The surprise on the bill
I recently audited my GitHub usage for a single day and found something that caught me off guard:
the Copilot Cloud Agent line item was larger than everything else on the bill combined — and I usually pick a low-cost model (Kimi).
Here is the raw data from July 11, 2026.
The numbers
| SKU | Units | Price/unit | Gross amount |
|---|---|---|---|
| Actions Linux | 101 min | $0.006 | $0.61 |
| Codespaces compute 2-core | 5.06 hr | $0.18 | $0.91 |
| Codespaces storage | 0.18 GB-hr | $0.07 | $0.01 |
| Copilot AI Credits | 103.44 credits | $0.01 | $1.03 |
| Copilot Cloud Agent | 863.9 credits | $0.01 | $8.64 |
Daily total: $11.20
Cloud Agent share: ~77%
Even though both Copilot AI Credits and Copilot Cloud Agent charge the same $0.01 per credit, the agent consumed 8.4× more credits than all my other Copilot AI usage that day.
Daily cost breakdown
Hover over the bars to see exact amounts.
Model usage for the month
The model-level view tells the same story. The Coding Agent model bar is an order of magnitude taller than every other category combined.
| Model category | Cost (Jul 1–12) |
|---|---|
| Coding Agent model | $12.24 |
| Code Review model | $0.92 |
| Auto: GPT-5.3-Codex | $0.61 |
| Claude Sonnet 4.5 | $0.59 |
| Claude Haiku 4.5 | $0.41 |
| Other models | $0.18 |
That is not because I selected an expensive model; it is because the agent layered many model calls into one autonomous session.
A simple visual comparison
The daily spend in text form:
Copilot Cloud Agent $8.64 ████████████████████████████████████████
Codespaces compute 2-core $0.91 ████
Copilot AI Credits $1.03 █████
Actions Linux $0.61 ███
Codespaces storage $0.01 ▏One Cloud Agent day cost roughly as much as running six Codespaces sessions or thirteen Actions workflows.
Why it costs so much more
The key insight is that Copilot Cloud Agent is not the same product as Copilot AI Credits.
When you use Ask, Edit, or inline chat, you pay for discrete model calls. When you use Cloud Agent, you pay for an autonomous runtime that can:
- Read and search across many files
- Run shell commands and tools
- Iterate, reflect, and retry
- Keep a large context window alive across multiple turns
- Run for up to ~59 minutes per session So even with Kimi — a cheap model on a per-token basis — the sheer number of calls and the length of the session add up.
Think of the two SKUs this way:
| Product | What you are paying for |
|---|---|
| Copilot AI Credits | Individual model invocations that you trigger directly |
| Copilot Cloud Agent | The entire agentic workflow: models plus compute plus tool orchestration |
Optimization tips
If you want to keep Cloud Agent useful without letting the bill run away:
- Scope sessions tightly. Give the agent one concrete task instead of “fix everything in the codebase.”
- Limit the working set. Point it to specific files or directories; large context windows multiply input tokens.
- Watch long-running tasks. Each Cloud Agent session can run for close to an hour. Cancel early if it is drifting.
- Review the model actually used. Check the AI usage dashboard to confirm it is running on the model you selected.
- Use direct chat for quick questions. Reserve Cloud Agent for tasks that genuinely require multi-step reasoning and tool use.
Takeaway
The price of Copilot Cloud Agent is not driven by the model alone. It is driven by session count, context size, and the number of tool/model interactions inside each run.
If your Cloud Agent bill feels high even with a low-cost model, that is normal: the agent SKU bundles compute, tooling, and repeated model calls into one credit pool. Treat it as a workflow product, not a chat product, and scope sessions accordingly.