GitHub Copilot Auto Model Selection Gets Three Cost-Quality Tiers
GitHub Copilot auto model selection now has three tiers: efficiency, balance, and intelligence. Here's what each tier does and who should use it.
GitHub has added three configurable tiers to Copilot's auto model selection feature: efficiency, balance, and intelligence. The update is rolling out now in Visual Studio Code, Copilot CLI, and the GitHub Copilot app. Each tier tells Copilot how to weigh cost, output quality, and response time when it picks a model for each prompt. You are still billed based on the model actually selected, and paid subscribers keep a 10% discount on auto-billed usage.
What happened
| Detail | Fact |
|---|---|
| Feature | Three tiers for Copilot auto model selection |
| Tiers | Efficiency, Balance, Intelligence |
| Rollout platforms | Visual Studio Code, Copilot CLI, GitHub Copilot app |
| Billing basis | Model auto selects, regardless of tier chosen |
| Paid subscriber discount | 10% on usage billed through auto |
GitHub’s auto model selection already picked a model for each Copilot prompt automatically. This update lets you steer that selection by choosing a tier that matches your priorities. The three options work as follows:
- Efficiency: favors low cost and speed, suited to simple, repetitive tasks.
- Balance: weighs cost, quality, and latency together, GitHub’s recommended default for everyday coding work.
- Intelligence: puts output quality first, designed for complex tasks where a slower or pricier model is worth it.
Importantly, all three tiers draw from the same set of available models. Choosing a tier changes the priority, not the model pool. GitHub gives a concrete example: even with the intelligence tier active, a simple job like adding a docstring to a function may still trigger a small, efficient model because auto evaluates each prompt individually.
How billing actually works
Your bill reflects the model auto selects, not the tier you pick. If you set intelligence but Copilot judges a prompt simple enough for a lightweight model, you pay the lightweight model rate. Paid subscribers continue to receive a 10% discount on any usage billed through auto, across all three tiers.
GitHub describes this as the first step toward giving users more control and visibility into the tradeoffs behind model selection, suggesting more configuration options are coming.
Why it matters
Teams running Copilot at scale can now tune spend without abandoning auto selection entirely. A junior developer doing repetitive edits probably belongs on efficiency. A senior engineer debugging a gnarly async race condition might want intelligence for that session. Having a per-tier setting rather than a per-model setting is simpler to manage across a team.
The fact that billing follows the actual model and not the tier is worth noting carefully. Setting intelligence does not mean every prompt gets an expensive model. It means Copilot will lean toward quality when the prompt warrants it. For cost-conscious teams, this is actually good news since the efficiency tier still gets to use larger models when the task genuinely needs one.
This kind of cost-quality tradeoff control mirrors what we see emerging across AI tooling broadly. As we covered in our look at AI agent energy and cost overheads, the real operational cost of AI tools is often opaque until you have a bill in hand. Giving teams a dial to set expectations upfront is a practical step forward.
Our take
Three tiers is a sensible starting point. The naming is clear enough that non-technical stakeholders can understand the intent, which matters when a team lead is setting defaults for a whole organisation. The billing mechanic (pay for what auto picks, not what tier you chose) is honest and avoids a situation where “intelligence” mode becomes a blank cheque.
What we would want to see next: per-project or per-workspace tier settings, and actual visibility into which model was selected for a given prompt and what it cost. GitHub hints that more visibility is coming. Until then, teams should check usage reports after switching tiers to understand what the real cost delta looks like in practice.
If you are helping clients integrate tools like Copilot into their development workflows, this is a good moment to revisit the setup. Our AI integration service covers exactly this kind of tooling configuration and cost governance.
What to do about it
- Open Copilot settings in VS Code, Copilot CLI, or the GitHub Copilot app and check whether the tier selector is live for your account yet.
- Set efficiency for team members doing high-volume, routine tasks, and balance or intelligence for those working on architectural or complex debugging work.
- Pull your usage report one week after switching tiers to see whether actual model selection shifted and what the cost impact is.
- Wait for GitHub’s promised visibility improvements before building any hard budget assumptions around tier behaviour.
Start with balance if you are unsure: it is GitHub’s own recommended default and gives you a cost and quality baseline to measure from.
Frequently asked questions
What are the three tiers in GitHub Copilot auto model selection?
The three tiers are efficiency (prioritises low cost and speed), balance (weighs cost, quality, and latency equally), and intelligence (prioritises output quality for complex tasks).
Does choosing the intelligence tier mean I pay more?
Not necessarily. Billing is based on the model Copilot actually selects for each prompt, not the tier you chose. A simple prompt may still trigger a cheap model even on the intelligence tier.
Where is the Copilot auto model selection tier available?
The feature is currently rolling out in Visual Studio Code, Copilot CLI, and the GitHub Copilot app.
Do paid Copilot subscribers still get a discount when using auto model selection?
Yes. Paid subscribers continue to receive a 10% discount on usage billed through auto, across all three tiers.
