GitHub Copilot coding agent pricing: what the 2026 AI credits change actually costs
GitHub moved Copilot to usage-based AI credits on June 1, 2026. What the coding agent costs per session now, what each plan includes, and where Actions minutes come in.
If your mental model of Copilot pricing is "flat $10 a month," it is out of date. On June 1, 2026, GitHub moved Copilot to usage-based billing: the old premium-request counting is gone, and what you get with a plan now is a monthly allotment of AI credits that drain based on token consumption. For the coding agent specifically, that is only half the bill. The agent runs in a GitHub Actions environment, so its sessions consume Actions minutes too.
This post lays out what each plan includes, how the credits actually burn, and what a coding agent session plausibly costs. We keep running into people who think the agent is a flat-fee feature, and as of June it just isn't billed that way.
Prices and mechanics checked September 2026 against GitHub's own docs. They have changed once this year already, and the promotional seat allotments lapsed on September 1; check the linked sources for current numbers.
Full disclosure: we build Sinatra, a background coding agent that competes with Copilot's, and our pricing takes the opposite approach. Every claim about Copilot below comes from GitHub's published docs and blog, linked as we go. We also compared the two products feature-by-feature in Devin vs GitHub Copilot coding agent and the background agents roundup.
What changed on June 1, 2026
Under the old model, Copilot metered "premium requests": each plan included a monthly count, a coding agent session consumed one request multiplied by the model's rate, and overage was $0.04 per request. Predictable, if a little opaque.
The new model bills like an API. Each plan includes a monthly allotment of AI credits, one credit is a cent, and credits are consumed based on token usage: input tokens, output tokens, and cached tokens, at published per-model rates. A short chat costs almost nothing. A coding agent session that reads a large repo and iterates on a change can consume a meaningful fraction of a plan's monthly allotment in one go.
What each plan includes
| Plan | Price | Included AI credits / month | Credits at face value |
|---|---|---|---|
| Copilot Pro | $10/mo | 1,500 | $15 |
| Copilot Pro+ | $39/mo | 7,000 | $70 |
| Copilot Max | $100/mo | 20,000 | $200 |
| Copilot Business | $19/user/mo | 1,900 per user | $19 |
| Copilot Enterprise | $39/user/mo | 3,900 per user | $39 |
Two things in that table are easy to miss. The individual plans hand you more credit value than the plan costs — Pro is $10 for $15 of credits, Pro+ $39 for $70 — so the sticker price is a discount on the meter rather than a cap on it. The seat plans are the opposite: a Business seat is $19 for exactly $19 of credits. Existing Business and Enterprise customers ran on higher promotional allotments of 3,000 and 7,000 credits from June 1 through September 1, 2026; that promotion has ended and those seats are now on the standard amounts.
Once the included credits are gone, there is no free fallback. Individuals can buy more credits; organizations decide via budget controls whether members can keep spending past the allotment or get cut off. Seat credits pool across the organization rather than stranding in each user's account, so a team's real allowance is the sum of its seats, and one heavy user draws on it.
How fast do credits burn?
Credits map to token consumption at per-model rates that look a lot like API price sheets. As of this writing, the published ranges per million tokens:
| Model tier | Input | Output |
|---|---|---|
| Lightweight (GPT-5 mini class, Haiku class) | $0.20 to $1.00 | $1.20 to $5.00 |
| Versatile (Sonnet class, mid GPT) | $0.75 to $3.00 | $3.75 to $15.00 |
| Powerful (Opus class, top GPT) | $2.00 to $10.00 | $10.00 to $50.00 |
Long-context tiers run at twice the default rates. Cached input is much cheaper than fresh input, roughly a tenth of the price, which matters a lot for agent sessions because most of what an agent "reads" on each turn is repeated context.
GitHub does not publish a per-session average for the coding agent, so here is an illustration with the assumptions stated. Take a mid-size session on a versatile-tier model at $3 per million input and $15 per million output: the agent reads two million tokens across its turns and writes two hundred thousand. With no cache hits that is $6 of input and $3 of output, $9 for one session, against the 1,500 credits — $15 — a Pro plan includes for the whole month. With realistic cache behavior (say 80 percent of input served from cache at a tenth of the price) the same session lands closer to $4.70. Either way, a Pro plan is a handful of agent sessions per month, not a month of agent usage.
The second meter: Actions minutes
The coding agent executes inside an ephemeral GitHub Actions environment, and those runner minutes are billed against your plan's included Actions minutes, then at standard Actions rates past that. This is not unique to the agent anymore: as of June 1, 2026, even Copilot code review consumes Actions minutes for private repositories.
Minutes are cheap relative to tokens, and public repos get free ones. But if you point the agent at a private monorepo with a slow dependency install, the compute for every session lands on the same Actions bill as your CI, and it is worth knowing that before the invoice explains it to you.
Who this pricing works for
Usage-based pricing has real virtues. Light users stop subsidizing heavy users, and the meter tells you the truth about what a session cost. If your team lives in GitHub Issues, already budgets Actions minutes, and hands the agent a few well-scoped issues a month, the included credits may genuinely cover you, and the integration is the smoothest on the market because it is the platform vendor's own.
The people who get surprised are the ones who took "assign it an issue" to mean "use it freely." An agent that runs for twenty minutes reading a large codebase is an expensive API workload, and the June change made that visible on the bill.
If metering is the dealbreaker
The alternative model is to pay a flat price for the orchestration and let the model bill ride on something you already pay for. That is how we built Sinatra: you assign it a Linear issue, or add the Sinatra label to a GitHub issue, and the run bills the Claude or ChatGPT subscription you already have (or your own API key) with no markup on tokens. The free tier is 5 tasks a day on your own credentials, every day, and the paid tier is $20 per member per month, flat. We are biased, and the pricing model is most of why the product exists.
If you want neither vendor: Google Jules includes 15 free tasks a day on Gemini models, and OpenHands is open source with an at-cost hosted option. The full comparison covers all of them.