Quick answer
Use the mini model for search, summarising, and small edits, and the strong model for multi-file changes and debugging that needs reasoning. The mini model is faster and cheaper per token, but it can take more turns on hard tasks, which erases the saving. Measure tokens per completed task, not price per token.
The real trade-off
| Dimension | Mini | Strong model |
|---|---|---|
| Latency | Lower | Higher |
| Cost per token | Lower | Higher |
| Turns on hard tasks | More | Fewer |
| Context discipline | Needs tighter prompts | Tolerates more |
| Best use | Search, summaries, small edits | Refactors, debugging, design |
A cheaper model that needs three attempts is not cheaper. The unit that matters is tokens per completed task.
A workflow that uses both
- Use mini to locate the relevant files and summarise them.
- Paste that summary into a strong-model session with the specific change.
- Use mini again for mechanical edits and test fixes.
- Return to the strong model only when the mini model loops.
Verify on your own tasks
Run the same task in both models with a fresh session and compare total tokens, not just output. If the mini model uses more than about twice the tokens, the strong model is the better deal for that task type.
UsageAtlas splits usage by model, so the comparison uses your real work rather than a synthetic benchmark.
Choose the model by task shape
| Task | Model class | Why |
|---|---|---|
| Locating a symbol | Small | Cheap search, short answer |
| Renaming across files | Small to mid | Mechanical edits |
| Debugging a race | Reasoning | Needs hypothesis testing |
| Designing an interface | Reasoning | Long output, many constraints |
| Writing tests | Mid | Repetitive with some judgement |
The expensive model is not always the faster route. A reasoning model that spends two thousand tokens deciding where to look can be slower and costlier than a small model with a precise file path. Measure with tokens per task before making one model the default.
Questions, answered
Is the mini model good enough for coding?
For small, well-scoped edits yes. For multi-file reasoning it often needs more turns, which can cost more overall.
Does model choice affect the usage window?
Yes. Different models consume the allowance at different rates, though both draw on the same window.
Can I switch models mid-session?
Yes, with /model. The context stays, and the new model's pricing applies from that turn.