Copilot's usage-based billing went live June 1, 2026. Visual Studio Magazine ran a headline on Day 1: "Slammed by Copilot Usage-Based Billing on Day 1, Facing $180 Bill for June." Today they followed up: "Copilot Billing Shock Hits Developers."
The problem isn't the new billing model itself — it's that most developers have no idea what each action actually costs. GitHub published the per-token pricing in their docs, but didn't map it to real-world usage patterns.
Here's the complete breakdown.
The Credit Math
1 AI Credit = $0.01 USD.
Your plan includes a fixed credit allowance:
- Copilot Pro: 1,500 credits ($15 value) for $19/mo
- Copilot Pro+: 7,000 credits ($70 value) for $39/mo
- Copilot Max: 20,000 credits ($200 value) for $99/mo
Code completions are free — they don't consume credits. Everything else does.
Model Pricing (What GitHub Buried in Their Docs)
| Model | Output Cost/1M tokens | Credits/1M out |
|---|---|---|
| Gemini Flash | $0.30 | 30 cr |
| GPT-5 mini | $2.00 | 200 cr |
| GPT-4o | $10.00 | 1,000 cr |
| Claude Sonnet 4.6 | $15.00 | 1,500 cr |
| Claude Opus | $75.00 | 7,500 cr |
Real-World Cost Table (The Part That Shocks People)
| What You're Doing | Model | Est. Output Tokens | Credits | Cost |
|---|---|---|---|---|
| Simple rename | Sonnet | ~5K | 7 | $0.07 |
| Add unit tests | Sonnet | ~15K | 22 | $0.22 |
| Fix a bug | Sonnet | ~30K | 45 | $0.45 |
| Complex refactor | Sonnet | ~80K | 120 | $1.20 |
| Architecture review | Opus | ~50K | 375 | $3.75 |
| 10-file agent run | Opus | ~200K | 1,500 | $15.00 |
| Heavy agent day | Opus | ~500K | 3,750 | $37.50 |
That last row is your entire Pro+ monthly credit budget. In one day.
Why This Is Happening
Copilot's default in agent mode often routes to frontier models (Opus, GPT-5). It uses the same model for a quick rename as for multi-file architectural changes. Nobody configured it otherwise because under flat-rate billing, it didn't matter. Now it does.
The Fix (5 Minutes)
Add this to your VS Code settings.json:
"github.copilot.chat.defaultModel": "gpt-4o",
"github.copilot.advanced": { "length": 500, "temperature": 0 },
"github.copilot.enable": {
"markdown": false,
"yaml": false,
"json": false
},
"github.copilot.chat.agent.runCommand": false
Also: go to github.com/settings/billing and set a hard spending limit to $0 overage. This stops Copilot from charging your card after credits run out.
The Model Selection Rule That Saves 70%
Route by task complexity:
- Tests, docs, renames → Gemini Flash or GPT-5 mini ($0.30–$2/M output)
- Feature dev, bug fixes → Sonnet or GPT-4o ($10–15/M output)
- Architecture, complex multi-file → Opus or GPT-5 (only when genuinely needed)
65% of daily coding tasks don't need a frontier model. Matching model to task is where the 70% savings comes from.
Going Further
I put together a complete kit: full model cost reference card, VS Code config templates (copy-paste ready), a cost calculator spreadsheet, and a migration guide for Cursor/Continue.dev if Copilot is still too expensive after optimizing.
→ GitHub Copilot Billing Survival Kit — $37
Or use the free tips above — the settings change alone cuts most bills by 40-70%.
Happy to answer questions in the comments.
Top comments (0)