Note: The numbers in this post are illustrative and based on a representative enterprise scenario. They are designed to reflect realistic usage patterns and cost dynamics — not data from any specific organization.
Introduction
GitHub is making a significant shift in how it bills for Copilot usage. Starting June 1, 2026, Copilot usage will be measured in AI Credits (AICs) instead of Premium Requests (PRUs). This isn’t just a rename — it’s a fundamental change in how costs are calculated.
Under the current model, every request to Copilot costs the same flat rate regardless of which AI model you used or how much compute it consumed. Under the new model, you pay for the actual AI compute consumed. A quick autocomplete and a deep multi-file reasoning session are no longer priced the same.
So the big question is: what does this actually mean for your team’s bill?
This blog walks through a scenario-based analysis using April 2026 usage data, explaining step by step how the numbers add up, where costs come from, and what you can do about it before June 1.
First, Let’s Understand the Two Billing Models
Before getting into the numbers, it helps to be clear on what’s actually changing.
| Metric | Current Billing (PRUs) | Usage-Based Billing (AICs) |
|---|---|---|
| Unit | Premium Request (PRU) | AI Credit (AIC) |
| Rate | $0.04 per PRU | $0.01 per AIC |
| Measurement | Flat per request | Actual AI compute tokens consumed |
Think of it this way. Under PRU billing, every Copilot interaction — whether you asked it to autocomplete a line or analyze an entire codebase — costs the same $0.04. It’s like a flat-fare bus ticket: simple and predictable, but it doesn’t reflect how far you actually traveled.
Under AIC billing, the cost reflects what the AI actually did. A simple code suggestion might consume 50 AICs (costing $0.50), while a complex reasoning session with Claude Opus 4.6 might consume 5,000 AICs ($50). The meter runs based on real compute — not just a headcount of requests.
This is why model choice becomes so critical under UBB. A user making 100 requests with Claude Haiku will cost a fraction of what another user making 100 requests with Claude Opus costs — even though their PRU counts look identical.
The Big Picture: How Much More Does UBB Actually Cost?
Let’s look at a representative enterprise in April 2026 and see what the transition does to the bill.
Under current PRU billing, the total came to roughly $620:
- The team consumed about 10,500 PRUs worth ~$420 in gross cost
- Almost all of that was covered by included PRUs from their licenses (−$418 discount)
- Net overage was just ~$2
- Add their license cost of ~$618, and the total is ~$620
That looks comfortable. The included pool nearly covered everything. But now look at the same month under AIC billing:
Under usage-based AIC billing, the total jumps to roughly $3,938:
- The team consumed ~400,000 AICs worth ~$4,000 in gross cost
- The included AIC pool covered only ~$680 of that
- Net overage: ~$3,320
- Same license cost of ~$618
- Total: ~$3,938
That’s a 6.4x increase — an extra ~$3,318 from the same month of work.
Why such a dramatic jump? Under PRU billing, the included pool was nearly dollar-for-dollar because most interactions were lightweight. But under AIC billing, the heavy use of premium models generated far more raw compute than the included pool could absorb. The team consumed ~5.9x more AICs than their licenses covered.
Why Does the Included AIC Pool Fall Short?
This is the part that surprises most teams. They assume their license covers normal usage — just like PRUs mostly did. But AICs work differently.
GitHub provides a billing preview simulator that lets you upload your current PRU usage report and see a projected AIC bill. When teams run this for the first time, the result is often a wake-up call: the included AIC pool is calculated based on seats, not usage intensity.
Here’s how it works. Each Copilot Business seat includes 1,900 AICs/month. Each Copilot Enterprise seat includes 3,900 AICs/month. The pool is the sum of what all your seats bring in.
Tip: Use GitHub’s billing preview tool to upload your own usage report and simulate your AIC bill before June 1. Don’t wait for the first invoice to find out your cost profile.
Who’s Actually Using What? Breaking Down the License Pool
The Users section of GitHub’s billing view is where you see how the included AIC pool is built — and how quickly it gets exhausted.
The organization has 20 active users distributed across two license tiers:
| License Type | Users | Included AICs |
|---|---|---|
| Copilot Business | 5 | 9,500 |
| Copilot Enterprise | 15 | 58,500 |
| Total AIC Pool | 20 | 68,000 |
Each Copilot Business seat includes 1,900 AICs/month, and each Copilot Enterprise seat includes 3,900 AICs/month. With ~400,000 AICs consumed against only 68,000 included, the team is consuming ~5.9x their included allocation. Every AIC beyond that pool costs $0.01 — and those overages accumulate fast.
How Do Different Models Actually Compare?
