GitHub Copilot adds cost-vs-quality tiers to auto model selection
Set a spend ceiling or force quality per prompt while auto-routing stays on; paid plans get 10% off auto-select usage. Plus Agnes open weights trend.
Copy markdown
Three tiers, one dropdown
Copilot's auto model selection now takes an Efficiency, Balance, or Intelligence setting. Efficiency keeps costs low for fast, simple tasks; Balance weighs cost, quality and latency for everyday work; Intelligence favors quality for the hard stuff. It's a ceiling you set once, not a model you pin.
It still routes per prompt
Within whatever tier you pick, Copilot scores each prompt and grabs the cheapest model that can handle it, so even Intelligence can drop to a smaller model for a trivial edit. You bound the range; the router spends inside it.
What it does to your bill
You're charged for whichever model the router actually uses, whatever the tier. Paid subscribers get 10% off usage billed through auto model selection, which makes staying on auto-select cheaper than manually pinning a premium model.
Where it's live
Rolling out now in VS Code, the Copilot CLI, and the GitHub Copilot app. GitHub calls it “the first step” toward more auto-selection controls, so expect finer knobs later.
Elsewhere: Agnes 3.0 Flash open weights
Agnes AI's 33B reasoning model is trending on Hugging Face as an Apache-2.0 preview that runs on a single H100 (~66GB at BF16) with a 256K-token context. It posts roughly 85 on GPQA Diamond and lists at $0.05/$0.15 per 1M tokens via API — flagship-class reasoning without flagship hardware.