What Happened
GitHub flipped a switch today that quietly resets the economics of paid coding assistance. Effective June 1, every paid GitHub Copilot plan moves off the legacy "premium requests" model and onto a unified token-metered system called GitHub AI Credits, where one credit equals exactly $0.01 USD and usage is calculated from input, output, and cached tokens at each model's published API rate.
The headline plan prices stay flat: Copilot Pro is still $10 a month, Pro+ stays at $39, Business holds at $19 per seat, and Enterprise remains $39 per seat. What changes is what those dollars buy. Pro subscribers now get 1,000 included credits a month, Pro+ users get 3,900, Business seats receive $19 worth of credits and Enterprise seats receive $39 worth. The flat-fee monthly cap on "premium" interactions, which had previously translated into a fixed number of model requests, has been retired in favor of a metered pool that drains with every paid model call. Code completions and the Next Edit Suggestions experience remain free across all plans and do not consume credits, which preserves the core daily workflow most developers actually rely on.
The shift was telegraphed in late April through the GitHub Blog and Microsoft's developer communications, then confirmed with the technical billing docs that went live in May. As of today, Business and Enterprise customers are automatically enrolled in a three-month promotional included-usage window that runs through the end of August, designed to soften the migration. Admins gain new budget controls at the enterprise, cost-center, and individual-user level, and a new "pooled included usage" feature lets unused credits flow across a team rather than expiring per seat. When the pool is exhausted, organizations choose whether to throttle, cap spend, or allow paid overage at the published per-token rate for each model.
Why It Matters
This is the moment the AI coding industry stops pretending that frontier models can be sold like flat-fee productivity software. For three years the dominant pricing motion across Copilot, Cursor, Codeium, and a long tail of competitors has been a familiar SaaS shape: a single monthly fee, an implicit usage cap, and the silent assumption that average users would subsidize power users. Today's transition makes that subsidy explicit and then dismantles it. The cost of inference has not fallen fast enough, and the average tokens-per-task generated by agentic workflows has risen too quickly, for the old math to keep working.
The economic logic behind the change is straightforward. A single complex agent run on a frontier model like Anthropic's Claude Opus 4.7 or OpenAI's GPT-5.5 can burn through $0.50 to $2.00 of credits depending on context length and tool-call depth. A user who runs ten such tasks a day on Pro+ is consuming roughly $5 to $20 of inference daily against a $39 monthly subscription. Under the old request-based system, that user was paying the same as a colleague who fired Copilot Chat twice a week. Under the new credit system, heavy users will be visibly metered and either capped, throttled, or asked to top up. GitHub's framing positions this as transparency; the structural reality is that flat-fee economics never made sense at frontier inference costs, and the entire industry was waiting for the largest player to move first.
For competitors, the timing is delicate. Cursor's Composer 2.5 dropped on May 18, Codex Pro's two-times promotional multiplier expired May 31, and DeepSeek made its 75-percent V4-Pro discount permanent on May 22. Each of those moves reshapes the per-token landscape in which GitHub's new credit system will be benchmarked. The most-cited number on developer forums is the implied effective rate: 1,000 credits for $10 buys roughly the same volume as 5 to 8 hours of moderate agent work on a mid-tier model, depending on prompt discipline. That is enough for individual contributors who use Copilot as an editor companion, but materially short of what production agentic workflows demand. The market is, in effect, being split into editor-assistant pricing and agent-orchestrator pricing, and today is the day that split becomes visible on the invoice.
Reaction
The GitHub Community discussion thread for the announcement crossed into the thousands of comments within days of the April reveal, and the tone in the run-up to the June 1 cutover has been markedly cooler than GitHub probably hoped. Three concerns dominate. The first is opacity around per-token rates: while the rates exist in the docs, the practical translation from "an average prompt I write" to "credits consumed" remains hard to predict, and developers are uncomfortable being asked to budget against a unit they cannot intuit. The second is the loss of fallback behavior. Under the legacy system, exhausting premium requests would gracefully degrade to a lower-cost model; under the credit system, the workflow halts at the budget cap unless an admin or user explicitly allows overage.
