Copilot's Meter Is On: What Usage-Based AI Credits Cost You

TL;DR

• What changed: Copilot now bills on usage-based AI Credits (1 credit = 1 cent), metered on the tokens an agent burns, with per-user budgets and a new Copilot Max tier. Live June 1, 2026.
• The real story: sticker prices barely moved. The metering itself is the tell that agentic coding became an infrastructure cost, not a flat-fee feature.
• What drives the bill: output tokens and model choice. Routing work by model tier is the main lever for keeping spend sane.
• The catch: a visible meter makes developers code defensively. That cuts both ways, disciplined cost control or just reaching for the agent less.

On June 1, 2026, GitHub stopped pretending agentic coding fits inside a flat fee. Copilot moved off premium request units and onto something called GitHub AI Credits, where one credit equals one cent and your bill tracks the tokens an agent actually burns.¹² The sticker prices barely budged. Pro is still $10 a month. What changed is that the meter is now running, and that single fact tells you more about where AI coding is headed than any pricing table.

Here is the thing most of the coverage missed in the rush to post burn-rate horror stories. The numbers are not the story. The metering is the story. When a vendor that spent three years selling unlimited flat-rate access turns on a meter, it is admitting that the product underneath the subscription stopped being software and started being compute. And compute has a meter, the same way your AWS bill and your CI minutes do.

I want to walk through what actually changed, why GitHub did it, what it costs a real developer, how people are already changing their behavior, and the part that matters most: whether any of this is GitHub uniquely extracting rent, or an industry-wide correction that was coming no matter who shipped it first.

What Actually Changed

Under the old model, Copilot billed in premium request units. Each interaction cost one unit, scaled by a model-power multiplier, and that was that.¹³ A one-line autocomplete and a forty-minute autonomous refactor could land on the same line of your invoice. The new model bills on real token consumption: input, output, and cached tokens, priced at published per-model API rates.¹²

The conversion rate is deliberately legible. One AI Credit is one US cent.² That makes the math easy and, I suspect, makes the meter feel more present. You are no longer spending abstract “units.” You are spending pennies, and you can watch them go.

The split between what is metered and what is free matters more than the headline. Code completions and Next Edit Suggestions stay unlimited on every paid plan and consume zero credits.² Everything agentic draws from the pool: chat, the CLI, cloud agents, Spaces, Spark, and Copilot code review.² Code review is the one to flag for teams, because it consumes GitHub Actions minutes on top of AI Credits. That is a dual meter, billed separately to the repository.⁴

The plan structure as of June 1:⁵²

Plan	Price	Included credits	Approx. value
Pro	$10/mo	1,500	~$15
Pro+	$39/mo	7,000	~$70
Business	$19/user/mo	1,900	varies
Enterprise	$39/user/mo	3,900	varies
Max	$100/mo	20,000	~$200

Copilot Max is the new tier, aimed at high-volume power users with priority model access.⁵ There is a catch for anyone who wants to sign up cold: new Max signups were paused on April 20, 2026, and as of June 1 it is available only as an upgrade for existing subscribers. New Pro, Pro+, and Student signups were paused too, with GitHub promising a return “in the coming weeks” and no firm date.⁵

One more change with teeth: per-user budget caps went generally available, and a zero-dollar budget on a user immediately blocks that user’s access.⁴⁶ For organizations, credits pool at the org or enterprise level instead of per seat. A hundred Business users share 190,000 credits as one bucket, not a hundred separate ones.⁷

Why Vendors Are Moving to the Meter

GitHub’s own explanation is the most honest thing in the announcement, and it is worth reading as a confession rather than marketing. Copilot, they wrote, evolved into “an agentic platform capable of running long, multi-step coding sessions,” which means “a quick chat question and a multi-hour autonomous coding session can cost the user the same amount.”¹

That sentence is the whole argument. A flat seat works when every user costs roughly the same to serve. It breaks the instant one feature can cost a hundred times more than another and you cannot tell in advance which one a given user will reach for.

