Connect with us

AI

GitHub Copilot’s Best Month Came From Billing the Agents

GitHub Copilot’s record June came from token billing of coding agents, and Microsoft later said Copilot revenue jumped over 60% quarter over quarter.

Published

on

GitHub Copilot logged its best month ever in June after the company began billing the tool by the token on June 1, 2026. Chief technology officer Vladimir Fedorov told staff on June 24 that June was “by far our best month ever,” then declined to share figures because Microsoft’s quarter was about to close. Satya Nadella filled in the blank on the July 29 earnings call: Copilot revenue accelerated over 60% quarter over quarter.

The wire version treats that jump as a pricing win against Cursor, OpenAI’s Codex, and Anthropic’s Claude Code. The customer GitHub finally started charging is the coding agent that was already hammering the host.

GitHub Stopped Rationing Copilot on June 1

On April 27, 2026, GitHub chief product officer Mario Rodriguez announced the company was moving Copilot to usage-based billing on June 1. Premium request units gave way to GitHub AI Credits, drawn down by input, output, and cached tokens at each model’s listed API rate. Seat prices did not move. Copilot Pro stayed $10 a month, Pro+ $39, Business $19 per user, and Enterprise $39 per user.

Rodriguez wrote that Copilot was no longer the in-editor assistant of a year earlier. It had become an agent platform that runs long, multi-step sessions across a whole repo. Under the old counters, a short chat and a multi-hour autonomous run could cost the same, and GitHub had been eating the inference bill. Usage-based billing, he said, “reduces the need to gate heavy users.”

That is the mechanism behind Fedorov’s record month. GitHub did not invent a new crowd of human subscribers in 23 days. It stopped capping the shops that already ran agents all day, then sent them a meter. Monthly Pro and Pro+ accounts moved on June 1 on their own. Annual plans stayed on request pricing until they expired. Code completions and next edit suggestions stayed unlimited on paid plans and still do not touch the credit pool.

The Agents Were Already on the Platform

The load showed up well before the new bill. On April 3, GitHub chief operating officer Kyle Daigle answered a post from the developer and streamer ThePrimeagen, who had thanked GitHub for absorbing “the amount of shit code that has been added over the last 3 months,” including “10s of billions of lines of code that will never see the light of a CPU.”

Yup, platform activity is surging. There were 1 billion commits in 2025. Now, it’s 275 million per week, on pace for 14 billion this year if growth remains linear (spoiler: it won’t.) GitHub Actions has grown from 500M minutes/week in 2023 to 1B minutes/week in 2025, and now 2.1B minutes so far this week. So we’re pushing incredibly hard on more CPUs, scaling services, and strengthening GitHub’s core features. And as a fine purveyor of hand-crafted shit code for many years, I’m not gonna weigh in on that.

Kyle Daigle, GitHub COO, on X

Peter Steinberger, who built the OpenClaw agent stack, posted the same day that he kept hitting GitHub API quota limits and that the interface “hasn’t been designed with agents in mind.” GitHub has been around since April 10, 2008. Microsoft bought it in 2018. The people who designed commit, pull request, and REST limits were counting human hands, not loops that open pull requests while you sleep.

THE YEAR THE METER CAUGHT UP

  1. April 3, 2026: Daigle posts 275 million commits a week and 2.1 billion Actions minutes in that week, after 1 billion commits in all of 2025.
  2. April 27, 2026: Rodriguez says agent sessions made the old request caps unsustainable and sets June 1 for token billing.
  3. June 1, 2026: Monthly Copilot plans move to AI Credits. Business and Enterprise accounts get promo credit piles for June through August.
  4. June 24, 2026: Fedorov tells staff June is the best month GitHub has ever had and will not discuss the figures.
  5. July 29, 2026: Nadella says Copilot revenue accelerated over 60% quarter over quarter after the new model went in.
  6. August 17, 2026: A new traffic peak knocks the Central US data center over for 7 hours and 47 minutes, including Copilot.

By September, some Copilot users were still writing their own command-line tools so project boards and task flows would stop chewing credits. The meter did not scare the traffic off. It taught the heavy users to watch the burn.

What a GitHub AI Credit Buys

One GitHub AI Credit equals $0.01. At launch, included credits matched the sticker price dollar for dollar. Current GitHub docs add a flex pile on individual plans, which is why a $10 Pro plan now includes 1,500 AI credits a month on Pro, not 1,000. Unused credits do not roll. The pile resets at 00:00:00 UTC on the first day of each calendar month, not on your personal billing date.

CURRENT COPILOT CREDIT ALLOWANCES

Plan Monthly price Included AI credits How the pile works
Copilot Pro $10 1,500 1,000 base plus 500 flex
Copilot Pro+ $39 7,000 3,900 base plus 3,100 flex
Copilot Max $100 20,000 10,000 base plus 10,000 flex
Copilot Business $19 per user 1,900 per user Shared org pool
Copilot Enterprise $39 per user 3,900 per user Shared org pool

Those individual totals are worth $15, $70, and $200 of model use at the $0.01 rate. For companies, credits are pooled at the billing entity level. A shop with 100 Business seats gets 190,000 credits in one bucket, so a few agent-heavy seats can draw what light users leave behind. Adding seats grows the pool at once. Removing seats does not shrink it until the next cycle. Existing Business and Enterprise accounts also got a promo of $30 and $70 in monthly credits for June, July, and August, then dropped back to the standard piles.

Admins can set budgets at user, cost center, org, and enterprise level. A $0 user budget blocks that person even if the company pool still has credits. When the pool is gone, the company either pays published rates or stops. There is no fallback to a cheaper model. Paid users get a 10% discount on model cost if they leave Copilot on auto model selection in chat, the CLI, the Copilot app, or the cloud agent.