The third concern is harder to dismiss as a pricing complaint. A number of vocal developers, including several visible on Visual Studio Magazine's coverage, argued that the transition delivers, in their phrasing, "less for the same price" for power users on annual Pro+ plans. Annual subscribers who pre-paid for a year of fixed access now face a usage meter mid-contract, and refund logistics have been a particular point of friction in the community thread. Enterprise architects raised a distinct concern: pooled budgets create a new political surface inside engineering organizations, where a small number of agent-heavy teams can drain shared credits and force throttling on colleagues who barely use the tool. GitHub's pre-built admin dashboards address the visibility side of this, but the underlying allocation question becomes a management problem rather than a product one.
Counterpoints exist. Several engineering leaders writing publicly noted that pooled usage with cost-center attribution is exactly what large organizations have been asking for, and that token-metered billing aligns Copilot with how teams already budget cloud spend. The promotional Business and Enterprise usage through August gives finance teams an empirical baseline before any real overages hit the books. For the median developer using Copilot mainly for autocomplete, today's change is invisible: completions and Next Edit Suggestions stay free, and most casual chat usage fits comfortably inside the included pool.
What's Next
Three trajectories will define the rest of the year. The first is competitive pricing pressure. With GitHub now the largest vendor charging per token, every smaller competitor has to decide whether to match, undercut, or differentiate on bundled value. Cursor's recent Composer 2.5 release pushed toward agent-first workflows at a flat fee, betting that frontier model gross margins improve fast enough to absorb usage growth. DeepSeek's permanent V4-Pro discount and Alibaba's Qwen 3.7-Max preview keep open-weight pressure on per-token rates from the supply side. The window for flat-fee unlimited agent access at frontier quality is closing rapidly, and the firms still offering it are doing so on borrowed runway.
The second trajectory is admin tooling. GitHub's new budget controls, cost-center allocation, and pooled usage across a business are the early pieces of what will become a full FinOps-for-AI discipline inside enterprises. Procurement teams that learned to manage cloud spend during the AWS era are now learning to manage token spend, and the vendors who provide the cleanest visibility and forecasting layers will capture disproportionate share. Microsoft's underlying decision to consolidate Copilot, Azure inference, and the broader developer stack on a common credit currency positions it well here, particularly as Anthropic explores running Claude inference on Microsoft's custom Maia 200 chips, which would tighten the per-token cost structure on Azure-hosted models.
The third trajectory is product behavior. When users see a meter, they prompt differently. Expect a measurable shift toward concise prompts, prompt templates, smaller context windows, and explicit model selection rather than always reaching for the most expensive frontier option. Tooling that helps developers reduce token consumption without losing output quality, including prompt compression, context pruning, and cheaper-model-first agent routing, becomes commercially attractive in a way it was not when the meter was hidden. The agent-first architectures that vendors have spent the past eighteen months building will be the first place this discipline is enforced, because their cost-per-task is the easiest to attribute.
Closing Thoughts
The transition that took effect today is less a pricing tweak than a category reset. For most of Copilot's history the product was sold as a fixed companion that lived inside the editor. The new credit system tacitly admits that the most valuable use of Copilot is no longer autocomplete but agentic execution, and agentic execution at frontier quality costs real money per task. GitHub did not raise headline plan prices, which preserves the perception of stability, but it pulled the cap off the underlying compute, which is the variable that actually determines profitability.
For the broader industry, the playbook is now visible. Vendors will keep the editor-assistant tier flat and inexpensive because that is where adoption lives, and they will move agent and orchestration features onto metered credits because that is where compute economics break. The companies that win the next phase will be the ones that make the meter feel fair, give engineering leaders forecasting tools, and reduce the marginal token cost of a useful task faster than competitors. Today's switch at GitHub is the starting gun for that race, and every other coding-assistant vendor now has to decide how quickly they follow.