GitHub Copilot's New Bill Will Shock Your Dev Team

Two weeks ago, GitHub Copilot moved every plan to usage-based billing against a new currency called GitHub AI Credits. One credit equals a penny. The meter runs against input tokens, output tokens, and cached tokens at per-model rates. The flat-rate Premium Request Units that rationed Copilot for the last 18 months are gone. The seat license your engineering org budgeted around in Q4 is now an opening deposit toward a metered bill.

The June 1 transition was the headline. The September 1 promotional cliff is the part most engineering leaders haven’t priced into Q4 yet.

GitHub is running a promotional credit allowance on Business and Enterprise plans through August 31. Enterprise seats are seeing 7,000 promotional credits per user instead of the standard 3,900. That promo period ends in 11 weeks. The bill arrives in 12. If your finance team is modeling Q4 Copilot spend against the summer usage curve, they’re forecasting at almost double the credit pool the plan actually ships with.

The 90-day window is the budget-control window. Per-user and team-level spending caps just went generally available in GitHub admin. They are not enabled by default. The admins who configure them between now and August 31 walk into Q4 with a defined ceiling. The admins who don’t walk into Q4 reading a meter that already moved.

Quick Verdict

Question	The Answer
When did the new billing model start?	June 1, 2026. Two weeks ago.
What is the new unit?	GitHub AI Credit. 1 credit = $0.01.
What does an agentic coding session cost?	$30 to $40 per session at published rates.
What does Pro $10/month get you?	1,500 credits ($15). Not enough for one frontier-model agentic session.
Enterprise standard allowance?	3,900 credits per user per month.
Enterprise promotional allowance?	7,000 credits per user per month, through August 31, 2026.
When does the promo end?	September 1, 2026. Hard cliff.
Most expensive frontier model on the meter?	GPT-5.5 at $5/$30 per million input/output tokens.
Cheapest model on the meter?	MAI-Code-1-Flash at $0.75/$4.50 per million tokens.
What just went GA?	Per-user and team-level budget controls in GitHub admin.
Default state?	Off. Manual configuration required.
What to do this week	Configure caps. Pick default models. Rebuild Q4 forecast against 3,900 credits, not 7,000.

What Actually Changed on June 1

The old Copilot pricing model was a flat-rate seat license rationed by Premium Request Units. Pay your seat fee, get a quota of PRUs against the heaviest agentic features, and stop worrying about per-token rates. The headline number on the invoice did not move with usage.

The new model is a meter against a credit balance. The seat price stays the same on paper. The seat now buys a fixed monthly bucket of AI Credits and nothing more. Usage above the bucket either bills at the published per-token rate (if the admin opted in to overage) or the agentic features stop responding (if the admin didn’t).

According to GitHub’s billing documentation, one credit equals one cent of consumption priced against the listed API rate for each model. Run more tokens, burn more credits. Pick a more expensive model, burn credits faster. The math is no longer hidden inside a quota system. It is the invoice.

I covered the immediate transition shock the day before the meter started in Copilot’s Flat Rate Dies June 1: What to Do Now. Two weeks in, the post-mortem data is sharper than the pre-launch preview bills. The shock landed. The cost curve is steeper than the published rate card suggests. And the promo allowance is masking the steady-state baseline.

The $30-to-$40 Agentic Session Problem

Most engineering leaders haven’t sat with this number long enough. A single agentic coding session, the kind of multi-step run where Copilot reads context, plans a change, edits across files, runs tests, and iterates, routinely consumes $30 to $40 of credit at published per-model rates.

That single number breaks the Pro plan in one sitting.

Pro is $10 a month and ships with 1,500 included AI Credits — $15 worth. Run one frontier-model agentic session and the included allowance is gone before the session finishes. The next session either bills against an attached payment method or the agentic features stop responding. Pro+ at the next tier up gets 7,000 credits, or $70 of metered consumption. At $30 to $40 per frontier session, that buys roughly two agentic sessions a month. Route the same work through MAI-Code-1-Flash at a sixth of the per-token rate and the same 7,000 credits stretch to maybe 12 to 15 sessions. That ceiling sounds usable until you map it against a productive engineer running 15 to 25 agentic sessions a day.

The Business plan ships with 1,900 credits per user per month. Enterprise ships with 3,900 credits per user per month at the standard rate. Through August 31, Enterprise is running a promotional 7,000 credits per user per month. The promotional number is the number your team is metering against right now. The standard number is the number your contract reverts to on September 1.

The Q4 forecast modeled against 7,000 credits is the wrong baseline. The actual steady-state baseline is 3,900. Everything above that prices out at the per-token rate, which lands inside the same band that ran Uber’s entire 2026 AI budget out in four months.

What is the GitHub Copilot AI Credits system?

