Choosing a model
Beta, early access on request: which models to turn on for Maestro, from measured results, and which job each one suits.
- Recommended default
- gpt-5-mini
- For the hardest work
- gpt-5
- Measured
- October 2026, 85 tasks
Maestro is in release 3.0 as Beta and is not turned on by default. To use it, ask MaestroHub for early access: you get a licence that carries Maestro. A system administrator then connects an AI provider under Second Brain → Providers. These results come from a development build in October 2026.
gpt-5-mini is the default. It comes in the model catalog as a built-in model, beside gpt-5, once an OpenAI provider is connected. It passed 80% of tasks Maestro was never tuned on and 95% of the tuned set, at about $0.06 a task. That's no more than gpt-4.1-mini costs, for much better results.
Only models that pass Maestro's own checks are built in. Both passed every "Draft it for me" form (every connector's function and connection settings) and the chat checks in October 2026. Any other model can be imported and used; it hasn't been tested.
Results
| Model | Tasks Maestro was tuned on | Unseen tasks | Cost per task (list price) |
|---|---|---|---|
| gpt-5 | 89% | 86% | $0.35 |
| gpt-5-mini | 95% | 80% | $0.06 |
| gpt-5-nano | 69% | 74% | $0.02 |
| gpt-4.1-mini | 89% | 57% | $0.07 |
| gpt-4o-mini | 68% | 50% | $0.04 |
| gpt-4.1-nano | 37% | 25% | $0.03 |
- Unseen tasks are the better guide to your own requests: 20 tasks written from this documentation by a separate author and never used to tune Maestro.
- Cost is at list price, before the discount providers give for repeated prompt text, so real bills are lower.
Which model for which job
| What you use Maestro for | Recommended | Also works | Avoid |
|---|---|---|---|
| Building pipelines and apps from a description | gpt-5-mini | gpt-5 | gpt-4.1-mini, gpt-4o-mini, nano models |
| Editing, running and debugging pipelines | gpt-5-mini | gpt-5, gpt-4.1-mini | gpt-4o-mini, gpt-4.1-nano |
| Questions about live plant data | gpt-5-mini | gpt-5-nano, gpt-4o-mini, gpt-4.1-mini | gpt-4.1-nano |
| Connections, functions, schemas, dashboards | gpt-5-mini | most models | nano models |
- gpt-5-nano answers data questions well, but it built apps in only 3 of 10 runs, and once acted on a deliberately vague request instead of asking. Keep it to read-only questions.
- gpt-4.1-mini and gpt-4o-mini handle questions and routine edits, but they struggle with requests unlike the ones they were tuned on.
- When Maestro is stuck on a request, move that one message up to gpt-5 with the model menu under the composer. See Build an app.
Safety does not depend on the model
No model, strong or weak, wrote or deleted anything without the person's approval. MaestroHub holds every change on a confirmation card until the person says yes. The model decides what to propose; the person decides what happens.
How this was measured
- Setup: each model ran Maestro on the same build, in a workspace seeded with about 2,000 plant topics.
- Grading: each task was graded on what it left in the instance (topic values, saved pipelines, app drafts) and on the figures in the answer.
- Runs: gpt-4.1-mini ran every task three times; the others ran once. Treat differences under about 10 points as noise.
- Partial runs: gpt-5's unseen score comes from 14 of the 20 tasks.
- Not yet measured: Anthropic and Google models in Maestro, and newer OpenAI models.