Kibble Blog

How much does GitHub Copilot cost per engineer per month?

By Kibble · 10 September 2026 · 6 min read

Since 1 June 2026 a GitHub Copilot seat is a floor, not the cost. Copilot Business is $19 per user per month with $19 of AI credits included, and usage past the allowance is charged by tokens at a published rate per model. The three labelled scenarios below cost about $0.50, $8 and $80 a month on the default GPT-5.6 Luna, and about $4.50, $80 and $770 on Claude Sonnet 5. They are made-up inputs priced at GitHub's rates, not a measured average; every assumption is listed so you can substitute your own.

What changed in Copilot billing?

Until May 2026 Copilot counted premium requests: each interaction cost one request times the model's multiplier, a plan included a monthly number of them, and extra requests were $0.04 each. On 27 April 2026 GitHub announced usage-based billing, effective 1 June 2026: consumption is now measured in AI credits, one credit is one US cent, and credits are consumed by input, output and cached tokens at the listed rate for each model. Each paid plan includes credits equal to its price. Administrators choose whether usage past the allowance continues at published rates or stops, and can set budgets at the enterprise, cost-center and user level. Code completions and next edit suggestions consume no credits on any paid plan.

Those are the figures in GitHub's announcement, read on 10 September 2026. Promotional credits for the first months and negotiated terms are outside this table; check the current plan page before budgeting.

What are the token rates?

The rates below come from GitHub's models and pricing page on the same date, in US dollars per million tokens, for the default context tier. Long-context tiers cost more. Earlier OpenAI models have no cache-write charge, so on GPT-5.4 the tokens written to the cache are ordinary input.

The Claude rows match Anthropic's own list rates, which is what makes the scenarios below comparable with the Claude Code article. A newer model can carry different cache pricing, so do not apply a multiplier to an unlisted version.

What usage does each scenario assume?

The three scenarios are the same made-up inputs as the Claude Code article, so the two can be read side by side. Light, moderate and heavy are labels for these numbers, not percentiles from a population of engineers. An agent-day is one day with the usage shown; inactive days contribute zero. The mix comes from Claude Code transcripts, and Copilot's cache behaviour on a given model may differ, which is exactly why a collector prices what the logs record rather than assuming a ratio.

For each category, divide tokens by one million and multiply by that category's rate, add the four amounts for a daily figure, then multiply by agent-days. The moderate scenario on GPT-5.6 Luna is $0.004 fresh input + $0.15 cache writes + $0.18 cache reads + $0.132 output = $0.466 a day, or $8.39 across 18 days: 839 credits.

What do the scenarios cost per month?

These are the exact totals for the invented inputs above, rounded to cents. One credit is one cent, so $79.92 is 7,992 credits.

This holds token usage constant to isolate price. It is not a benchmark of how well each model does the same task: models take different numbers of turns and produce different outcomes. Use your own distribution for a budget, and investigate a change in the bill before treating it as waste.

Where does the included allowance run out?

A Business seat includes $19 of credits a month. On the default model, only the heavy scenario passes it. On Claude Sonnet 5, only the light scenario stays inside it. Whether the excess is charged at published rates or the engineer is stopped is the administrator's budget setting, not the engineer's choice.

So the per-engineer number is the seat plus the overage, and the overage depends on the model as much as on the volume. A team that stays on the default model has a Copilot bill that is mostly seats; a team that picks Opus for everything has one that is mostly credits. Under the old premium-request model the model choice was a multiplier on a fixed unit; now it is the unit.

How do you see this for a real team?

GitHub's usage metrics API has reported ai_credits_used per user per day since 19 June 2026, to organization owners and enterprise administrators. GitHub calls it a metrics signal rather than a billed total, and it carries no split by model, feature or surface; the billing page is the invoice. That per-user figure answers who, and stops there.

A local collector fills in the why. Kibble reads the session logs Copilot CLI and VS Code already write, prices the recorded counters at list rates, and shows the result per engineer, model and repository, next to Claude Code and Codex. The result is an estimate, labelled as one, and it is never added to the invoice: two views of the same usage are never summed. Which surfaces the logs cover, and what they cannot say, is set out in the tracking guide.

How does this compare to the invoice?

The invoice includes the seats of people who used nothing, taxes, promotional credits, negotiated discounts and other GitHub charges such as Actions minutes for Copilot code review. These scenarios exclude all of that. They also exclude completions, which cost no credits, and the cloud coding agent and code review, which consume credits but leave no counters on a laptop.

Kibble (usekibble.com) is a Claude Code, Codex and GitHub Copilot usage dashboard for teams: it applies this arithmetic to collected counts, keeps repository attribution an estimate, and keeps the estimate apart from what the vendor charged.