It’s useful to step back and look at all models side by side. In our scenario, the team was actively using:
- Claude Opus 4.6 (premium reasoning)
- Claude Sonnet 4.6 (balanced capability and cost)
- Claude Haiku 4.5 (lightweight and fast)
- GPT-5.3-Codex (code-optimized)
- Gemini 3 Flash (fast, multimodal)
- Auto-routed variants (Copilot choosing the model automatically based on context)
Here’s the pattern that emerges when you compare requests versus AICs by model:
- Claude Opus 4.6 consistently dominated AIC consumption — even on days where its request count was moderate, it generated the largest share of compute. A handful of Opus sessions could outspend a full day of Haiku usage.
- Claude Sonnet 4.6 was a meaningful contributor too, but at a noticeably lower cost per request. It struck a good balance — capable enough for most complex tasks, without the heavyweight token consumption of Opus.
- Claude Haiku 4.5 and GPT-5.3-Codex showed up as genuinely cost-efficient options. Plenty of daily requests, minimal AIC impact. These models are well-suited for repetitive tasks: code completion, documentation, simple Q&A, small refactors.
The takeaway is practical: not every task needs the most capable model. A developer who defaults to Opus for everything — including tasks where Haiku would do fine — generates significantly more AIC cost than one who matches the model to the complexity of the task. Check the pricing table for specific AIC costs by model to understand the cost implications.
Cost Centers: Giving Teams Visibility Into Their Own Spend
One of the most useful features in the UBB billing framework is Cost Centers — a way to group users by team, project, or department and track their AI spending separately.
Think of it like departmental budgeting for AI compute. Instead of one big enterprise bill at the end of the month, you can see exactly how much each team spent, and where they stand against their budget.
More importantly, cost centers create awareness before the bill arrives. When a team knows they’re 80% through their AIC budget by the 20th of the month, they can make intentional decisions: switch to lighter models for the rest of the month, defer complex sessions, or plan for the overage proactively. Without cost centers, none of that visibility exists until the invoice lands.
Setting up cost centers before June 1 lets teams build a baseline now, so they go into the UBB era with real data — not estimates.
So What Should You Actually Do?
Let’s be direct about the action items — and the reasoning behind each one.
1. Be intentional about model selection
Not every task needs Claude Opus 4.6. Code completion, documentation, simple refactors, Q&A on known topics — Claude Haiku or Sonnet handles these well at a fraction of the AIC cost. Reserve Opus for genuinely complex reasoning where its capabilities make a real difference: multi-file architecture analysis, complex debugging sessions, nuanced code generation. This one habit change, applied consistently across a team, can significantly cut monthly AIC spend.
2. Look at per-user AIC consumption, not just PRU counts
Two users with similar PRU counts can have wildly different AIC footprints. If you’re only watching PRUs today, you’re missing the signal. Export your usage data now, identify your top AIC consumers, and have a conversation about whether their model choices match the complexity of their tasks. It’s not about restricting access — it’s about informed usage.
3. Calculate your included AIC pool vs actual consumption
Take your seat count, multiply Business seats by 1,900 and Enterprise seats by 3,900. That’s your monthly included pool. Then look at your current usage data and estimate how many AICs your team actually consumes. The gap between those two numbers is your projected monthly overage under UBB — and it’s the number your budget conversation should start with.
4. Set up Cost Centers before June 1
If you wait until after the billing switch to set up cost centers, you’ll be flying blind for the first month. Set them up now, let usage data accumulate against them for a few weeks, and go into June with a real baseline for each team rather than a surprise.
5. Think about session complexity, not just session count
The number of Copilot interactions is no longer the key cost metric. What matters is how much context is being passed to the model and how complex the responses are. Long agentic sessions, whole-repository context, and extended multi-turn chains are where AICs accumulate quickly. Encourage developers to scope their sessions — not every question needs the full project as context, and breaking complex tasks into focused steps often produces better results anyway.
6. Run the UBB simulator with your own data
GitHub’s Cost Management tool lets you upload your current PRU usage report and see a simulated AIC bill. Run it with a recent month of real data. The result will be far more concrete than any estimate and will give you the numbers you need to have an informed budget conversation with leadership before June 1.
Closing Thoughts
The shift to Usage-Based Billing isn’t just a pricing change — it’s a transparency change. For the first time, organizations will be able to see exactly what their AI compute actually costs, broken down by model, by user, by product, and by team.
That visibility comes with a price for teams that have been using premium models freely under the comfort of flat PRU billing. But it also creates a genuine opportunity: to understand what’s driving spend, right-size model selection, and make informed decisions about AI usage rather than treating it as a cost-free resource.
The organizations that will navigate this transition best are the ones that start analyzing their usage patterns now — before June 1 — rather than reacting to the first UBB invoice.
