GitHub Copilot goes metered: what developers will actually pay per 1M tokens
On June 1, Copilot drops premium request units and switches to AI Credits - 1 credit, $0.01, direct token billing. Code completions stay free. Everything else gets a price tag. Here's what that actually means for your monthly bill.
Photo by Pankaj Patel on Unsplash
Starting June 1, Copilot Pro's $10/month buys you 1,000 AI Credits ($0.01 each) instead of a fixed request count. Chat and agent calls consume credits at the real model rate - Claude Sonnet 4.6 at $3.00/1M input, GPT-5.5 at $5.00/1M. Code completions stay unlimited.
For most developers this changes nothing. For anyone running agentic sessions daily, a Copilot Pro allowance lasts under two weeks. Amazon Q remains the only tool in this space that still doesn't bill per token.
What GitHub actually changed
The April 27 announcement from GitHub CPO Mario Rodriguez was unusually direct about the reason: "Today, a quick chat question and a multi-hour autonomous coding session can cost the user the same amount. GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable."
That's the clearest admission yet that flat-rate AI tool pricing is breaking. Copilot was designed when 'using Copilot' meant tab completions. Now it means running multi-hour parallelized agent sessions against your entire codebase - and GitHub was eating the bill for all of it equally.
The mechanics: premium request units (PRUs) are out. AI Credits are in. One credit costs $0.01. Token consumption - input, output, cached tokens - converts to credits at published per-token rates for each model. Plan prices are not changing. What changes is that your monthly plan fee now buys you a credit balance rather than a fixed number of requests, and heavy usage draws from that balance at the real model rate.
Every model, every rate
GitHub now publishes exact per-token pricing for all models inside Copilot. The numbers match what providers charge directly - no markup on the model layer. The rates below are per million tokens from GitHub's docs (April 29, 2026).
| Model | Input /1M | Cached /1M | Output /1M |
|---|---|---|---|
| Claude Sonnet 4.6 | $3.00 | $0.30 | $15.00 |
| Claude Opus 4.7 | $5.00 | $0.50 | $25.00 |
| Claude Haiku 4.5 | $1.00 | $0.10 | $5.00 |
| GPT-5.5 | $5.00 | $0.50 | $30.00 |
| GPT-5.4 | $2.50 | $0.25 | $15.00 |
| GPT-5.4 mini | $0.75 | $0.08 | $4.50 |
| GPT-5.2 / GPT-5.2-Codex | $1.75 | $0.18 | $14.00 |
| GPT-5 mini | $0.25 | $0.03 | $2.00 |
| Gemini 3.1 Pro | $2.00 | $0.20 | $12.00 |
| Gemini 3 Flash | $0.50 | $0.05 | $3.00 |
GitHub also has two fine-tuned models: Raptor mini ($0.25 input/$2.00 output) and Goldeneye ($1.25/$10.00). These handle autocomplete and are not billed in credits on paid plans.
What costs what: autocomplete, chat, and agents
The thing a lot of coverage got wrong: code completions are still free. If you use Copilot mainly for inline suggestions, nothing changes for you on June 1. The credit system only touches explicit model calls.
Here's what each mode actually costs on Claude Sonnet 4.6 - the model most developers reach for in Copilot chat:
| Mode | Tokens (est.) | Cost per unit | Credits |
|---|---|---|---|
| Autocomplete (inline) | N/A | Free | 0 |
| Chat exchange (light) | 1.5K in, 300 out | $0.009 | ~0.9 |
| Chat exchange (with file context) | 8K in, 1K out | $0.039 | ~3.9 |
| Agent task (simple refactor) | 15K in, 3K out | $0.09 | ~9 |
| Agent task (multi-file) | 50K in, 10K out | $0.30 | ~30 |
| Agent session (complex, 2h) | 200K in, 40K out | $1.20 | ~120 |
The jump from chat to agentic is where this gets real. A 10-message chat session costs about $0.09. A complex two-hour agent session on the same model costs $1.20 - and that's at Sonnet 4.6 rates. Switch to Claude Opus 4.7 ($5 input, $25 output) and multiply everything by roughly 2-2.5x. The developers who are going to hit their credit limits first are not the chat users. They're the people running Workspace on hard problems.
One thing GitHub's new model also removes: the fallback. On the old PRU system, hitting your limit meant Copilot switched you to a cheaper model. On credits, you stop until next month or you buy more. That's a harder wall.