Agent mode is that feature. A single agentic request does not fire one model call. It fires a loop: read the repo, plan, edit, run tests, read the failure, edit again. Each turn is a fresh inference at full token cost. Ten to a hundred model calls per request is normal, and none of that compute amortizes across a flat seat the way a few autocompletes do.

A quick chat question and a multi-hour autonomous coding session can cost the user the same amount.

The infrastructure framing makes this concrete. One read going around is that AI assistants are becoming cloud infrastructure, and cloud infrastructure has a meter.⁸ Early estimates put Microsoft’s loss under the old flat model near $20 per user per month once agentic usage scaled.⁸ I cannot verify that figure against a primary source, so treat it as a directional estimate from industry coverage, not a confirmed number. But the direction is right. A subsidy that large does not survive contact with a feature that ten-x’s the cost to serve.

The Cost Math, Done Honestly

This is where the horror stories live, so let me separate what is verified from what is anecdote.

The verified part is the per-token rates Copilot publishes (per one million tokens, June 2026):⁹

Model	Input	Output
Claude Haiku 4.5	$1.00	$5.00
Gemini 2.5 Pro	$1.25	$10.00
GPT-5.4	$2.50	$15.00
Claude Sonnet 4.x	$3.00	$15.00
Claude Opus 4.x	$5.00	$25.00
GPT-5.5	$5.00	$30.00

Run the arithmetic against Pro’s 1,500 included credits, which is $15 of budget. At Sonnet’s $15-per-million output rate, a single heavy agentic session that generates a lot of output tokens can eat a meaningful slice of a Pro user’s monthly allowance in one morning. Output tokens dominate the bill, not metering itself, which is why model choice turns out to be the real lever.¹⁰ A generation routed to Opus or GPT-5.5 can cost multiples of the same work on Haiku or Gemini Flash.

The anecdotes are louder and need a label. Community burn-rate reports from GitHub’s own discussion thread include one developer who reported 822 credits consumed in a single request, another who reported 8 percent of a monthly allotment gone in two hours of regular development, and a third who reported 53 percent of a monthly quota consumed by four code-agent requests in one session.¹¹ These are user reports, not GitHub benchmarks. They are directionally real and they are not representative. Treat them as “this can happen,” not “this is your bill.”

The press-tier numbers run hotter still. The Register, on June 2, 2026, reported a single feature-implementation request costing over $6 in AI Credits on Pro+, and quoted a developer calling the change “a staggering shift from a predictable subscription to a stressful meter-based service.”¹⁰ TechCrunch surfaced cost-trajectory anecdotes of $29 a month climbing toward $750, and $50 toward $3,000.¹² Microsoft did not verify those, and even the critical coverage concedes they come from continuous vibe-coding agent loops at frontier-model rates with no workflow discipline.¹²¹⁰ They are the worst case, not the median.

The Behavioral Effect

Here is the part that does not show up in a pricing table and matters more than any single number. When you put a visible meter in front of a developer, the developer stops coding freely and starts coding defensively.

People are already rationing. The same discussion thread that produced the burn-rate reports also shows developers mid-month migrating to OpenRouter, Cursor, RooCode, and LM Studio.¹¹ GitHub’s billing-change thread accumulated more than four hundred comments and roughly nine hundred downvotes within forty-eight hours.¹⁰ That is not a pricing complaint. That is a workflow grievance. The meter changed the feel of the tool.

Rationing has a constructive form, harness engineering, and practitioners are already writing it up. The argument: respond to metering with deterministic cost control, because “everything that can be enforced deterministically should be, because instructions in your prompt cost tokens on every single request.”¹³ Stratify work by model. Use included or zero-cost models for the basics, cheap models like Haiku or Gemini Flash for research and routine edits, mid-tier Sonnet for implementation, and reserve Opus for reasoning that actually needs it. One practitioner claims structured delegation cuts costs by roughly 85 percent versus single, monolithic premium-model sessions.¹³ That number is a blog claim, not a benchmark, but the shape of the advice is sound: route deliberately, cap the expensive features, do not let an agent loop on Opus unsupervised.

