Module 5 · ModelsVerified 2 Nov 2026
Choosing a model
There is no best model, only a best model for this task. Learn the categories once and the names stop mattering.
| Type | What it is | Strengths | Trade-off |
|---|---|---|---|
| Fast model | Optimised for speed and low cost. | Quick edits, boilerplate, simple questions, short explanations. | Can miss subtle bugs or long chains of reasoning. |
| Reasoning model | Thinks before it answers (see effort levels). | Multi-step debugging, refactors, architecture and trade-off analysis. | Slower and uses more AI credits, especially at high effort. |
| Large-context model | A bigger context window to hold more at once. | Big codebases, long logs, long-running conversations. | More tokens in means more cost; quality can still fade over very long chats. |
| Auto | Routes each request to a suitable model for you. | Everyday work. Weighs task complexity and live model availability, and respects your organisation's model policy. | Less control over exactly which model answers (hover a reply to see which one did). |
| Copilot Fusion | Orchestrates one or more models with a draft, critique and revise workflow. | Substantial, well-scoped tasks where an extra review pass is worth the wait. | Slower and can use noticeably more AI credits than a single request. Availability depends on your account. |
Model advisor
Pick your task
Fast model
The answer is obvious, so speed and cost matter more than depth.
Suggested effort
Low
Tip
Be specific so a smaller model has nothing to guess.
Generic guidance by model category. Exact model names and availability change often, so check your model picker.
| Task | Model type | Effort | Tip |
|---|---|---|---|
| Boilerplate, renames, small edits | Fast model | Low | Be specific so a smaller model has nothing to guess. |
| Explain an error or a short snippet | Fast model | Low | Paste the full stack trace and use #selection. |
| Implement a clearly specified function with tests | Auto | Medium | Point at an existing example with #file. |
| Multi-step debugging across modules | Reasoning model | High | Give logs and a failing test, and ask for ranked causes before any fix. |
| Architecture review or trade-off decision | Reasoning model | High | State constraints (team size, SLOs, budget) so it can reject options. |
| Large refactor touching many files | Reasoning model | Medium–High | Use Plan mode, then switch to Agent with the plan attached. |
| Summarise a huge log or a whole module | Large-context model | Low–Medium | Trim noise first. Use /compact to keep long chats lean. |
| High-stakes migration in unfamiliar territory | Copilot Fusion | Handled by Fusion | Give it a precise goal and acceptance tests. Do not use it for trivial tasks. |
| Review a security-sensitive diff | Copilot Fusion | Handled by Fusion | Treat output as a reviewer's opinion, not a verdict. Verify findings. |
| Docs, changelogs, commit messages | Fast model | Low | Provide a template, and use a prompt file to reuse it. |
| A mixed day of small tasks | Auto | Adaptive | Set Optimize for to Efficiency, Balance or Intelligence if your picker offers it. |
- Start a new chat per task. Long histories are re-sent on every request.
- Run /compact (optionally with a hint such as /compact keep the API decisions) before a long chat gets expensive.
- Raise thinking effort only for hard problems, and drop back to Low or adaptive afterwards.
- Default to Auto, and save Fusion and high effort for work where an error would cost you hours.
- Attach only the context you need: #file or #selection beat dumping a whole folder.
- Put recurring rules in instruction files instead of re-typing them. Shorter prompts mean fewer tokens.
- Watch your usage in the account view and use the Cache Explorer to see how well prompt caching is working.
Bring your own key
You can add your own model provider for chat (not for inline completions). Useful when your team already pays for one.