GitHub Copilot Auto Model Selection Gets Efficiency, Balance, and Intelligence Tiers

What shipped

On September 14, 2026, GitHub rolled out tiered control for Copilot auto model selection in VS Code, Copilot CLI, and the Copilot app: efficiency, balance, and intelligence. You pick the objective; auto then routes each prompt to whichever model best fits it.

What changed — and what didn't

Nothing about billing changed, and that's the point. All three tiers draw from the same model pool, and "usage is charged based on the model auto selects, regardless of tier." GitHub's own pricing docs put that pool at a 100× input-token spread: GPT-5.6 Luna costs $0.20 per 1M input / $1.20 output, while GPT-6 Astra on long context runs $20 / $75. Claude Fable 5.1 sits at $10 / $50. A tier is a routing objective, not a spend cap — and GitHub notes a "docstring" task may still get a small model even on Intelligence, meaning the tier constrains nothing you can audit.

The sweetener: paid subscribers get a 10% discount on usage billed through auto — a price advantage that exists only if you surrender model choice.

Why a builder cares

This is Copilot's second opaque router. Code review already picks its model and won't tell you which; now auto offers a knob that looks like cost control but isn't, plus a discount that pulls you off pinned models. If you're budgeting AI credits (1 credit = $0.01), pin the model for anything throughput-heavy. "Intelligence" as a default is how a $0.20 prompt becomes a $75 one without a line item explaining why.