GitHub Copilot AI Credits is the unified billing currency introduced on June 1, 2026, replacing the flat-rate Premium Request Unit quota that previously rationed Copilot’s heavier features. One credit equals one cent of consumption, billed against the public API rate of whichever model the request runs against, including input tokens, output tokens, and cached tokens. Each plan ships with an included monthly credit allowance — Pro at 1,500 credits, Pro+ at 7,000, Business at 1,900 per user, Enterprise at 3,900 per user at the standard tier — with Enterprise temporarily running a promotional 7,000 credits per user through August 31, 2026. Usage above the included allowance either bills at the published per-token rate or, if the admin has not enabled overage, the agentic features stop responding for the rest of the month. Per-user and team-level budget caps are generally available in GitHub admin and require manual configuration. The system shifts Copilot from a predictable seat-license cost to a metered consumption cost that scales directly with engineering velocity and model choice.

The Model Cost Gap Is the Hidden Bill Multiplier

The biggest line-item variance is the one most teams haven’t audited yet. The model your developers default to multiplies the bill by 6x or more without any change in feature usage.

GPT-5.5 lists at $5 per million input tokens and $30 per million output tokens. MAI-Code-1-Flash lists at $0.75 per million input tokens and $4.50 per million output tokens. Same coding task. Same engineer. Same agentic session length. The session that runs against GPT-5.5 burns roughly 6.7 times the credit of the same session against MAI-Code-1-Flash on the output side, where most of the token volume lands in code generation work.

Model	Input ($/1M tokens)	Output ($/1M tokens)	Relative cost per session
MAI-Code-1-Flash	$0.75	$4.50	1x baseline
Mid-tier coding	$1.50 to $3.00	$6.00 to $15.00	2x to 3x
GPT-5.5	$5.00	$30.00	6x to 7x

The wrong default model multiplies your bill. The right default model still gets the coding work done, just against a credit consumption curve that fits the included allowance.

This is the procurement piece most engineering orgs have skipped. The Copilot model picker is now a budget control, not just a quality control. Defaulting the org to GPT-5.5 for everyday completions and routine refactors is the same decision as defaulting the company travel policy to first-class flights for routine trips. Reserve the frontier model for the workloads that actually need it. Route everything else through MAI-Code-1-Flash or a mid-tier model. The credit pool stretches. The frontier capability stays available for the work that justifies it.

I made a similar call in Your AI Stack Has an Expiration Date. Model-agnostic routing is the architecture choice that survives any pricing shift. Inside Copilot, the routing layer is policy and defaults set at the org level. Set them.

Per-User Budget Controls Just Went GA. Turn Them On.

The piece of the rollout that did not get the press the meter shock did is the admin tooling that just shipped alongside it. Per-user and team-level budget controls are generally available inside GitHub admin. The admin can cap monthly spend at the user level, at the team level, or at the org level. When the cap hits, the agentic features for that user or team stop responding for the remainder of the cycle. No surprise overage. No CFO meeting in October.

The default state is off.

This is the part that should be on every engineering manager’s calendar this week. The budget controls do not enable themselves. The org that ships into Q4 without them ships into the post-promo September 1 cliff with no ceiling on per-engineer spend. The org that configures them between now and August 31 ships into Q4 with a defined burn rate per seat and a hard floor under the monthly invoice.

The Uber pattern at scale is the cautionary case. Uber publicly imposed a $1,500 monthly cap on individual engineer Claude Code spend after burning the entire 2026 AI budget in four months. The cap shipped after the budget breach. The same cap configured before the budget breach would have changed the entire arc of the story. GitHub just gave every Copilot Enterprise admin the equivalent control. The admin who enables it before September 1 is the admin who avoids the Uber arc on the Copilot line.

The mechanics are straightforward. Set a per-user monthly credit cap at a number that fits inside the included allowance with margin for the most productive 20% of engineers. Set a team-level cap for the same threshold across the team’s seat count. Configure overage policy explicitly — pay through, block, or notify-and-block. Audit the cap weekly in July. Audit it daily in August. By the time the promo expires on September 1, the org has eight weeks of cap data against actual usage and the Q4 forecast is built against the right baseline.

The September 1 Cliff Is Not a Forecast. It’s a Calendar Date.

The promotional credit allowance is a marketing decision with a hard calendar boundary. Every Enterprise plan that signed before June 1 or upgraded into the new model in June is currently running against 7,000 promotional credits per user per month. On September 1, that allowance reverts to the standard 3,900 credits per user per month.

The cliff is not a forecast risk. It is a contract term. The Q3 spend report your finance team pulls in early October will show August at the promo allowance and September at the standard allowance. The delta between the two is either covered by per-user budget caps and model defaulting, or it shows up as overage charges that nobody budgeted against.

