Copilot AI Credits Explained: What GitHub’s Usage Billing Really Costs

GitHub moved Copilot to usage-based AI Credits in June 2026. How credits are counted, what changed on September 1 and October 1, and how to forecast your bill.

Disclosure: this guide is based on GitHub’s billing documentation and announcements as reported by several outlets. We have not audited a large organisation’s Copilot invoices. Plan figures are attributed and may change; check GitHub’s current documentation before budgeting.

On June 1, 2026, GitHub replaced Copilot’s premium request units with Copilot AI Credits, a usage-based currency worth one cent each. Code completions stayed free and unlimited. Everything heavier, especially agent mode and premium models, now draws from a monthly credit allowance priced at the underlying model’s API rate.

The honest summary for anyone managing a Copilot budget: base seat prices did not change, but what a seat buys did. A generous promotional boost masked the change for three months. When it expired on September 1, Business allowances fell by about 37% and Enterprise by about 44%. Then on October 1, existing customers had to start prepaying for seats. If your developers lean on agents, your real cost per seat now depends on behaviour, not on the price list.

Table of Contents

Copilot AI Credits lighthouse guiding ships at dusk
Completions stay free. Agents are where Copilot AI Credits disappear, so that is where the budget needs a light on it.

What Copilot AI Credits Are

Copilot AI Credits are GitHub’s unit of Copilot usage under usage-based billing. One credit equals $0.01. When you use a feature that consumes credits, GitHub measures the tokens involved, including input, output and cached tokens, prices them at the published API rate of the model doing the work, and deducts the equivalent number of credits.

GitHub’s chief product officer, Mario Rodriguez, announced the change in April and described it as a step toward a sustainable, reliable Copilot business. GitHub’s community post framed it as aligning pricing with the higher compute consumption of agentic workflows. In plain terms: agents use far more tokens than autocomplete, and flat pricing could not absorb that.

Code completions, the inline suggestions most developers think of as Copilot, remain unlimited and do not consume Copilot AI Credits. That is the most important reassurance for teams that use Copilot mainly as a smart autocomplete. For them, very little changed.

InfoWorld noted this was the second pricing recalibration for Copilot in less than a year. Expect more as model prices and usage patterns keep shifting.

Copilot AI Credits by Plan

Plan Seat price Included credits per month Promotional (June 1 to Sept 1)
Copilot Free $0 Limited free usage Not applicable
Copilot Pro $10 1,000 ($10) Not applicable
Copilot Business $19 1,900 ($19) 3,000 ($30)
Copilot Enterprise $39 3,900 ($39) 7,000 ($70)

The pattern is simple once you see it: each paid seat includes credits equal to its price. A $19 Business seat buys $19 of model usage on top of unlimited completions. Credits on individual plans are not pooled; check your organisation’s settings for how Business and Enterprise allowances are shared.

SAMexpert’s licensing guide described a split between base credits, which match the subscription price and do not change, and a flex allotment, a variable extra amount GitHub adjusts as model pricing and efficiency evolve. The promotional boost was the most visible example of flex credits.

Copilot AI Credits: The Three 2026 Changes

Codercops summarised the year as three separate changes, and it is worth keeping them apart because they hit at different times.

June 1: credits replace premium requests

Premium request units ended and Copilot AI Credits began. Plans with a premium request allowance were migrated to the equivalent credit allowance automatically. Existing Business and Enterprise customers received the promotional boost.

September 1: the promotion ends

The promotional band ran for the first three months of usage-based billing. On September 1, Business seats dropped from 3,000 to 1,900 credits, a 36.7% cut, and Enterprise from 7,000 to 3,900, about 44%. Ecorpit’s analysis called it the real cost floor. The same date, new Business and Enterprise seat assignments began requiring upfront payment before a user gained access.

October 1: prepayment for existing customers

DevOps.com reported that existing Business and Enterprise customers got a month’s extra runway before the upfront payment requirement applied to them on October 1. Codercops noted that many teams who thought October was the change had actually felt the bigger impact on September 1.

Alongside the billing changes, GitHub began merging Copilot Cloud Agent, the chat experience on github.com and GitHub Mobile Chat into a single experience, rolling out no earlier than September 28.

How Copilot AI Credits Get Spent

Understanding the drivers is the key to controlling the bill.

