GitHub Copilot's AI Credits, the actual math — what a credit buys and where the hidden costs hide

On June 1, 2026, GitHub moved every Copilot plan to usage-based billing on a currency it calls AI Credits. The announcement gives you totals — Pro gets 1,500, Pro+ gets 7,000 — and a conversion rate, 1 AI credit = $0.01. Put those two facts together and the arithmetic breaks: $10 ÷ $0.01 is 1,000, not 1,500.

That gap is the whole story. Once you see what’s in it, the rest of the system reads clearly.

Why the numbers don’t divide: base + flex

Each paid plan’s credits come in two parts, and the docs and the blog describe them in different units, which is what makes them look contradictory:

• Base credits are “matched 1:1 with your subscription price” — your dollars, turned into credits at 100 per dollar. These “never change.”
• Flex allotment is “variable additional usage on top of your base.” Your base is spent first; “if you go beyond your base, the flex allotment is applied automatically at the same rates.”

So the “1,500” you were quoted is base + flex total, and the “$10” is base only. Both are correct; they’re measuring different things.

All figures from the individual-plans and org/enterprise billing docs. Note the structural difference: individual plans bundle a flex multiplier (Pro effectively gives you $15 of usage for $10), while Business and Enterprise are published at pure 1:1 with no flex shown. Org and Enterprise customers do get a temporary cushion — a promotional boost for the first three months (June 1 – September 1, 2026) lifts Business to 3,000 credits and Enterprise to 7,000, reverting to standard after. (Free and Student plans get “an allowance of GitHub AI Credits” with no published number — treat their budget as unknown.)

One more org-only mechanic worth knowing: credits pool. “An enterprise with 100 Copilot Business users gets a shared pool of 190,000 AI credits rather than 100 individual buckets.” Light users subsidize heavy ones inside the same org — which is good for teams with uneven usage and bad for forecasting any single seat.

What a credit actually buys (the part that matters)

A credit is a fixed $0.01, but the work one credit buys swings wildly, because consumption is metered as plain per-token dollar pricing — “each token is priced based on the model used, and the total is converted into AI credits.” There is no 0×/1×/10× multiplier here; that was the old request-based system. Now the only thing that moves your burn rate is which model you pick and how many tokens it touches.

Here’s the lever, made concrete. Take Pro’s full 1,500 credits ($15) and spend it entirely on output tokens, one model at a time. Using GitHub’s published per-million-token rates (arithmetic ours, output-only for illustration — real sessions also bill input and cached tokens):

Same plan, same $15, and a 15× spread in how much agent work you get out of it depending on whether you reach for GPT-5.5 or GPT-5 mini. The model picker is not a quality dial with a cost side-effect; under credit billing it is your budget dial. Microsoft’s new MAI-Code-1-Flash at $4.50/1M output is clearly positioned as the value default — roughly 7× cheaper per token than GPT-5.5 — though GitHub doesn’t (yet) document it as the automatic selection.

The practical read: a Pro plan is generous if you live on mid-tier and flash models, and genuinely tight if you run frontier models on multi-file agent sessions. The plan didn’t change how good the output is; it changed who pays attention to the model dropdown.

The two costs that don’t touch your credit balance

This is where the bill surprises people, because neither line shows up as credits.

1. Code review is double-metered. A Copilot code review charges “in two ways: token consumption is billed in AI credits, and the agentic infrastructure that powers the review consumes GitHub Actions minutes.” The Actions minutes come out of your repo’s existing entitlement (and overage bills at standard Actions rates), separately from the credits the review’s tokens burn. GitHub does not publish a per-review minute figure, and “the model is selected automatically and is not disclosed, so per-token costs may vary between reviews” — so this is the one line you genuinely cannot forecast in advance. Teams that auto-trigger review on every PR should watch their Actions bill, not just their credit balance.

2. Completions are free — and they’re the only free thing. “Code completions and next edit suggestions are not billed in AI credits. They remain unlimited for all paid Copilot plans.” The old idea of “included models” is gone; no model is 0-credit now. So the ghost-text autocomplete that is most people’s daily Copilot is fully covered, while the agentic/chat work is what draws down credits. If your Copilot usage is 90% Tab-completion, the billing change barely touches you. If it’s agent sessions, it changes everything.

When the credits run out

You hit zero, and then it’s a setting, not a surprise — if you configured it. The overage options are binary:

Critically, “there is no automatic fallback to lower-cost models when a budget is exhausted” — it stops, it doesn’t downgrade. A “$0 user-level budget blocks the user immediately,” which is the lever admins use to cap a seat hard. Budgets exist at user, cost-center, organization, and enterprise levels. One gotcha for individuals: you can’t buy additional credits through the iOS or Android apps — that purchase path is web-only.

If you’re on a monthly Pro or Pro+ plan, you were auto-migrated on June 1. On an annual plan, you stay on the old request-based pricing until renewal — which is also the only place the legacy 0×/1×/10× “model multipliers” still apply. And note that for new individual users this is currently moot: sign-ups remain paused for Student, Pro, Pro+, and Max while the transition lands.

What to actually do

• Find your model default and set it deliberately. It’s the single biggest variable in your bill — a 15× range. If you don’t need frontier reasoning for a task, don’t spend frontier credits on it.
• Separate your usage mentally. Completions: free, ignore. Agent sessions and chat on frontier models: that’s your credit burn. Code review: credits plus Actions minutes.
• Set a budget before the first full month, not after. The default behavior you want — hard stop vs. pay-to-continue — is a choice you make in advance; the surprise invoice is the one you didn’t configure.
• If you’re a team, watch the pool and the Actions line. Pooled credits hide individual heavy use, and review-driven Actions minutes are the cost least visible on the credit dashboard.

The deeper point: Copilot didn’t get more expensive on June 1 so much as it got legible — every token now has a price you can see. That’s strictly better for anyone willing to read the meter, and a quiet tax on anyone who won’t. For where this leaves Copilot against the other agents now that every major vendor is metered or seat-priced, our June verdict has the cross-vendor call.

Companion reading

• Best AI coding agents — June verdict — where Copilot lands now that the whole field went usage-based.
• State of AI coding agents — June 2026 — the billing swap in the context of the month’s other moves.
• Claude Code usage limits, explained — the other major vendor’s metering model, for comparison.