The math is mechanical. An engineering org running 200 Enterprise seats against the 7,000 promo allowance has 1.4 million credits per month to allocate. The same org against the 3,900 standard allowance has 780,000 credits. The gap is 620,000 credits per month, or $6,200 in monthly overage exposure per seat block of 200 if usage patterns hold. At a 5,000-engineer Uber-scale org, the gap is roughly 15.5 million credits per month, or $155,000 in monthly overage exposure starting September 1 if nothing changes between now and then.

The honest forecasting move is to model Q4 against 3,900 credits per user per month and treat the 7,000 promo as a learning period. The dishonest forecasting move is to extrapolate August spend forward into Q4 and assume the same allowance carries over. The dishonest forecast is what generates the October CFO meeting.

I covered the broader pricing flexibility piece in The AI Price War Starts Now. Don’t Lock In Wrong.. The Copilot case is a microcosm of the same dynamic. The vendor signal on the floor price is public. The buyer who builds the contract architecture against the signal beats the buyer who plans against the promotional number.

What should an engineering admin do in the 90-day budget control window?

The 90-day window between June 1 and September 1 is the period the org has to convert the promotional Copilot allowance into a steady-state budget posture. Three moves cover the window. First, enable per-user and team-level budget caps in GitHub admin at thresholds that fit the standard 3,900-credit allowance, not the promotional 7,000-credit allowance, so the cap configuration survives the September 1 reversion. Second, set the org default model away from frontier tier toward MAI-Code-1-Flash or a mid-tier coding model for everyday work, and reserve GPT-5.5 access for specific workloads that justify the 6x cost multiplier. Third, instrument the Copilot admin spend dashboards weekly through July and daily through August, so the Q4 forecast is built against actual post-cap, post-default usage patterns rather than against the unconstrained June and early July baseline. The three moves are signable inside any GitHub Enterprise contract this quarter. None require a vendor concession. All three are reversible if the data points the other direction.

The Promo Number Is the Wrong Baseline

Most Q4 Copilot forecasts being built right now are built against the wrong number. The teams modeling spend against current usage are extrapolating the 7,000-credit promo allowance forward into Q4 budgeting. That forecast is going to be off by roughly 80% on the day the promo expires.

The right baseline is the 3,900-credit standard allowance. The forecast that captures the post-September 1 reality is built against the standard allowance, with the promo period treated as a credit float that subsidizes the learning cycle and disappears on the calendar date.

The clean forecasting model is straightforward. Take June, July, and August spend data. Compute usage per engineer in credits. Compare against the 3,900 standard ceiling. The percentage of seats currently running above 3,900 credits is the percentage of seats that will hit overage starting September 1 if nothing changes. That number is the action item. Per-user caps and model defaults bring it down before the promo cliff. Doing nothing means the same number lands as an overage line on the October invoice.

The CFO question in early Q4 is going to be specific. “Why is our Copilot bill 40 percent higher in September than in August at the same seat count?” The honest answer for the org that planned against the promo allowance is going to be that the budget model treated a 90-day marketing offer as steady-state pricing. The org that planned against the standard allowance has a different answer — that the September bill is the budget number that has been forecast since June.

The Anti-Hype Read

Three honest cautions before this turns into a board memo.

The model cost numbers in this post are the public rate card and could shift. GitHub publishes the per-model token rates inside the Copilot billing documentation, and those rates can be updated. The 6x to 7x cost gap between GPT-5.5 and MAI-Code-1-Flash is the current state. The structural point — that model choice is now a budget control — survives any specific rate revision. Audit the rate card weekly through the promo period and adjust defaults if the pricing structure shifts.

The promo allowance might get extended. GitHub could decide to extend the 7,000-credit Enterprise allowance past September 1, particularly if the developer reaction to the cliff turns into an enterprise renewal risk. Plan against the published calendar. Treat any extension as upside. Do not budget against a hoped-for extension that has not been announced.

The budget caps themselves carry productivity tradeoffs. A per-user cap set too low blocks productive engineers from finishing work in the last week of the month. The cap configuration is a real engineering management decision, not just a finance one. Set the cap at the 80th percentile of the team’s healthy usage curve, not at the average. Allow legitimate overage workflows that route through an approval process. The cap is a guardrail, not a budget ceiling that forces every engineer into the average.

None of these cautions changes the recommendation. The September 1 cliff is real. The standard allowance is the right baseline. The admin tooling that protects the budget is generally available and off by default. Turn it on.

Three Moves Before the September 1 Cliff

Sized for any CTO, VP of Engineering, or platform lead running GitHub Copilot Enterprise at meaningful scale. Doable inside the next two weeks.