How this plays out over a month
Copilot Pro gives you 1,000 AI Credits ($10) per month. Copilot Pro+ gives you 3,900 ($39). Here's roughly what different usage patterns cost in credits - all using Claude Sonnet 4.6 for model calls:
| Developer type | Typical monthly use | Credits/mo | Plan fit |
|---|---|---|---|
| Autocomplete-first | Mostly tab completions, occasional chat | ~50 | Copilot Pro |
| Regular chat user | ~50 chat sessions/month, no agents | ~450 | Copilot Pro |
| Mixed (chat + light agents) | Daily chat + 5 agent tasks/month | ~550 | Copilot Pro |
| Heavy agent user | Daily chat + 20 agent tasks/month | ~2,200 | Copilot Pro+ or overflow |
| Power (daily agentic, Opus 4.7) | Multi-hour sessions most days | 4,000+ | Likely overflow |
Credit estimates based on GitHub's published token rates for Claude Sonnet 4.6. Actual usage varies by context window size, file count, and model selection.
How Copilot compares to Cursor, Windsurf, and Amazon Q
The honest version of this comparison is that the tools use very different pricing structures and it's genuinely hard to compare apples to apples. GitHub now has the most transparent pricing in this category - you can see the exact per-token rate for every model. Cursor and Windsurf don't publish granular token rates for users.
| Tool | Base price | Overflow / overage | Billing model |
|---|---|---|---|
| GitHub Copilot Pro | $10/month | $0.01/credit (~API rate) | Included credits + metered |
| GitHub Copilot Pro+ | $39/month | $0.01/credit (~API rate) | Included credits + metered |
| Cursor Pro | $20/month | API rate on overflow | Flat + usage overage |
| Cursor Pro+ | $60/month | API rate on overflow | Flat + usage overage |
| Windsurf Pro | $20/month | API rate on overflow | Flat + usage overage |
| Amazon Q Developer Pro | $19/user/month | None | Pure seat-based |
| JetBrains AI Pro | ~$8.30/month | $1/AI Credit top-up | Included credits + top-up |
Amazon Q is the one genuinely different model here. $19/seat, no overage, no token billing, no per-model rates. If you're running heavy agentic workloads and want a fixed bill, that's the argument for Q - not that it's necessarily cheaper for light use, but that the cost is predictable regardless of how much the agent churns.
Cursor and Windsurf are converging toward the same hybrid model as Copilot - flat fee with metered overflow. None of them publish granular per-token pricing for users. GitHub is actually the most transparent of the three now, which is a position I didn't expect them to land in.
Who actually gets hit by this
Most Copilot users - the ones using it for autocomplete with occasional chat - will not notice a change. The math just doesn't add up to a problem. 50 chat sessions at ~4 credits each is 200 credits. A 1,000-credit monthly allowance handles that with room to spare.
The developers who will hit limits are the ones who have started treating Copilot like Claude Code or Codex - long agentic sessions, multiple parallel runs, heavy model use on frontier models like Opus 4.7 or GPT-5.5. A single complex Workspace session on Opus 4.7 at $5 input / $25 output can easily cost 200-400 credits. Run a few of those per week and you're over your Pro+ allowance.
GitHub is giving Business and Enterprise plans a promotional boost for June, July, and August: Business gets $30 of credits instead of $19, Enterprise gets $70 instead of $39. That's their way of buying goodwill during the transition for teams most likely to notice the change.
Run your own Copilot cost math
Token costs for Claude Sonnet 4.6, GPT-5.4, GPT-5.5, and every other model inside Copilot are on TokenCost's pricing page. Compare models or estimate monthly spend before the June 1 switch.
Questions
Does the metered billing affect code completions?
No. Code completions, Next Edit Suggestions, and tab completions remain unlimited on all paid Copilot plans. Metered AI Credits only apply to explicit model calls - chat, edit mode, agent mode, Copilot Workspace, and Spark.
How many credits does a typical agent task use?
A simple refactor on Claude Sonnet 4.6 costs roughly 9 credits ($0.09). A complex multi-file task runs 30-120 credits ($0.30-$1.20). Using Opus 4.7 instead of Sonnet 4.6 multiplies those numbers by about 2-2.5x because of the higher input and output rates.
When exactly does the billing change?
Monthly plan users switch automatically on June 1, 2026. Annual plan subscribers stay on the old premium request unit system until their current term expires, then transition to monthly/credit billing when they renew.
What happens if I run out of credits?
Access to premium models stops until credits replenish next month. There is no longer a fallback to a cheaper model - that option was removed. You can purchase additional credits at $0.01 each, or enterprise admins can set budget caps at the org, cost center, or user level to prevent overruns.