GitHub Copilot cut team AI credits up to 44%. The price didn’t move.

On 1 September 2026 a GitHub Copilot Business seat stopped including 3,000 AI credits a month and started including 1,900. Enterprise dropped from 7,000 to 3,900. No seat price moved. Business is still $19 per seat per month, Enterprise still $39. GitHub had granted the higher numbers as a three-month promotion covering June, July and August, and the promotion ended on the day it said it would.

That is a 36.7% cut for Business and a 44.3% cut for Enterprise. It lands on agentic work only: chat, the Copilot CLI, the cloud agent, Spaces, Spark, third-party coding agents. Code completions and next edit suggestions stay unlimited and consume no credits.

One number explains the rest. An AI credit is worth $0.01. So 1,900 credits is $19 and 3,900 credits is $39, which is exactly what those seats cost. After the promotion, a Copilot team seat gives you back its own price in model usage, and that is the whole allowance. You are paying $19 for seat management, policy controls, unlimited completions, and $19 of tokens at cost.

What ended on 1 September

GitHub announced the change on 27 April 2026. Premium request units, the old flat unit where one interaction cost one PRU with a per-model multiplier, were replaced by AI credits on 1 June. Credits are drawn down by token count, input and output and cached, at each model's published API rate.

The same post set the promotion: existing Business customers got $30 in monthly credits, existing Enterprise customers $70, for June, July and August. September is the first month billed at the standard allowance.

GitHub was direct about why. "Agentic usage is becoming the default, and it brings significantly higher compute and inference demands," the post says, and GitHub "has absorbed much of the escalating inference cost behind that usage." The promotion was the last of the absorbing.

The included allowance is the sticker price, on every plan

Line the plans up next to their base credit allotments and the pattern holds without exception. Copilot Pro costs $10 and carries 1,000 base credits. Pro+ costs $39 and carries 3,900. Max costs $100 and carries 10,000. Business costs $19 and carries 1,900. Enterprise costs $39 and carries 3,900. In every case the base allowance converts to the subscription price at a penny a credit.

Individual plans then add a second bucket GitHub calls a flex allotment: 500 credits on Pro, 3,100 on Pro+, 10,000 on Max. GitHub describes it as an amount "designed to adapt as AI economics evolve," which is a way of saying it can move. Business and Enterprise have no flex bucket. The promotion was doing that job, and it expired.

Divide included credits by price and the ranking is uncomfortable for anyone who bought the team plan for the usage.

Five Pro+ seats cost $195 a month and carry 35,000 credits, $350 of model usage. Five Business seats cost $95 and carry 9,500 credits, $95 of usage. The team plan buys admin: policy, seat assignment, audit, and one thing the individual plans do not have, which is pooling. Business and Enterprise credits pool at the billing entity, so a hundred Business seats share 190,000 credits and a heavy user can spend a light user's share. Five individual subscriptions are five separate buckets that expire unused. That pooling is worth something to a team with uneven usage. It is not worth 79 cents on the dollar to a five-person shop where everyone codes.

What 1,900 credits buys in tokens

Copilot charges each model's published API rate. Claude Sonnet 5 inside Copilot runs $2 per million input tokens and $10 per million output. Anthropic's own list price for Sonnet 5 is $2 and $10. Same numbers. The rate is pass-through; what ended was the discount.

So $19 of credits is 9.5 million Sonnet 5 input tokens, or 1.9 million output tokens, or some mix. A single agent task that reads a few files, plans, edits and re-reads can move hundreds of thousands of input tokens on its own. Run the cloud agent a few times a day on a real repository and 1,900 credits is a fortnight, not a month. That is the arithmetic behind every "my September bill looks wrong" thread, and it is the same shift we looked at when AI tools went metered in June.

Nothing stops at the line unless you set it

This is where a smaller allowance turns into a bigger invoice. GitHub's budget documentation says the setting called "Stop usage when budget limit is reached" applies to enterprise spending limits, cost center budgets and organization budgets, and is off by default. There is no budget in place until an administrator creates one. Without one, a request that exhausts the pool keeps going and starts metering.

User-level budgets behave differently: they always enforce a hard stop with no option to continue. If you want a ceiling that actually holds, that is the control to use, and somebody has to go and set it. Credits do not carry over either. Unused ones are forfeited at 00:00:00 UTC on the first of the month.

Who this does not touch

If your team uses Copilot mostly as autocomplete, the 1 September change costs you nothing. Completions and next edit suggestions are excluded from credits and stay unlimited on every paid plan. The cut is a tax on agents, and it scales with how much of your work you have handed to them.

One thing worth flagging as unconfirmed. Several write-ups have said the individual flex allotments also expire in September, taking Pro to 1,000 credits and Pro+ to 3,900. GitHub's individual billing page still lists Pro at 1,500 and Pro+ at 7,000 as of 8 September 2026 and publishes no end date for flex. Treat the individual cut as a rumour until GitHub dates it. Treat the flex bucket itself as temporary, because GitHub has told you in writing that it adapts.

What to watch

The first invoices under the standard allowance arrive at the end of September. Two things will show up in them. Teams that never set a budget will find out what their agent usage actually costs, at pass-through rates, with no cushion. And teams that do the division will notice that five individual seats now carry more model usage than five Business seats, which is an awkward place for a product sold on being the team option.

The wider pattern is the one worth tracking. Vendors are converging on subscriptions that include their own price in tokens and charge list rates above that, which turns the seat fee into a wallet and moves all the variance onto the buyer. Token prices falling does not help much when the lock-in is the real cost, and neither does topping up a balance when the top-up itself carries a fee.

If you bill clients for AI-assisted work, the line item just got easier to calculate and harder to absorb. A credit is a penny. Count them.

Sources

• GitHub Docs, Usage-based billing for organizations and enterprises, retrieved 8 September 2026
• GitHub Docs, Usage-based billing for individuals, retrieved 8 September 2026
• GitHub Docs, Plans for GitHub Copilot, retrieved 8 September 2026
• GitHub Docs, Models and pricing for GitHub Copilot, retrieved 8 September 2026
• GitHub Docs, Budgets for usage-based billing, retrieved 8 September 2026
• The GitHub Blog, GitHub Copilot is moving to usage-based billing, 27 April 2026
• Anthropic, Claude API pricing, retrieved 8 September 2026

This is not financial advice.