By ByteBot
GitHub updated Copilot’s auto model selection on September 14 with three configurable tiers — Efficiency, Balance, and Intelligence — that let you tell the selector how to weigh cost against quality for every prompt. If you are on AI Credits billing and wondering why your Copilot spend looks nothing like what you expected since June, this setting deserves your attention.
What the Tiers Actually Do
Here is the part that matters most: all three tiers draw from the same pool of available models. The tier does not restrict or unlock anything. What changes is the optimization target the auto selector uses when it evaluates your prompt.
On Efficiency, the selector prioritizes low-cost models and routes to lighter options unless the task clearly demands more. On Balance, it weighs cost, quality, and latency together — the closest thing to the old behavior before tiers existed. On Intelligence, it gives preference to higher-capability models and is willing to spend more credits for a better answer.
There is an important nuance: even on Intelligence, a simple request like adding a docstring to a function will likely still route to a small model. The tier is a priority signal, not a floor. The real difference appears on moderate-complexity prompts where the system has a genuine choice — and that describes most of what you actually send Copilot.
The Numbers Are Real
Visual Studio Magazine ran the same prompt across all three tiers and published the AI Credits cost for each. The results: 0.88 credits on Efficiency (routed to Gemini 3.7 Flash) versus 4.29 credits on Intelligence (routed to Claude Sonnet 5). That is a 4.9x difference on a single prompt.
Multiply that across a full day of development and the tier you choose becomes a meaningful budget variable. Copilot Pro includes 1,500 credits per month at 0. That is roughly 1,700 prompts on Efficiency or about 350 on Intelligence before you hit overages.
This is the direct consequence of the June 2026 billing shift to token-based AI Credits. When every model request counted as one interchangeable premium unit, model choice was irrelevant to your bill. Now there is up to a 6.7x price spread between the cheapest and most expensive models in Copilot's pool, and auto model selection has been making that call for you. Tiers let you set the direction instead.
Which Tier to Use for What
The practical breakdown:
- Efficiency: Docstrings, boilerplate generation, import fixes, simple renames, code formatting. Any task where a capable small model gets the job done. High-volume work like CI-integrated code review or bulk documentation runs belongs here.
- Balance: Feature implementation, standard debugging, code review, explaining unfamiliar code. The right default for most developers doing daily work — and the tier that behaves most like what you were used to before this update.
- Intelligence: Architectural design reviews, complex multi-file refactors, debugging subtle distributed system failures, security audits. Tasks that genuinely need a frontier model, not tasks that just feel complex because you are stuck.
One honest caveat: the router cannot perfectly detect complexity. A tricky async race condition described in a short casual prompt may still get routed to a lighter model even on Intelligence. Tiers shift the probability in your favor; they do not guarantee a frontier model shows up every time you need one.
This Is Not HydraFusion
Since we covered HydraFusion when it launched earlier this month, worth being explicit: these are separate features. Tiers apply to standard auto model selection, which picks one model per request. HydraFusion — still in /experimental in Copilot CLI — replaces that flow entirely with multi-model orchestration that can run a draft-and-critique workflow across two different model families in a single turn. If you enable HydraFusion, the tier setting does not apply to those requests.
How to Configure It
The feature is rolling out now to VS Code, Copilot CLI, and the GitHub Copilot app. JetBrains and Visual Studio (Windows) are not yet included. Look for the tier option in the model picker next to the Auto selection. The 10% auto model selection discount continues under all three tiers — check the official changelog for the latest rollout status.
Enterprise admins can set a default tier for their organization via the managed-settings.json file, with support for per-team overrides added in the September 2 enterprise settings update. Users retain the ability to change tiers per conversation regardless of what the admin sets.
If you have not revisited your AI Credits usage since the June billing shift, this is a good moment to do it. Start with Balance as your default, switch to Efficiency for high-volume routine tasks, and reserve Intelligence for the work that actually earns it.