The model matters. Credits are charged at the underlying model’s published API rate. A premium frontier model costs many times more per token than a small one. GPT-6.1 Sol, for example, is available in Copilot on Pro+, Max, Business and Enterprise plans, and lists at $2 per million input tokens and $10 per million output tokens through OpenAI’s API.

Agents multiply tokens. An agent that reads files, runs tests and iterates sends the growing context back to the model on every step. A single agent task can consume tens or hundreds of thousands of tokens.

Cached tokens still count. Usage includes input, output and cached tokens, though cached tokens are typically priced lower. Long repeated context is cheaper than fresh input but not free.

Completions do not count. Inline suggestions remain unlimited. A developer who never uses chat or agents can stay well inside any allowance.

Six Costly Copilot AI Credits Traps

  1. Budgeting from the promotional months. June to August usage looked affordable because allowances were inflated. Ecorpit’s advice was to divide your expected usage by the September pool, not the August one.
  2. Defaulting to the most expensive model. Developers often pick the strongest model for every chat. For quick questions, a cheaper model costs a fraction of the credits.
  3. Unbounded agent runs. An agent left to iterate on a stubborn bug can burn through a month’s allowance in an afternoon.
  4. Ignoring the hard stop. Under the new approach, running out of credits can stop the service unless overage is enabled. Yahoo Finance’s coverage made that point directly. A team that runs dry mid-sprint loses its tool at the worst moment.
  5. Forgetting seat prepayment. Since September and October, seats must be paid upfront. Adding contractors mid-month now needs procurement planning.
  6. Treating all seats the same. Usage varies enormously between developers. A few heavy agent users can consume most of a pooled budget.

Forecasting Copilot AI Credits Spend

A workable forecast takes four steps.

First, pull usage reports from the past month and separate completions from credit-consuming features. Only the latter matter for the bill.

Second, group developers by behaviour: completion-only users, occasional chat users and heavy agent users. Most teams find a long tail of light users and a small group responsible for most of the spend.

Third, estimate monthly credits per group at the September allowances. Light users will sit well under 1,900 credits on Business. Heavy agent users may need several times that.

Fourth, decide how to cover the gap: overage spending, a higher plan for the heavy group, or usage policies that steer them toward cheaper models. Enterprise DNA recommended recalculating every coding assistant’s cost using expected agent requests and required platform seats rather than promotional credits or headline pricing. That is the right frame.

Controlling Copilot AI Credits

FindSkill highlighted the most useful default: a $0 overage cap that stops surprise bills. With the cap at zero, users cannot exceed their included Copilot AI Credits. That prevents overspending at the cost of interrupting heavy users when they hit the limit.

For most organisations, a better setting is a modest overage budget with alerts. It keeps work flowing while making overspend visible. Review the alerts weekly and talk to the people driving them; often a change of default model solves most of the problem.

Policy helps too. Recommend a cheaper default model for chat, reserve premium models for complex tasks, and set expectations for agent use, such as reviewing progress after a fixed number of iterations rather than letting runs continue indefinitely.

Copilot AI Credits vs Rival Coding Tools

Usage-based billing is now the norm across coding assistants. Anthropic includes Claude Code with its $20 Pro plan and offers Max plans at $100 and $200. Google meters Antigravity through its AI Pro, Ultra and top plans. Meta’s Muse Code starts at $5 a month with usage limits. Cursor and others use their own credit schemes.

The structure GitHub chose is distinctive in one way: credits are tied directly to published API rates. That makes Copilot AI Credits transparent in principle, because you can calculate what a model call should cost. In practice, few developers know how many tokens a given agent task will use, so the transparency only helps if someone on the team does the arithmetic.

A Worked Example for a 20-Person Team

Numbers make this concrete. Take a hypothetical team of twenty developers on Copilot Business at $19 a seat, so $380 a month and 38,000 included credits in total if allowances are pooled at the organisation level. Check your own settings, because pooling rules determine whether one heavy user can draw on others’ unused credits.

Suppose twelve developers mostly use completions and occasional chat, using perhaps 300 credits each. Six use chat heavily and run small agent tasks, at around 2,000 each. Two run long agent sessions most days and use 8,000 each. That adds up to 3,600, 12,000 and 16,000, or 31,600 credits in total.

Under the promotional allowance of 3,000 per seat, the team had 60,000 credits and plenty of headroom. Under the standard 1,900, it has 38,000, still enough on these assumptions, but the margin is much thinner. One more heavy agent user, or a busy release month, pushes the team over. These figures are illustrative, not measured, and your mix will differ. The method is what matters.

