GitHub Copilot's First Token Cycle Just Closed. The Developer Bill Came In at 10x to 50x.

GitHub flipped Copilot to usage-based billing on June 1, 2026. Today is the day the first full 30-day cycle closes. The screenshots are already circulating. Developers on the Pro and Pro+ plans, the ones who got told the headline subscription price is unchanged, are looking at projected charges that run an order of magnitude past what a request-based plan ever cost. One developer posted a $750 projection on a seat that used to cost $29. Another posted $3,000 on a seat that used to cost $50. The agentic users at the right tail are higher than that. As of yesterday, the trade press is calling it a confirmed 10x to 50x surge for heavy agentic seats.

The headline number is dramatic. The underlying mechanic is more interesting, and it is why we are writing this up the day the cycle closes instead of as part of the weekly roundup. GitHub did not raise the sticker price. GitHub stopped paying the meter on everybody's behalf.

The Meter Math

Every Copilot plan still includes a fixed monthly fee. On top of that, every plan now ships with a monthly allotment of GitHub AI Credits, where one credit equals one cent. Pro at $10 ships with 1,500 credits ($15 of meter at retail). Pro+ at $39 ships with 7,000. The new Copilot Max tier, launched alongside the billing change, costs $100 a month and ships with 20,000. Enterprise gets 3,900 credits per seat. Beyond the allotment, you pay the listed rate at the model's standard token price.

The model menu underneath is exactly what you would price out of an API console. GPT-5.5 runs $5 input, $30 output per million tokens. GPT-5.4 runs $2.50 input, $15 output. Claude Opus 4.5 through 4.8 runs $5 input, $25 output, plus cache-write. Claude Sonnet runs $3 input, $15 output. Gemini 2.5 Pro runs $1.25 input, $10 output. Microsoft's house MAI-Code-1-Flash runs $0.75 input, $4.50 output and is the cheapest serious model on the list, which is not an accident. Code completions and the next-edit suggestion stream stay free, because those still run on small models GitHub is comfortable eating the cost on.

One agentic task on Claude Opus 4.7 with a 40,000-token diff is running 60 to 100 credits per interaction. Call that 80 credits on average. A developer running ten of those a day on twenty working days burns 16,000 credits in a month. On a Pro+ plan, the first 7,000 come included; the next 9,000 cost 9,000 cents, or $90, on top of the $39 subscription. Push that workflow to all-day autonomous agent runs and the meter compounds. The 10x and 50x numbers are not a misprint. They are what happens when somebody who used to send 600 chat messages a month under the old request cap sends the equivalent token volume on a Claude Opus run.

Why GitHub Did This

Under the old request-based plan, GitHub absorbed the inference cost differential between a one-line completion and a multi-hour agentic loop. Both counted as one request. With coding agents now writing pull requests instead of completing variable names, the cost spread between the cheapest and the most expensive request inside a single Copilot seat ran to four orders of magnitude. GitHub was the one eating it. Microsoft was the one paying Anthropic and OpenAI for the tokens underneath. The math stopped working at the unit-economics level, and the company said so in the announcement.

The decision is rational. The optics are bad. GitHub gets to keep the $10 Pro sticker price for marketing copy and let the developer's own usage decide whether the seat costs $10 or $750 a month. The flat fee is now a floor, not a ceiling. That is exactly how every cloud bill works, and exactly how no developer-tooling seat has worked before.

The Buyer-Side Cliff Just Hit the IC Seat

Three days ago we wrote up the tokenmaxxing cliff in our IPO math piece: enterprise buyers are done burning unbounded AI budget and have started capping the line item. Uber capped Claude Code at $1,500 per employee per month per tool after the 2026 AI budget melted in four months. Lindy moved its entire production stack from Claude to DeepSeek. Vercel watched DeepSeek's share of token volume on its AI Gateway jump from under 1 percent to 17 percent in May while DeepSeek's share of spend stayed near 1 percent. The cliff was on the enterprise side. The Copilot meter change is the same cliff a level down, and it lands on the individual contributor instead of the CIO.

What an enterprise buyer can do is renegotiate, cap by seat, swap models, and route through a gateway. What a Copilot Pro user can do is open the model picker and choose a cheaper model on the next prompt. That is the part of this story that matters for pricing. Inside the first 30 days of the new billing regime, every Copilot seat is now a model-shopper. The free completions stay free. The chat box now has a per-token cost stamped on it that the developer reads in their own dollars.

