GitHub Copilot AI Credits: How Token Billing Works and What It Costs

Photo by James Harrison on Unsplash

On June 1, 2026, GitHub Copilot stops counting premium requests and starts counting tokens. The shift is described in GitHub's official announcement, which confirms that every paid Copilot plan moves from premium request units (PRUs) to usage-based billing measured in GitHub AI Credits. Base subscription prices do not change. What changes is how quickly you can burn through what your subscription covers.

That distinction matters more than the headline suggests.

What an AI Credit actually is

A GitHub AI Credit is a unit of account, not a flat allowance of requests. GitHub's pricing documentation sets the conversion at 1 AI Credit = $0.01 USD, and quotes every model rate per million tokens of input, cached input, cache write, and output. A token, as AI News explains, is roughly three-quarters of an English word, so a 10,000-word codebase pasted into a single query consumes around 12,000 to 13,000 input tokens before the model has written a single character of response.

The practical effect: your bill now depends on which model you pick, how long your prompts are, how much context the agent loads, and how much the model writes back. None of those variables existed in the PRU world, where a multi-hour agentic task and a one-line question both counted as a single premium request.

What stays free, and what doesn't

Not everything in Copilot now meters tokens. According to GitHub's announcement, code completions and Next Edit Suggestions remain included in every paid plan and do not consume AI Credits. They stay unlimited. GitHub also designates GPT-4.1 and GPT-5 mini as included models that draw no credits, per the pricing documentation.

Everything else is metered. Chat with frontier models, agent runs, Copilot Workspace tasks, and Copilot code review all draw against your credit balance. Code review carries an additional cost: it also consumes GitHub Actions minutes, billed at the standard Actions per-minute rate, on top of whatever model it invokes. GitHub does not disclose which model code review picks, which makes per-task costs there genuinely unpredictable.

The model price list, in plain numbers

The range between the cheapest and most expensive model on Copilot is wide enough to matter for budgeting. From the pricing page:

• GPT-5.4 nano is the cheapest OpenAI option at $0.20 per million input tokens, $0.02 per million cached input, and $1.25 per million output tokens.
• GPT-5.5 sits at the top of the OpenAI lineup: $5 per million input, $0.50 per million cached input, and $30 per million output.
• Claude Haiku 4.5 is Anthropic's cheapest available tier at $1 per million input, $0.10 per million cached input, $1.25 per million cache write, and $5 per million output.
• Claude Opus 4.5 through 4.7 is the single most expensive option on the menu: $5 per million input, $0.50 per million cached input, $6.25 per million cache write, and $25 per million output.

Two footnotes are easy to miss. GPT-5.4 pricing only applies to prompts at or below 272K tokens, and Gemini 2.5 Pro and Gemini 3.1 Pro pricing only applies to prompts at or below 200K tokens. Cross those thresholds and the per-token math changes.

Output tokens are the expensive ones. A long agentic run that emits thousands of lines of generated code, reasoning traces, and tool calls will rack up output charges far faster than a chat session that returns a few paragraphs.

How much your plan actually covers

GitHub kept subscription prices flat, which is the part of the announcement most likely to mislead casual readers. GitHub's announcement confirms Copilot Pro remains $10 a month, Pro+ remains $39 a month, Business remains $19 per user per month, and Enterprise remains $39 per user per month.

The credit allowance attached to those plans is small. Reporting from EnterpriseDNA puts Copilot Pro's monthly allowance at $10 in AI Credits and Pro+ at $39 in AI Credits, matching the subscription fee almost exactly. At Claude Opus output rates of $25 per million tokens, a Pro user's entire monthly credit pool buys roughly 400,000 output tokens before overage kicks in. At GPT-5.5 output rates, the same $10 buys about 333,000 output tokens.

Credits do not roll over. EnterpriseDNA confirms that any unused balance is forfeited at the end of the billing cycle, which removes the option to bank a quiet month against a busy one.

The buffer that disappeared

The gap between the old system and the new one is larger than the price tags suggest. AI News notes that under PRUs, users could routinely spend three to eight times the token value their subscription covered without triggering any penalty, because a request was a request regardless of how much compute it consumed behind the scenes.