1. Configure per-user and team-level budget caps in GitHub admin against the 3,900-credit standard allowance. Do not configure against the 7,000-credit promo number. The cap configuration that survives September 1 is the one written against the steady-state baseline. Set the per-user cap at roughly 3,500 credits with a 10% buffer for the most productive engineers. Set a team-level cap at the sum of per-user caps across the team’s seat count. Configure overage policy explicitly — pay through, block, or notify-and-block — and document the decision so the finance team and the engineering managers are aligned on the policy before September.

2. Set the org default model to MAI-Code-1-Flash or an equivalent mid-tier coding model. Reserve GPT-5.5 access for specific workloads through an explicit policy, not as the default. The 6x to 7x cost multiplier between the frontier and the mid-tier coding model is the single largest variance in the Copilot bill that does not require renegotiating the contract or changing the headcount. Default routing is policy. Set it. Document which workloads justify the frontier-tier escalation and put the escalation behind a manager approval if the org culture supports it.

3. Rebuild the Q4 Copilot forecast against the standard 3,900-credit allowance. Treat the June through August window as a learning period that delivers usage data, not as a steady-state baseline for forward forecasting. Compute per-engineer credit consumption weekly across the promo window. Identify the percentage of seats currently running above 3,900 credits. That percentage is the overage exposure on September 1 if nothing changes. The Q4 budget submitted to finance has to either match the post-cliff baseline or document the explicit policy that accepts the overage. The forecast extrapolated forward from the promo allowance is the forecast that breaks in October.

What Comes After the Cliff

A few predictions worth committing to.

The standard 3,900-credit allowance becomes the default Copilot budget unit inside enterprise procurement by Q4 2026. Finance teams that have been treating the Copilot line as a flat seat license through 2025 are now treating it as a metered consumption line with a credit pool per seat. The mental model shifts from seat count to credit pool by year-end. The Q1 2027 budget reviews will look fundamentally different from the Q1 2026 reviews because the underlying unit changed.

The model defaulting decision becomes a standard governance decision inside engineering platform teams. Today most orgs let individual developers pick their model inside Copilot. By 2027, the model pick is a platform policy with org-wide defaults, escalation paths for frontier-tier workloads, and audit trails on the workloads that triggered the escalations. The pattern matches the way enterprise procurement governed cloud instance type selection ten years ago. Same shape. Same governance maturity curve.

The per-user budget cap becomes a standard contract control across every metered AI developer tool by 2027. Copilot just shipped it. Claude Code, Cursor, and the rest of the developer-tool stack will ship equivalents inside two quarters because the enterprise procurement function will require the control as table stakes for any usage-based renewal. The orgs that build the cap configuration discipline now are the orgs that scale across vendors cleanly.

My Read

The June 1 transition to AI Credits was the headline story. The September 1 promo cliff is the budget story. Most engineering orgs that managed the immediate transition by absorbing the meter shock haven’t built the post-promo budget posture, and the 90-day calendar window between the two dates is the only window the org has to do it without overage.

The 7,000-credit Enterprise promo allowance is the wrong baseline. The 3,900-credit standard allowance is the right baseline. The forecast built against the promo number is going to be off by roughly 80% on September 1 and finance is going to find the gap inside the October invoice. The forecast built against the standard number captures the cliff before it lands.

Per-user and team-level budget caps just went generally available in GitHub admin. They are not enabled by default. They are the single most important administrative control in any Copilot Enterprise deployment this quarter and the one most likely to be deferred until the post-cliff overage forces the conversation. Configure them now. Set them against the standard allowance. Document the overage policy explicitly.

The model default is the second control. GPT-5.5 at $5/$30 per million input/output tokens runs 6x to 7x the credit cost of MAI-Code-1-Flash at $0.75/$4.50 for the same coding task. Defaulting the org to the frontier tier without an explicit policy is the same decision as defaulting the travel policy to first class without a justification process. Set the default. Reserve the frontier tier for the workloads that need it.

The Q4 Copilot forecast is the third control. Rebuild it against the standard allowance. Use the promo window as a learning period. Identify the overage exposure before September 1 instead of after. The CFO question in early October is going to be specific. The honest answer for the org that planned against the promo number is going to be that the budget model treated a 90-day marketing offer as steady-state pricing. The org that planned against the standard number has a different answer.

The 90-day window started two weeks ago. Eleven weeks remain. Configure the caps. Set the defaults. Rebuild the forecast.

The cliff is on the calendar.

---

Related Reading:

• Copilot’s Flat Rate Dies June 1: What to Do Now
• Your AI Coding Budget Is About to Break
• The AI Price War Starts Now. Don’t Lock In Wrong.
• Your AI Stack Has an Expiration Date
• The AI Trust Deficit: Copilot’s Public Perception Problem