Copilot inputs cause a credits explosion

[aiis2]

🏷️ Discussion Type

Question

Body

Copilot inputs cause a credits explosion. You'll discover this issue when customizing the official Deepseek model using Copilot. Let's test this with a small scenario. Modify 10 lines of code and observe usage information from Deepseek model vendors: 2.95 million tokens input, 6003 tokens output. This explains why the credit limit is consumed so quickly. Looking at billing, using Claude 4.6 Sonnet as an example, every million tokens earns 300 credits, meaning this input consumes nearly 900 credits. So why are Copilot's inputs so high?

[gopiprakan]

The high token usage is usually caused by how Copilot constructs the context sent to the model, not just by the 10 lines you changed.

When you ask Copilot to modify code, it may include:

The entire file being edited
Related files and dependencies
Open editor tabs
Workspace context
Previous chat history and instructions
Additional metadata used for code understanding

As a result, a seemingly small change can generate a very large prompt. If the workspace contains large files or extensive context, the model may receive millions of input tokens even though the actual code modification is only a few lines.

In your example, 2.95M input tokens for a 10-line change suggests that a substantial amount of repository and conversation context was included in the request. Since billing is primarily based on input and output tokens, the majority of the credit consumption comes from the input side rather than the generated response.

It would be helpful if GitHub provided:

A detailed token breakdown per request.
Visibility into which files and context were included.
Controls to limit the amount of workspace context sent to the model.
Warnings before submitting unusually large prompts.

Greater transparency would make it much easier for users to understand and manage their credit usage when working with Copilot and large codebases.

[mo0feng]

Thank you for the tips. I understand your description—I've already used up all 7,000 credits. My project has about 400,000 lines of code, and the session is new. First, I run the plan (not deepseek), then switch to the agent to execute "Start implementation" (deepseek).

[alive-xd]

2.95M input tokens for a 10-line change doesn't seem right. If that's accurate, there should be more transparency around what's actually being sent to the model.

Is Copilot including the entire repository, previous chat history, or additional context behind the scenes? A token usage breakdown would make it much easier to understand why credits are being consumed so quickly.

[06naresh]

Root Cause

Copilot does not only send the 10 lines you edited.

It constructs a large context window that may include:

- Entire file contents

- Related files and dependencies

- Other open editor tabs

- Workspace metadata

- Previous chat history and instructions

Why Credits Spike

Billing is based on input + output tokens.

Even small edits can trigger millions of input tokens

if the workspace or repository is large.

Example: 2.95M input tokens → ~900 credits consumed.

What Can Help

1. Request transparency:

- Token breakdown per request

- Visibility into which files/context are included

2. Limit context:

- Option to restrict Copilot to current file only

- Controls to exclude unrelated tabs/dependencies

3. Add safeguards:

- Warnings before unusually large prompts

- Ability to cap token usage per request

Practical Steps

- Monitor usage in billing dashboard

- Keep fewer files/tabs open when prompting Copilot

- Re‑authenticate if usage looks abnormal

- Raise a support ticket for manual review if credits drain too fast