One developer opened their bill preview in April. The same usage patterns that cost $39.07 under PRUs came back at $902.72. Not a typo. That's GitHub's own preview tool showing what the new billing model would have charged. June 1, this becomes reality for every Copilot subscriber.
What's actually changing?
Until now, Copilot ran on "Premium Request Units" (PRUs). Different models consumed different PRUs, but you could use them freely within your monthly cap. On the Pro plan, $10/month was the ceiling.
From June 1, that becomes GitHub AI Credits. 1 AI Credit = $0.01. Your subscription fee equals your monthly credit limit.
| Plan | Monthly Price | Included Credits |
|---|---|---|
| Copilot Pro | $10/mo | 1,000 credits |
| Copilot Pro+ | $39/mo | 3,900 credits |
| Copilot Business | $19/user/mo | 1,900 credits |
| Copilot Enterprise | $39/user/mo | 3,900 credits |
What consumes credits?
| Still Free | Credit-Consuming | |
|---|---|---|
| Inline code completions | ✓ | |
| Next Edit Suggestions | ✓ | |
| Copilot Chat | ✓ | |
| Copilot CLI | ✓ | |
| Cloud agents / Spaces | ✓ | |
| Code review (+ Actions minutes) | ✓ |
How much more will this actually cost?
Same plan price, so why does $39 turn into $902? Here's the thing.
GitHub's official line: "A quick chat question and a multi-hour autonomous coding session costing the same amount was no longer sustainable." GitHub had been absorbing a large portion of rapidly growing inference costs under PRUs — it just couldn't keep doing that.
The heavier your agent usage, the bigger the shock. One European developer saw their projected bill jump from €67 to €966. These numbers are showing up in GitHub's own preview tool before June even starts.
Model choice matters a lot here. The same task can cost 15× more depending on which model you use.
| Model | Output cost (per 1M tokens) | Best for |
|---|---|---|
| GPT-5 mini | $2.00 | Simple questions, quick edits |
| Gemini 3 Flash | $3.00 | Balanced everyday tasks |
| GPT-4.1 | $8.00 | General coding |
| Gemini 2.5 Pro | $10.00 | Complex reasoning |
| Claude Sonnet 4.6 | $15.00 | High-quality code |
| Claude Opus 4.7 | $25.00 | When only the best will do |
| GPT-5.5 | $30.00 | Top-tier reasoning |
Real cost example (Xebia analysis)
Running a 1M-token session (80% input / 20% output) with Claude Sonnet 4.6 costs roughly $3.20 per session. Claude Opus 4.7 runs about $8.10. If an agent is scanning whole repositories, that adds up fast across a workday.
Heads up for annual subscribers
Annual Pro/Pro+ subscribers don't auto-migrate on June 1. Your PRU-based plan continues until renewal. But model multipliers increase on June 1, so PRUs will drain faster even before you switch.
What to do right now
- Download your April usage report
Go to GitHub Settings → Billing → Usage, pull April's data, and compare it with the preview billing tool. That's your real cost baseline for June. - Set a spending cap before June 1
Budget controls exist at the enterprise, org, cost center, and user level. Set an alert at 75% consumption so you don't hit limits mid-workflow. - Match models to task complexity
Don't throw GPT-5.5 or Claude Opus at every question. GPT-5 mini ($2/1M output) or Gemini Flash ($3/1M output) handles most routine tasks fine. Reserve premium models for complex design and reasoning. That alone can cut costs 60–80%. - Minimize context window size
Don't dump entire repositories into agent context. Explicitly reference only the files you need. Token consumption drops significantly. - Compare alternatives
If you run agents heavily, direct Claude Max ($100–200/mo) or raw API access might be more economical. Run a simulation with your April data.