That buffer is gone. So is the fallback experience. GitHub's announcement confirms that the previous behavior, where a user who exhausted premium requests could fall back to a cheaper model and keep working, has been removed entirely. Usage is now governed strictly by credit balance and admin budget controls.

For anyone running long agent sessions, the change lands hard. A developer cited in the official community thread and quoted by EnterpriseDNA estimated that agentic coding sessions, where Copilot plans, researches, and executes multi-step tasks, routinely consume $30 to $40 per session. A Pro subscriber on the $10 plan can exhaust the month's allowance in a single working afternoon.

The agent problem nobody is naming

Here is the awkward part of the rollout. GitHub's cloud agent reached general availability earlier in 2026, and Microsoft spent the first half of the year promoting agentic coding as the direction of the product. The new billing model prices that exact workflow like raw compute. EnterpriseDNA flags this directly: the pricing change undercuts the value proposition of a feature GitHub itself was promoting as the future of development, from day one of its GA window.

The critique from developers, captured by TechCrunch, is sharp: Microsoft encouraged indiscriminate agentic use, then switched the pricing model in a way that penalizes precisely that pattern. GitHub Chief Product Officer Mario Rodriguez told TechCrunch that "Copilot is not the same product it was a year ago," which is true in both the direction GitHub means it and the one its users mean it.

The community reaction was unambiguous. EnterpriseDNA reports the official announcement thread drew more than 400 comments and close to 900 downvotes.

The bills people are reporting

Some of the numbers circulating are eye-catching. TechCrunch documents developers projecting monthly bills jumping from $29 to $750, and in one case from $50 to $3,000. Those figures almost certainly reflect heavy agentic workloads rather than typical chat-based use, but they illustrate the upper bound of what unconstrained agent use can cost when each token is metered.

The scale of real-world AI coding spend explains why GitHub made the change. The Information, as summarized by AI News, reported that Uber's CTO disclosed the company had already exhausted its 2026 AI budget, noting that 11% of updates to Uber's code are now written by AI. At that scale, flat-rate seats become a structural loss for the vendor.

TechCrunch also reports that developers are actively migrating to alternatives, including OpenAI Codex via direct API, Anthropic's Claude Pro at $20 a month flat, DeepSeek V4 Pro, and locally hosted open-source models.

What annual subscribers should know

The transition has a carve-out. GitHub's announcement confirms that annual Pro and Pro+ subscribers remain on legacy PRU pricing until their plan expires. When the annual term ends, they transition to Copilot Free unless they pick a new monthly plan. That makes the renewal decision, not the June 1 cutover, the moment the new economics actually arrive for those users.

GitHub also launched a preview billing dashboard in early May 2026 so users could see projected costs before the change took effect. Using it before renewal is the only reliable way to estimate what the new model will cost a given workflow.

How teams should think about budgeting

A few structural details from the pricing documentation and EnterpriseDNA's coverage matter for anyone administering Copilot at scale.

Business and Enterprise plans pool credits org-wide rather than isolating them per user. A team with a handful of heavy agent users and many lighter chat users will get more out of the pool than the same headcount on individual Pro seats. Admin budget controls become the actual lever for cost containment, since the fallback experience no longer exists. And because Copilot code review consumes both AI Credits and GitHub Actions minutes, the cost per pull request is the sum of two separate meters running against two separate budgets.

The single biggest budgeting trap is choice of model. Switching a default workflow from GPT-5.5 to GPT-4.1 (which is included) or to GPT-5.4 nano (priced at launch.20 per million input tokens) can change a team's monthly Copilot spend by an order of magnitude without changing how anyone works day to day. Routing agent runs to Claude Opus, by contrast, is the fastest way to drain a credit pool, with output tokens at $25 per million and cache writes at $6.25 per million on top.

The new model is not strictly worse. It is honest about what each interaction costs. What it removes is the slack that made Copilot feel like a flat-rate productivity tool. From June 1, 2026, it is a metered API with a chat interface bolted on, and the meter does not stop running while the agent thinks.