Three concrete behaviors are visible in the cycle that just closed. First, model downshift. Developers are switching from Opus 4.8 to Sonnet 4.6 or Gemini 2.5 Pro for tasks that do not need the top tier, because the price gap is now legible at the seat level. Second, harness substitution. Cursor, Claude Code, Aider, and Codex CLI are taking inbound from Copilot users who do their own routing and pay the upstream API rate. Third, MAI uptake on the Copilot surface itself. Microsoft's own coder model is the cheapest model in the picker, and people are clicking it.

What It Does to the Pricing Floor

The inference price floor we track is set by the marginal cost of a token at the API layer. The Copilot change does not move that floor. It exposes it. For the past three years the developer surface on top hid the per-token cost behind a flat-rate subscription, and the subsidy made the price floor a thing you read about in our coverage instead of a thing you felt in your credit card bill. That subsidy is gone now. Every developer with a Copilot Pro seat is, starting this month, a token economist whether they signed up to be one or not.

That is what makes today different from a normal pricing change. The Copilot installed base is somewhere between 15 and 20 million paid seats. Even if 10 percent of them sit in the heavy-agentic tail of the distribution, that is one to two million developers who just got a price signal that the floor matters. Open-weight models like GLM 5.2 and DeepSeek V4 land at one fifth the cost of frontier closed models on agentic benchmarks, which we covered in the tokenmaxxing piece and the GLM 5.2 piece. Until this week, the developer-tooling buyer rarely saw that gap in their own invoice. Now they do.

What This Does to Anthropic and OpenAI

The mixed read. Anthropic and OpenAI are the two suppliers underneath the most expensive Copilot interactions, which means they are the suppliers whose token revenue grows when a Copilot seat goes over the credit allotment. That is the immediate revenue tailwind. GitHub is now the highest-volume customer-facing meter on top of both stacks, and a Copilot bill shock translates to higher API draw at the source.

The medium-term read is the one that matters for IPO math. Both labs have S-1 paperwork in motion at run-rates that assume token volume keeps doubling. The Copilot meter is the first time millions of mid-tier developers see what an Opus call costs in their own currency. Every developer who switches from Opus to Sonnet, or from Sonnet to MAI, on the next prompt is a per-seat price compression event that the closed-frontier revenue forecast did not assume. The doubling curve does not stop tomorrow. The doubling curve starts having a model-mix problem the labs cannot fully control, because the meter that drives it is now visible to the seat.

Our Take

The Copilot meter change is the most important pricing event in developer tooling this year, and the bill cycle that closed today is the moment it stops being a thread on Hacker News and starts being a budget conversation inside every engineering org. It is not a Copilot problem. It is the first time the marginal cost of agentic inference is showing up on an individual contributor's invoice instead of inside a hyperscaler gross margin line. We have been writing about that price floor for a year. The meter just made it personal.

Practical implication for builders. If you ship a paid product on top of a hosted coding agent, your customer-facing meter has to assume the buyer can read it. The Copilot template is now the public reference for how developer tools price agentic compute, and it ships with a model picker, a credit balance, and a per-call cost in the UI. That is the surface every comparable product is about to copy. We audited our own paid API for the same reason in February, and the kindest thing you can do for an agent customer is not surprise them on the invoice. The pricing-floor work we have published is the long-form version of that lesson.

The next 30 days are the ones to watch. The second billing cycle closes on July 30, by which point seat-by-seat substitution behavior will be measurable inside GitHub's own model-mix telemetry. If MAI usage spikes on Copilot, Microsoft has bought itself an intra-stack hedge against the Anthropic and OpenAI bills it is paying. If Claude Sonnet eats Opus share, the buyer-side discipline cliff we wrote about is on the IC seat permanently. If most seats stay on Opus and just pay the overage, the doubling curve holds. The cycle that closed today is the first datapoint. The one that closes a month from now is the trend.

We are watching the Copilot model-mix disclosure inside any forthcoming GitHub blog post on cycle two, the next quarterly Microsoft earnings call (where the Intelligent Cloud line will quietly absorb the Copilot revenue delta), and any movement on the Anthropic and OpenAI gross margin commentary that lands in the confidential S-1 follow-up files. Cross-reference our inference money piece for the broader rotation: capital is flowing into the serving layer, the buyer is now reading the meter, and the seat priced in dollars is the new front line.