The destructive version is simpler. People leave, or they stop reaching for the agent at all, which defeats the point of buying an agent.

The Counter-Case, Not Buried

A fair reading has to engage the strongest version of “this is fine,” because parts of it are true.

Light users are unaffected or come out ahead. If your Copilot use is mostly inline autocomplete, completions and Next Edit Suggestions stay unlimited and free of credits on every paid plan.² Your bill is the same $10 and your editor feels identical. The change is a non-event. The flat-fee era cross-subsidized heavy agentic users at the expense of light ones, and metering corrects that. If you were the light user paying for someone else’s autonomous refactors, this is your raise.

The bundles are over-allocated against sticker price. Pro includes about $15 of credits for a $10 plan, Pro+ about $70 for $39, Max about $200 for $100.⁵² For anyone who stays inside the included pool, the effective per-token cost sits below direct API access. GitHub is still subsidizing usage up to the cap. The meter is real, but the included buffer is generous relative to what you pay.

Those horror-story numbers are unverified and atypical, and I said so above. The $750 and $3,000 projections come from undisciplined frontier-model loops, and neither GitHub nor Microsoft verified them.¹²

Switch-away math is weaker than the backlash implies. Leaving for direct OpenAI or Anthropic APIs means paying the same per-token rates while losing IDE integration, GitHub context, and the PR and Actions hooks, plus the included credit buffer you already paid for.¹⁰ Copilot prices Sonnet at the same rate as Anthropic’s own API with no pass-through discount, so the alternative is only clearly cheaper if you route through cheaper or open models via OpenRouter or self-hosting. For IDE-centric work, Copilot’s integration is genuine value, not lock-in theater.

And the strongest point: this is convergence, not GitHub uniquely reaching into your wallet.

This Is an Industry Pattern

Every major AI coding tool abandoned unlimited flat-rate for agentic use, because multi-turn agent inference is genuinely uneconomical at a flat fee. Cursor moved to credit-pool tiers. Windsurf moved to daily and weekly quota refreshes in March 2026, which sidesteps the monthly-sprint-depletion risk. Claude Code meters through five-hour rolling windows with weekly caps.¹⁴¹⁵ GitHub was arguably the last conservative holdout, and the irony is that competitors won relative standing without shipping anything: the overage-uncertainty narrative alone handed Cursor and Windsurf the “predictable” label and cost GitHub its position as the safe enterprise default.¹⁶

For enterprises, the price is not even the deciding factor. Hacker News commenters point out that enterprise retention runs on Microsoft EA contracts, data-governance agreements, and compliance reviews, and that switching providers is “a one-year endeavor, minimum.”¹⁷ Metered pricing stings individuals and startups far more than the large accounts locked in by procurement.

The SaaS-per-seat mental model is ending for any tool with an agent mode.

So the metering is rational, probably right, and broadly shared. It is also not free of cost to you, and the cost is subtle.

The Strategic Read

Metering transfers cost-forecasting risk from the vendor onto the developer. Under a flat seat, GitHub ate the variance. A user who ran a thousand agent sessions and a user who ran three paid the same, and GitHub absorbed the difference. Under the meter, you own the variance. Your bill is now a function of how you work this month, and you have to forecast it yourself.

That is the real shift, and it reframes the whole category. AI coding is leaving the SaaS-per-seat model and joining the compute budget, sitting next to AWS and CI/CD as a metered line item with forecasting, cost centers, per-team budget owners, and usage reporting. Any tool with an agent mode is on this trajectory. GitHub just made it visible first and loudest.

