Updated September 30, 2026
The first question about GPT-6.1 Sol is where to use it. The second is how to tell whether it completed your task correctly.
This guide answers both with two small exercises: reviewing fictional support requests in ChatGPT Work and building a tiny request-counting function in Codex. Each has inputs you can inspect and a clear acceptance checklist. These are proposed practice tasks, not benchmark results or accounts of tests we performed.
Where is GPT-6.1 Sol available?
OpenAI announced GPT-6.1 Sol for Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex. As of September 30, 2026, it is not yet available in Chat. The announcement also describes Sol Ultrafast for Codex as coming in the following days; do not assume it is already selectable. OpenAI’s Sol announcement
Open Work or Codex and check the selected model before beginning. If your account does not offer GPT-6.1 Sol, verify your plan, workspace settings and current official availability rather than assuming an ordinary Chat conversation is using it.
Developers can also use the model identifier gpt-6.1-sol through the API. The exercises below use Work and Codex, so they do not require an API integration. Official model documentation
First Work task: turn support notes into a useful brief
A fictional bike-repair shop wants to understand five incoming requests. Use invented records to learn the process before handling genuine customer information.
Copy this sample data:
R01 | Appointment | Wants a Saturday slot; no booking confirmed
R02 | Status | Asks when repair B17 will finish; completion date unknown
R03 | Appointment | Wants a Tuesday slot; no booking confirmed
R04 | Information | Asks whether the shop repairs folding bikes; policy not supplied
R05 | Status | Asks whether repair B18 is ready; status unknown
Now use this prompt:
Analyze these five fictional support records. Use only the supplied data. Count requests by category and calculate each category’s share of all records. Create a brief with the counts, three operational observations and a draft response for each record.
For missing information, write a specific question for the shop owner. Do not invent appointment availability, repair completion dates or services offered. Label each response as a draft. Do not send messages or update external systems. End with the calculations and a list of facts that require verification.
Keeping the dataset small makes checking faster. It also exposes a common problem: a fluent answer can sound helpful while quietly inventing business facts.
Check the Work result against an answer key
The categories should be Appointment: 2, Status: 2 and Information: 1. Their shares are 40%, 40% and 20%. The count must total five and the shares must total 100%.
The response drafts should not confirm Saturday or Tuesday availability. Neither repair has a supplied completion status. Folding-bike service remains an unanswered policy question.
Three reasonable observations would concern appointment demand, missing repair information and the need for a clear service list. They should remain limited to this five-record sample. It would be unjustified to call these the shop’s most common annual issues.
If the output fails, request a targeted correction:
Recheck every factual claim against the record IDs. Remove unsupported promises. Recalculate the category totals and return only the corrected sections plus an explanation of what changed.
Save both versions. The corrections teach you which instructions and checks matter for your work.
First Codex task: build the smallest testable helper
If you want to explore coding, turn the same example into a small function in a new practice folder. Keep it separate from an existing business application.
Use this prompt:
In this practice folder, create a Python function count_categories(records). Each record is a dictionary with id and category. Return a dictionary containing counts for each category present.
Reject duplicate IDs and empty categories with ValueError. An empty input list must return an empty dictionary. Use the Python standard library only. Add tests for the five-record example, empty input, duplicate IDs and an empty category. Run the tests if the environment supports Python. Report the exact command and observed result, or clearly say that execution was unavailable. Do not deploy, publish or modify files outside this practice folder.
The function has a deliberately narrow contract. That lets you inspect the implementation without first learning an entire framework.
Review code, execution and behavior separately
A sensible-looking function is only one part of the result. Review the changed files, then inspect the test command and its actual output.
The five-record test should return two Appointment records, two Status records and one Information record. Empty input should return an empty dictionary. Duplicate IDs and empty categories should raise the specified exception.
Ask whether the tests check behavior rather than merely copying the implementation’s assumptions. For example, does the duplicate-ID test use two different records that share one ID?
If tests could not run, treat the code as unverified. “Tests added” and “tests passed” describe different outcomes. An environment error should lead to a clear limitation, not a success claim.
Keep model settings secondary to evidence
The API documentation lists reasoning levels from low through max, with medium as the API default; none and minimal are unsupported. Those API details do not establish which settings your Work or Codex interface exposes. Sol model reference
For your first attempt, record the model and settings actually shown, keep the prompt and compare results using the same checks. A larger output or longer wait is not proof of correctness.
To practice the fundamentals behind these exercises, start with ChatGPT for Beginners — 2026 Edition, which includes prompting, file inputs and fact-checking. It is a foundation course, not a claim of dedicated GPT-6.1 Sol training.