WHAT DRAWS DOWN AI CREDITS

  • Chat and CLI: Copilot Chat and Copilot CLI spend credits on every model call.
  • Cloud agent runs: Long Copilot cloud agent jobs across many files drain the pile faster than a short question.
  • Spaces, Spark, and outside agents: Copilot Spaces, Spark, and third-party coding agents sit on the same meter.
  • Code review: Copilot code review spends AI Credits and GitHub Actions minutes at the normal per-minute Actions rate.
  • Still free on paid plans: Inline completions and next edit suggestions do not use AI Credits and stay unlimited.

That split is the real price list. Ghost-text autocomplete is still a seat product. Anything that looks like an agent is a token product. Teams that assign GitHub issues to Copilot agents are on the metered side of that line, which is why a project board can empty a monthly pile while the editor’s gray suggestions keep coming.

Microsoft Put a Number on That Silent June

Fedorov’s June 24 meeting sat six days before Microsoft’s fiscal quarter ended on June 30, which is why he would not “talk about the numbers.” Microsoft declined to comment on those internal remarks. Nadella did not stay quiet on the July 29 call.

THE JUNE GITHUB WOULD NOT QUANTIFY

  • Copilot revenue: Accelerated over 60% quarter over quarter after usage-based billing started in June, Nadella said.
  • Copilot users: 50 million, with business and enterprise seats still growing plus “significant consumption revenue” once the new model was live.
  • GitHub users: 225 million, including over 90% of the Fortune 500, on the same call.
  • Fedorov’s line: June was “by far our best month ever,” with no dollar figure attached.

Nadella’s phrasing matters. He did not only cite more seats. He cited consumption after the meter turned on. That is agent traffic becoming a line item, which is what Rodriguez said the old request caps could not hold.

Heavy Agent Sessions Drain the Credit Pool

Replies under GitHub’s April 27 announcement treated the meter as an eviction. People talked about quitting, about going back to typing, about a prompt that would cost more than a lunch. The June result and the over 60% revenue jump say the heavy users did not leave. They paid. Light users who live on autocomplete barely felt it, because that path never entered the credit ledger.

The bill now tracks how long the model works, which model you pick, and how much context you stuff into the run. A frontier model on a repo-wide agent session spends the pile in a way a 12-word chat never will. Companies that leave additional usage on will see Copilot show up like an API invoice. Companies that cap budgets will watch agents stop mid-task. Either way, the human seat fee is no longer the whole story.

Rivals already billed that way, which is why GitHub moved. Matching Cursor, Codex, and Claude Code on consumption did not just “align pricing.” It let GitHub keep the agent shops that request caps were about to throttle, then charge them for the tokens those shops were already burning.

Monthly Commits Climbed From 1.4 Billion to 2.9 Billion

The same loops that filled the credit meter also filled the git log. In his August 20 note, Fedorov wrote that monthly commits had grown from 1.4 billion in April to 2.9 billion. That is the April agent wave continuing after the “best month,” not a one-month spike that faded when the promo credits hit.

On August 17 the site took an outage that lasted 7 hours and 47 minutes. It hit github.com, sign-in, Actions, APIs, pull requests, issues, and Copilot. Fedorov said a new traffic peak met a Central US component that would not scale, then Copilot client retries made recovery worse. It was the second bad August incident after an Actions failure on August 6. “Both incidents were capacity failures at their core,” he wrote. “If you were trying to ship software that day, we let you down.”

GitHub added more than 3 million CPU cores and 120 petabytes of high-speed storage, and it pushed more of the platform onto Azure. Fedorov said Azure now serves roughly 58% of platform load and half of all Git operations, up from 12% of platform load in May. Microsoft also looked for extra cloud capacity after the June surge while that move was still underway. The host that stores the code, runs the Actions minutes, and meters the Copilot tokens is the same machine, and the agents do not wait for it to catch up.

Fedorov Still Sees No Need for a Big Price Hike

On June 24, with the record month in hand and the quarter still open, Fedorov said he did not personally think GitHub needed to raise prices much. He would not confirm a plan. Seat prices on June 1 had not moved. The new money is the overage, the flex, the pooled org burn, and the agent sessions that used to look like one “premium request.”

That stance only holds if autocomplete stays a loss-leader and the token meter keeps catching the loops. It breaks if agent shops cap budgets and walk, or if another outage like August 17 makes those shops keep a second host. GitHub’s bet is that the people who already parked their repos there will pay for the agents on top, one credit at a time, while the gray ghost text remains free.

Nadella told investors the agentic era is being built on GitHub. Fedorov’s June remark, the over 60% Copilot revenue jump, and the climb from 1.4 billion to 2.9 billion monthly commits are the same fact in three registers: the agents showed up, the meter finally saw them, and the old host is still learning how to stand under them.

Harry is the editor of Oton Technology, an independent site he owns and edits, covering the part of technology that people actually have to act on. After ten years in journalism, first reporting and then editing, he works from primary material by habit: the advisory rather than the write up of it, the filing rather than the press release, the changelog rather than the launch video. Every figure in an article carries its source and its date, and where a number comes from a vendor or an analyst model rather than a count, he says so plainly instead of letting it stand as established fact. What he leaves out is anything he could not verify himself, which on a beat full of unnamed supply chain claims removes a great deal. That standard applies across all the sections the site publishes for an international audience, from artificial intelligence and security to phones, computers, gaming, crypto and the software businesses depend on. He corrects errors in the open and labels them, because a site that hides its mistakes is asking readers to trust the rest on nothing.

Continue Reading
Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending