a model, not a rule that Sol wins every task.
OpenAI's model-selection guidance puts Sol between Luna and Astra. In practical terms:
| Model | OpenAI's positioning | A sensible first trial |
|---|---|---|
| GPT-6 Luna | Efficient, frequent, well-scoped work | Extract fields from a repeatable form or make a small, defined edit. |
| GPT-6 Sol | Everyday work that needs judgement, including coding and research | Investigate a bug, make a multi-file patch and explain the evidence. |
| GPT-6 Astra | Ambiguous problems and demanding deliverables | Plan a complex project or resolve a task whose requirements are uncertain. |
These are starting points for testing, not results from a head-to-head evaluation. The model, reasoning effort and surrounding tools all affect the outcome. A Luna task can become a Sol task when its edge cases demand judgement; a Sol task can justify Astra when the stakes and ambiguity rise.
Think of a coding task that begins with an unclear bug, requires reading several files, changes more than one component and ends with tests and a reviewable patch. That is a reasonable Sol trial. A short, repeatable extraction or formatting job may be a better candidate for a faster, cheaper option. The result depends on the task and the tools around the model.
The API model page lists a 1,050,000-token context window, a 128,000-token maximum output, text and image input, and text output. It lists reasoning effort from none through max. These are technical capabilities for integration planning. A large context window does not prove the model will find the right detail in a long repository or document set. That still needs a practical test.
For API developers, the distinction between endpoints matters. The GPT-6 Sol API documentation recommends the Responses API for built-in tools and function calling with Sol. It says Chat Completions supports function calling only when reasoning effort is set to none. A team moving an existing tool-based workflow should check that constraint before switching model IDs.
What does GPT-6 Sol cost?
For the API, OpenAI lists the following standard, short-context US-dollar prices per million tokens on its GPT-6 Sol model page:
| API usage | GPT-6 Sol list price |
|---|---|
| Input | US$2.00 |
| Cached input | US$0.20 |
| Cache writes | US$2.50 |
| Output | US$10.00 |
For an illustrative request with 100,000 input tokens and 25,000 output tokens, the uncached token charge would be US$0.45 at those rates. That excludes tools, retries, taxes and any other charges. It is an arithmetic example, not a typical customer bill.
Long context has an important exception: the API model documentation says prompts above 272,000 input tokens are charged at twice the input and cache rates and 1.5 times the output rate for the full request. Batch, Flex and Fast processing have different rates. Check the live pricing page before budgeting. API list prices are separate from ChatGPT subscription limits and Codex credits.
OpenAI says Sol and Luna have lower token prices than their GPT-5.6 predecessors. A lower rate does not automatically make an agent task cheaper. More attempts, longer context or additional tool calls can erase the saving. The measure that matters is the cost of an accepted result, alongside the review work needed to trust it.
GPT-6 Sol in Codex: what changes for existing users?
For users signed in with ChatGPT, Sol can be chosen in the model picker as the rollout reaches their account. In the Codex command line, OpenAI Docs shows the gpt-6-sol model ID. The model a person selected manually in Codex desktop remains selected until they change it, according to the release notes. A team should therefore check saved defaults, project instructions and recurring tasks individually rather than assuming every task moved to Sol on launch day.
OpenAI's model guide says GPT-5.5 will retire from ChatGPT, Work and Codex with ChatGPT sign-in on 14 October 2026, while remaining available through the API. For eligible Plus, Pro, Business, Enterprise and Edu users, it points to Sol as a replacement when available. This is a separate, upcoming migration decision; it is not evidence that Sol automatically reproduces every GPT-5.5 result.
How should a team decide whether to use Sol?
Run one matched trial before changing the team's default. Choose a task with a clear result, such as repairing a failing test in a small feature or drafting a report from a defined set of documents. Give the current model and Sol the same starting material, tools, permissions and acceptance criteria. Record the model used, total usage, time to completion, corrections and reviewer effort.
| Record for each run | Why it matters |
|---|---|
| Model and reasoning effort | A higher effort setting can change both output quality and usage. |
| Same task and starting state | Different inputs make the comparison hard to interpret. |
| Tools and permissions | A model with broader tool access has a different job. |
| Accepted result and tests | Completion is more meaningful than a confident-looking draft. |
| Total usage, elapsed time and reviewer corrections | This is the practical cost of getting usable work. |
For a coding trial, a useful acceptance rule might be: the patch fixes a reproducible bug, passes the relevant tests, stays within the permitted files and needs no undisclosed manual repair. Run the task more than once and include a second task with a different shape. A single favourable result should not become a universal model ranking.
This is a suggested method, not a claim that we tested the models head to head. Our AI evals guide explains how to build a small set of representative tasks. If the work uses tools, approvals and memory, the agent harness guide explains why the model name alone does not determine the outcome.
There is also a product decision. A developer evaluating Sol in Codex should assess the whole coding workflow: where the task runs, which commands it can use, how it reports changes and how a person reviews them. Our Codex versus Claude Code guide covers that wider choice. Its product details should be checked against the new rollout before treating an older comparison as current.
GPT-6 Sol frequently asked questions
Is GPT-6 Sol available in regular ChatGPT chat?
No. OpenAI Docs says the September 2026 Sol rollout is for ChatGPT Work and Codex, and that those models are separate from the models in Chat. Check the Work or Codex model picker for your account.
Can ChatGPT Plus users access GPT-6 Sol?
OpenAI's rollout notice says Sol is coming to Plus, Pro, Business, Enterprise and Edu users in Work and Codex. Availability still depends on rollout and workspace settings. Enterprise administrators must enable the new models.
Does a ChatGPT subscription include GPT-6 Sol API usage?
No. API-key usage is charged at API rates, separate from usage included with a ChatGPT plan, according to OpenAI Docs pricing. The API model page lists the rates for gpt-6-sol. Do not use those rates to estimate what a ChatGPT plan or Codex task will cost.
Should I choose GPT-6 Sol or GPT-6 Luna?
OpenAI's model-selection guidance positions Sol for work needing more judgement, including coding and multi-step tasks, and Luna for focused, frequent work. Start with the cheaper or faster option that meets your acceptance test, then move up when the result needs more reasoning or review. This is a testing approach, not a universal ranking.
What remains to be seen after launch?
The rollout is still subject to plan and workspace controls. Pricing, usage limits and model-picker availability can change. OpenAI's use-case guidance is a useful starting point, but independent testing and a team's own outcomes will determine whether Sol is the right default for its work.
For now, the decision is straightforward: look for GPT-6 Sol in Work or Codex, not ordinary Chat; choose one consequential but bounded task; and compare the reviewed result with what your current model produces. The launch gives teams another option. The evidence for switching comes from their own accepted work.
