Copilot's New Cost Tiers Draw From a Pool That Spans 25x

GitHub added three tiers to Copilot auto model selection on 14 September: Efficiency, Balance and Intelligence. The announcement frames them as a dial between cost and quality, so this is a platform change aimed squarely at LLM cost optimization rather than at features.

The docs source behind the change says something the announcement does not. The tiers move a preference, not a ceiling. Every tier draws from the same pool, and in Copilot CLI that pool spans 25x on price.

What shipped, and where it does not exist

Auto model selection is now two distinct systems rather than one. Auto with task optimization routes on real-time system health combined with an assessment of task complexity, and is generally available in Copilot Chat on the GitHub website, in VS Code, in Copilot CLI, in the GitHub Copilot app, and in the Copilot cloud agent. The three new tiers sit on top of that, and the docs restrict them to three of those five surfaces: VS Code, Copilot CLI and the GitHub Copilot app.

Everywhere else, Auto means a different thing. JetBrains IDEs, Eclipse, Xcode and Visual Studio get auto optimized for model reliability and availability, which chooses on system health and says nothing about task complexity. Copilot Chat on the website and the cloud agent get task optimization but no tier control. If your team standardised on JetBrains, this week's change does not reach them at all.

The pool the tiers draw from

The model list is a YAML table in the docs repository, and it is per-surface. Fourteen models appear across its four columns, but no surface sees all of them: the cloud agent pool is four models, the Copilot app nine, Copilot CLI eleven, Copilot Chat twelve.

Price those eleven CLI models against the pricing table in the same repository, in dollars per million tokens at standard context pricing. Input runs from $0.20 for MAI-Code-1.1-Flash and GPT-5.6 Luna to $5.00 for GPT-5.5, Claude Opus 4.8 and Claude Opus 5. Output runs from $1.20 to $30.00, with GPT-5.5 at the top of that range rather than the Claude models at $25.00. Both spreads are 25x, and the same eleven are available under Efficiency as under Intelligence — the docs are explicit that tiering changes which model is preferred, not which models are present.

The discount for using auto at all is 10%.

That is the falsifiable claim in this post: the tier bounds nothing. If GitHub publishes a price ceiling per tier, or your usage report shows Efficiency never routing above some model, then it is a budget control after all and this is wrong.

Where LLM cost optimization actually happens

Organization and enterprise usage is billed in AI credits at $0.01 each. A Copilot Business seat includes 1,900 per user per month and Copilot Enterprise 3,900, pooled across the billing entity and forfeited at 00:00:00 UTC on the first of each month.

Nineteen dollars of Business allowance, with the 10% discount applied, buys roughly 700,000 output tokens on GPT-5.5 or roughly 17.6 million on GPT-5.6 Luna. Spread across a month that allowance is about 63 credits per user per day. A user consistently above it is either doing more work than the allowance assumes or being routed to the expensive end of the pool, and the model name Copilot prints with each response is what tells you which.

Two controls do bound this, and neither is the tier. The first is the model allowlist: an enterprise or organization owner can disable models for Business and Enterprise seats, and auto excludes anything a policy has excluded. The second is the AI credits paid usage policy, which is enabled by default, so spend past the included pool happens unless an administrator turns it off. GitHub documents no org-level tier policy, which means the cost control shipped this week is the one an administrator cannot set. Item four of the PoC-to-production checklist names routing to cheaper models as a day-one production concern; buy rather than build and the routing becomes the vendor's while the governing stays yours.

Two details to check before you change anything

Evaluation models. Auto may serve them to users on plans for individuals, they appear under codenames rather than model or provider names, and GitHub's own reference states that testing revealed they may perform worse than other models on security-related prompts. Selecting Intelligence does not exclude them. Only the policy does, from the AI controls tab in user settings.

Cache boundaries. The docs say routing happens along natural cache boundaries, and give the reason: switching models mid-session has shown increased cost without a comparable gain in quality. The pricing table shows the size of that effect. A cache read on Claude Opus 5 costs $0.50 per million tokens and a cache write $6.25, figures Anthropic's own pricing page lists identically, so re-caching a context on Opus 5 costs 12.5 times what reading it back would have.

Verdict

Set a tier; the setting itself costs nothing. Do not file it under cost control.

1. Confirm the surface. Tiers reach VS Code, Copilot CLI and the Copilot app, and nothing else.
2. Set the allowlist before the tier. It is the only documented ceiling on what auto can pick.
3. Decide the paid usage policy deliberately. It permits overage by default.
4. Watch per-user burn against 63 credits per day on Business, 130 on Enterprise.
5. Read the model name printed with each response — Copilot CLI prints it in the terminal — and check that Efficiency routes where you assumed.
6. On plans for individuals handling security-sensitive code, disable evaluation models.

The buy-versus-build call in buy, build, or fine-tune usually turns on control. This is what the buy branch costs: a pool you do not set, a router you do not tune, and a dial that expresses a preference to it. An LLM consulting engagement that opens anywhere other than the allowlist and the paid usage policy has opened in the wrong place.