There is also a scheduled second wave. Promotional credit inflation for existing customers runs June 1 through September 1, 2026. Business gets 3,000 credits per user per month against a standard 1,900, and Enterprise gets 7,000 against 3,900.⁷ When those promo credits expire on September 1, allotments normalize down by a third to a half, and any team that modeled its costs against the inflated June numbers is set up for a second round of sticker shock. The cliff is already on the calendar.

What To Actually Do

One more, for the skeptics weighing portability as a hedge. Because Copilot prices Sonnet at the same rate as Anthropic’s direct API with no pass-through discount, pipeline-centric agentic work may get better cost control from tools that take your own API key, such as Cursor, Windsurf, Cline, or Continue.dev. The honest counter is that Copilot’s VS Code and GitHub PR and Actions integration is real value for IDE-centric work, and you lose it when you leave. Portability is a hedge, not a verdict.

The Bottom Line

The prices barely moved. The model did. GitHub turned on a meter and, in doing so, confirmed what the agentic-coding boom always implied: this is infrastructure now, and infrastructure costs money to run. The metering is rational, it matches what Cursor and Windsurf and Claude Code already did, and for most developers it changes nothing or saves a little.

But it moves the forecasting risk onto you, and the September 1 promo cliff means a second sticker-shock wave is already scheduled. The right response is not panic or exit. It is the boring discipline you already apply to every other line on the compute budget: measure what you use, cap what you cannot afford to leave uncapped, route the cheap work to cheap models, and check the math before the bill, not after.

Footnotes

1. GitHub. “GitHub Copilot is moving to usage-based billing.” GitHub Blog, June 1, 2026. github.blog ↩ ↩² ↩³ ↩⁴

2. GitHub Docs. “Usage-based billing for individuals.” June 2026. docs.github.com ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

3. GitHub Docs. “What changed with billing” (legacy request-based model). 2026. docs.github.com ↩

4. GitHub Changelog. “Updates to GitHub Copilot billing and plans.” June 1, 2026. github.blog ↩ ↩²

5. GitHub Docs. “Individual plans” (Copilot Max tier, paused signups). June 2026. docs.github.com ↩ ↩² ↩³ ↩⁴ ↩⁵

6. GitHub Docs. “Budgets for usage-based billing” (hard cap off by default). June 2026. docs.github.com ↩ ↩²

7. GitHub Docs. “Usage-based billing for organizations and enterprises” (pooled credits, promo inflation, overage choice). June 2026. docs.github.com ↩ ↩² ↩³

8. hitechies.com. “GitHub Copilot metered billing, June 2026” (infrastructure framing, loss estimate). 2026. hitechies.com ↩ ↩²

9. GitHub Docs. “Models and pricing” (per-token model rates). June 2026. docs.github.com ↩

10. The Register. “GitHub Copilot users threaten exit as metered billing kicks in.” June 2, 2026. theregister.com ↩ ↩² ↩³ ↩⁴ ↩⁵

11. GitHub Community Discussion #192948 (burn-rate anecdotes, community opinion). 2026. github.com ↩ ↩²

12. TechCrunch. “‘What a joke’: GitHub Copilot’s new token-based billing spurs consternation among devs.” May 30, 2026. techcrunch.com ↩ ↩² ↩³

13. htek.dev. “GitHub Copilot billing: harness engineering for cost control.” 2026. htek.dev ↩ ↩²

14. usagebox.com. “AI coding tool billing models 2026” (Cursor, Windsurf, Claude Code convergence). 2026. usagebox.com ↩

15. byteiota.com. “GitHub Copilot AI Credits billing: real costs and how to stay in control.” 2026. byteiota.com ↩

16. aitoolbriefing.com. “GitHub Copilot token billing 2026” (competitive repositioning). 2026. aitoolbriefing.com ↩

17. Hacker News discussion. “GitHub Copilot usage-based billing.” 2026. news.ycombinator.com ↩