Paste only this prompt into a fresh conversation with your chosen model. Keep the answer key here, outside that conversation.
Find your good-enough AI setting.
Copy a test into ChatGPT, then record the result here. Start with one task and compare a few settings.
BUILD YOUR OWN EVIDENCE
Your results
Compare the same task, surface, conditions and batch. No model has a score until you record one.
Run history
“Promising” means at least 3 recorded runs all passed this exact test. It is a small-sample screen, not a reliability guarantee. Before adopting a default, check at least 5 fresh examples of your real work, including an edge case. Keep failures in the record.
A STARTING POINT, NOT YOUR RESULTS
Model, effort, and workspace
The model is the capability you select. Reasoning effort controls how much reasoning it can devote to the job. Chat or Work sets the product workflow and available tools. Test the whole combination you intend to use.
| Model | Jobs to try | First setting |
|---|---|---|
| GPT-6 Luna | Summaries, extraction, focused coding | High |
| GPT-6 Sol | Everyday work, complex coding and workflows | Medium |
| GPT-6 Astra | Difficult work across code, apps and research | Light / Low |
GPT-6 Sol and Luna are in Work and Codex, not Chat. Availability varies. In Chat, record the exact available model or Auto label. An Auto result evaluates the routing experience, not a guaranteed fixed model.
Low/Light favors speed. Medium adds depth. High/Extra High can help harder tasks. Max increases reasoning depth. Ultra uses parallel subagents; Luna supports up to Max. Equal effort names across models do not imply equal performance.
Checked September 22, 2026. These are starting hypotheses, not benchmark scores. Official model guide · Subagents and Ultra
YOUR FIRST EXPERIMENT
Get an answer without testing everything
- Choose one frequent task. Use the same prompt and input in each fresh conversation. For Work, use separate empty output folders or identical copies of the starting files.
- Screen three combinations. Try Luna High, Sol Medium, and Astra Light when available in Work. For Chat, compare two or three settings actually offered there.
- Grade before repairing. Keep the original response and check the answer key. Pass = at least 8/10 and no critical failure. Include failed and interrupted runs.
- Repeat the promising candidates. Run each twice more in fresh contexts; alternate the order. If a task takes a long time, start with two candidates.
- Step down or escalate. If a model passes, try one lower effort. If it fails, test one higher effort or another model. Change one thing at a time when diagnosing why.
- Choose on quality, then burden. Among reliable-enough options, compare total model wait + your review/repair time, follow-ups, and measured usage. Unknown usage stays unknown.
- Confirm on fresh work. Use at least five new representative examples, including an edge case. A single canned test cannot establish reliability for a whole category.
KEEP IT FAIR
What can fool the comparison?
- One model getting extra hints, files, memory, browsing or tools. Keep inputs fixed; log unavoidable differences.
- Showing the answer key to the tested model. Copy only the task prompt.
- Rewarding confidence, length or impressive formatting instead of correctness and usable output.
- Only recording successful retries. Preserve the first attempt and correction effort.
- Giving a Work task a passing score for a plan. Open the file or app and check it.
- Letting Ultra's extra agents look like an isolated-model comparison. Record delegation and compare the whole workflow separately.
For subjective writing, grade anonymized A/B responses against the same rubric; swap their order and allow a tie. An AI grader can assist, but its judgment is not ground truth. Verify factual and executable checks yourself.
Method adapted from OpenAI evaluation best practices. The 8/10 threshold and repetition counts are practical screening rules chosen for this kit, not OpenAI guarantees.
WHEN SOMETHING NEW COMES OUT
Run a challenger test
Keep your current winner. Start a new batch and compare it with the new model on three saved tasks: a routine task, a typical task and a difficult or previously failed task. Use the same inputs in fresh conversations. Repeat close results and inspect failure cases.
Switch only where the new option preserves the required quality and improves your time or measured usage. It can win one kind of work and lose another. Repeat after meaningful model, tool or workflow changes.
For browsing research, first use the frozen source test here. Then use a separate live research test where you check sources and timestamps; web changes make that a product test with extra variability.
Personal routing card: For [task type], in [Chat/Work], use [model + effort] under [conditions]. Evidence: [date/batch, examples, pass rate, time]. Escalate when [specific failure]. Recheck after [change].
This kit does not submit prompts or run models. You control each test in ChatGPT. It has no API key or service dependency.