A Fair Copilot AI Credits Policy for Your Team

Credit-based billing works best when developers understand it, so a short, plain usage policy saves a lot of friction. It does not need to be long. One page covering a few points is enough.

State the default model for everyday chat and explain why. Developers are more willing to accept a cheaper default when they know the cost difference and know they can switch for hard problems.

Describe when premium models make sense: multi-file changes, unfamiliar codebases, tricky debugging. Give examples from your own work rather than abstract rules.

Set expectations for agents. A reasonable rule is to check in after a set number of iterations or a set time, and to stop and rethink if an agent has not made progress. Unbounded runs are where budgets disappear.

Explain what happens at the limit. If you use a hard cap, say so, and tell people how to request more. If you allow overage, say how much and who approves it. Surprises cause more frustration than limits do.

Finally, share the numbers. A monthly summary of usage by team, without naming individuals, helps everyone see where credits go and makes the policy feel like shared housekeeping rather than surveillance.

Copilot AI Credits for Individuals and Open Source Maintainers

Most coverage focused on companies, but individual developers felt the change too. On Copilot Pro, the $10 plan includes 1,000 Copilot AI Credits a month, which is $10 of model usage on top of unlimited completions. For someone who uses chat a few times a day with a mid-priced model, that is usually enough. For someone who runs agents on side projects every evening, it may last a week.

The simplest way to stretch a personal allowance is model choice. Use a small, cheap model for quick questions and explanations, and switch to a premium model only for tasks where it clearly helps. The difference in credits per conversation can be large.

Keep agent sessions focused. Give the agent one well-defined task, a clear success check such as a passing test, and stop it when it starts circling. Long, wandering sessions are the fastest way to empty a personal allowance.

Check your usage page weekly for the first month. Most people have no intuition for how many tokens an agent task uses, and a week of real data replaces guesswork. After that, a monthly glance is usually enough.

If you maintain open-source projects, check whether you qualify for any free or discounted access GitHub offers maintainers, and whether those programmes were affected by the move to Copilot AI Credits. Terms for these programmes change, so confirm them on GitHub’s current pages rather than relying on older announcements.

Who Should Use What

If you are Start with Why Link
Mostly using completions Copilot Pro or Business Completions stay unlimited See the changes
A heavy agent user Enterprise or a modest overage budget Larger allowance, fewer interruptions Forecast your bill
Managing a team budget A capped overage with weekly alerts Visibility without hard stops See controls
Comparing tools Claude Code or Google Antigravity trials Different plan structures Compare pricing

Related on this site: for more context, see Google Antigravity Six Months After 2.0, Claude Code vs Codex vs Cursor and Cursor Security Incident.

Frequently Asked Questions About Copilot AI Credits

What are Copilot AI Credits?
GitHub’s usage-based billing unit for Copilot since June 1, 2026. One credit is worth $0.01, and credits are deducted based on tokens used, priced at the underlying model’s API rate.

Are code completions still free?
Yes. Inline completions remain unlimited and do not consume credits on paid plans.

How many credits does each plan include?
Pro includes 1,000 credits, Business 1,900 and Enterprise 3,900 per seat per month, matching the seat price in dollars. Business and Enterprise had boosted allowances of 3,000 and 7,000 until September 1.

What happens if we run out?
Without overage enabled, credit-consuming features stop until the next cycle. You can set an overage budget to keep working, or a $0 cap to prevent any extra charges.

What changed on October 1?
Existing Business and Enterprise customers began paying upfront for seat assignments. New seat assignments had required prepayment since September 1.

Does Copilot Free still exist?
Yes. The free tier continues with limited usage.

Verdict

If your team mostly uses completions, Copilot AI Credits change very little and the seat prices are unchanged. If your developers rely on agents and premium models, rebuild your budget from the September allowances and set an overage policy before the next sprint, not after it.

The caveat is that GitHub has recalibrated Copilot pricing twice in a year and describes part of the allowance as flexible. Plan figures here reflect published documentation at the time of writing and may change.

Next step: export last month’s usage, split developers into light, moderate and heavy groups, and set a default model and overage cap that fit each group.

Sources and Further Reading

On this site: compare tools in AI coding assistants in 2026, cut model costs with LLM API cost optimization, connect tools through the best MCP servers, and secure agents with prompt injection and agent security.

2 Comments

Comments are closed.