GitHub Copilot Credits, Naveen Raj Kanagaraj

The situation

The transition landed. The invoice has not.

On 1 June, premium requests were replaced by metered AI Credits across every plan. Base prices held: Business $19, Enterprise $39 per user per month. Completions stay free; chat, CLI, agents, and code review are metered.

For June, July and August a promotional allowance replaced the standard one rather than adding to it: 3,000 credits on Business, 7,000 on Enterprise. It ends 1 September. The backlash so far is almost entirely individuals. Enterprise is still silent, which is the tell: the real exposure hasn't hit a finance team yet.

Included AI allowance, per user / month

The promotional amount replaced the standard one for three months · 1 credit = $0.01

Business · Jun–Aug, promotional3,000 cr · $30

Business · from 1 Sep, standard1,900 cr · $19

Enterprise · Jun–Aug, promotional7,000 cr · $70

Enterprise · from 1 Sep, standard3,900 cr · $39

On 1 September, included usage falls by 37% on Business and 44% on Enterprise. Seat prices do not change.

The diagnosis

We didn't raise prices. We removed the ability to forecast them.

A cost increase

is absorbed
negotiated at renewal
stays inside engineering

Quiet.

A forecasting failure

is escalated
reaches finance, triggers procurement review
ends in a spend freeze

Loud.

We produced the second and are being criticised for the first. Every fix that adjusts price solves the wrong problem. Credits burn on tokens at each model's published rate: output runs $5 per million tokens on Claude Haiku 4.5 and $25 on Claude Opus, and for code review the model is chosen automatically and not disclosed. The person choosing is a developer mid-task with no budget context.

The mechanism

The safety net came out in the same release as the meter.

Before

Allowance runs out
Model falls back
Work continues, degraded

Graceful. No decision required.

After

Allowance runs out
No fallback exists
Overage by default, or work stops

Binary. Overage unless an admin opts out.

The field is selling budget caps as the answer to unpredictability. The cap is also an off switch, and overage stays on unless an admin disables it. We tell admins predictable spend is one setting away, without telling them that setting can stop their developers mid-task. A user-level budget can block one developer while the pool still has credits.

The sensitivity

What 1,000 Enterprise seats pay in September.

Monthly cost at 1,000 Enterprise seats, by September consumption

Pooled allowance 3,900,000 credits, $39K included · 1 credit = $0.01 · directional, not a forecast

3,500 credits per seat$39K · no overage

5,250 credits per seat$52.5K · $13.5K over

7,000 credits per seat, full promotional use$70K · $31K over

Included in planOverage, billed by default

Below 3,900 credits per seat, September adds no overage at all. Above it, every credit is billed. At full promotional use a 1,000-seat enterprise carries ~$31K of monthly overage, billed by default, on an October invoice nobody forecast. The point isn't the estimate. It's that the customer cannot see in advance which side of that line they land on.

Positioning

From predictable seat to governed platform.

Was

"One price per developer. Stop thinking about it."

→

"Spend you can see, attribute, and govern before it happens."

Positioning statement · platform engineering lead

For the platform lead who now owns a variable AI budget they're accountable for, Copilot is the only AI coding platform that meters agentic work at the cost centre rather than the seat, so spend is attributable, capped, and defensible in a budget review.

What it deliberately does not claim: that the bill is predictable. It isn't, and claiming otherwise breaks trust a second time.

Messaging

One reposition. Three translations.

Developer

Sees the meter. Completions stay free; credit cost shows at the point you pick a model.

"See the cost before you commit."

Platform lead

Owns allocation. Pooled usage removes stranded capacity; budgets at enterprise, cost-centre, or user level.

"Allocate by cost centre, not by seat."

Finance

Sees the variance. On exhaustion the organisation decides: continue at published rates, or cap. Overage is the default.

"You choose the ceiling in advance."

Competitive

Everyone has this problem. We're first to show the customer.

Agentic coding consumes frontier inference. That cost is real for every vendor. The only question is who absorbs the variance, and for how long.

Flat-rate seat · most of the category

Predictable invoice, easy to buy
Vendor absorbs variance, then rations quietly
Spend can't be attributed, because there's no spend to attribute

Metered & governed · Copilot, from 1 June

Spend is attributable and capped in advance
The invoice moves
The trust cost is paid up front, in public, which is where we're standing now

Speculation. Flat-rate agentic pricing doesn't survive current inference economics. If competitors meter within 12 to 18 months, their customers arrive with no allocation model at all.

Execution

Land the cliff before it lands on them.

19 Aug

Admin notification

Every admin sees their own June–August consumption beside their standard September allowance.

19 Aug

Field briefed first

AEs and CSMs get the cap trade-off before customers get the email.

25 Aug

Guided configuration

CSM-led budget setup, ranked by projected gap rather than by ARR.

1 Sep

Say it plainly, in public

Changelog and docs state that promotional credits have ended. Silence converts a billing change into a trust story.

Early Oct

First unsubsidised invoice

Outreach to every account above forecast before they open it.

The dependency

One ask. Marketing can't close this alone.

The ask · restore a degraded capability floor at the cap

When the pool is exhausted and the organisation has chosen to cap, route to a low-cost model at reduced concurrency instead of stopping. The admin keeps a hard spend ceiling. The developer keeps a working tool.

Why this and not a discount. A taper postpones the conversation by one cycle and costs revenue in every cycle after. A floor is paid for once.

If it's declined. The claim weakens from governed to attributable. The fallback is alerting well ahead of the ceiling and a named approval path, which manages the stop rather than preventing it.

The trade-off

What this costs us.

We give up

The predictability message.

Competitors selling flat rate run it against us for at least two quarters, and it works
We should expect to lose price-sensitive mid-market deals
We should not pretend otherwise in the forecast

We keep

A claim that stays true.

Attribution and governance can be demonstrated in a console today
It will not be retracted later
The alternative fails on the first spiky month and costs the account, not just the deal

The bet. That in enterprise software trust compounds and price advantages do not. We absorb the reaction in public now, and the allocation layer becomes table stakes as the rest of the category meters. If that reading is wrong, this is the wrong strategy, and the measurement below says so by November.

Measurement

Four numbers, and what would prove me wrong.

Budget configuration rate

Organisations with a budget set at any level, by 30 September.

Cap proximity in week two

Share of orgs whose usage approaches their own ceiling within 14 days. The only signal that arrives early enough to recalibrate.

Forecast gap at first invoice

Admin projection against actual October spend. The direct measure, and the number I'd be held to.

Billing cited in Q4 churn

Non-renewals naming predictability. Lagging and contaminated. Read alongside 03, never alone.

Falsification. If the October gap is wide and configuration is high, the diagnosis was wrong and this memo should be withdrawn, not iterated.