Formal complaint: Copilot AI Credits are exhausted by day 11 despite lighter usage than previous monthseedback about the GitHub Discussions product>

[AmedeoPelliccia]

Discussion Type

General

Discussion Content

Hi GitHub Copilot team,

I would like to express a formal complaint and strong dissatisfaction with the new Copilot AI Credits model.

Today is only day 11 of the month, and I am already unable to continue my normal work with Copilot because my included AI credits are almost exhausted.

This is happening despite the fact that I have used Copilot much less intensively than in previous months. Under the previous premium-request model, I was generally able to work through the month without this kind of immediate blockage. With the new AI Credits model, the included allowance is being consumed far faster, even when I am deliberately avoiding heavy usage and not intentionally selecting the most expensive workflows.

From a user perspective, this feels like a severe degradation of the product.

Concrete example

One ordinary prompt containing a small SVG technical file moved my displayed included AI credits usage from 83% to 85%.

The prompt was not a large dataset, not a repository-wide task, not a long agentic session, and not a cloud-agent workflow. It was a normal technical documentation task involving a structured SVG file.

The selected model shown in the UI was Claude Fable 5.

After that single prompt, GitHub displayed:

You've used 85% of your included AI credits. Limit resets on Jul 1.

This means that one normal prompt visibly consumed a meaningful percentage of the monthly allowance.

I understand that percentage displays may be rounded, but the user-facing result is still unacceptable: a routine developer/documentation prompt can materially reduce the monthly allowance, without any itemized explanation.

The practical consequence

This is not a theoretical complaint.

The practical consequence is that I cannot continue my work normally on day 11 of the month.

For a paid professional development tool, this is a serious problem.

A developer should not have to stop working after 11 days while using the product less than in previous months. This makes the new model feel economically unfair, operationally unpredictable, and unsuitable for daily professional use.

The core issue: lack of transparency

The problem is not that AI inference has a cost.

The problem is that the cost is not inspectable.

For each billable interaction, I cannot see:

• how many input tokens were billed;
• how many output tokens were billed;
• how many cached tokens were billed;
• whether repository context was automatically included;
• whether conversation history was included;
• whether hidden or automatically injected context was added;
• whether any agent/tool sub-calls occurred;
• which exact model pricing was applied;
• how many AI credits were charged for that specific prompt;
• why that prompt consumed that amount.

Without this information, users cannot manage their usage in a rational way.

This is especially problematic for developers working with technical documentation, SVG, XML, Markdown, YAML, JSON, configuration files, or structured engineering artifacts. These are normal software and documentation workflows, not exceptional workloads.

Why this feels unfair

The new model appears to replace a more predictable monthly allowance with an opaque usage meter.

The nominal number of credits may look larger than the previous premium-request allowance, but in practice the value of the allowance can be much lower depending on hidden context, tokenization, model routing, cached context, repository context, and internal Copilot behavior.

That creates a serious trust issue.

From the user side, this feels like a bait-and-switch:

• the subscription still looks like a monthly professional plan;
• but the included allowance can be exhausted very early in the month;
• the user cannot see exactly why;
• the user cannot set a hard per-prompt limit;
• the user cannot reliably predict cost before sending a prompt;
• the user cannot audit the billing after the interaction.

That is not acceptable for a professional developer tool.

Requested clarification

Please clarify:

1. Why am I almost out of included AI credits by day 11 despite using Copilot less than in previous months?

2. Why can one ordinary SVG prompt move the displayed usage from 83% to 85%?

3. Does Copilot Chat automatically include repository context even when the user only pastes a file?

4. Does Copilot Chat automatically include previous conversation history in billable context?

5. Are cached tokens charged against the monthly included AI credits?

6. Are hidden system/tool/agent calls charged to the user?

7. Can users see a per-interaction breakdown of tokens and AI credits?

8. If not, why is usage-based billing being enforced without a usage ledger?

9. Does GitHub plan to introduce pre-send cost estimation?

10. Does GitHub plan to introduce hard per-prompt credit caps?

Requested changes

At minimum, Copilot should provide:

1. Per-interaction billing ledger

Each billable interaction should expose:

interaction_id: "..."
timestamp: "..."
feature: "Copilot Chat"
mode: "Ask"
model: "Claude Fable 5"

user_input_tokens: "..."
repository_context_tokens: "..."
conversation_history_tokens: "..."
cached_tokens: "..."
output_tokens: "..."

agent_subcalls: "..."
tool_subcalls: "..."
hidden_context_included: true/false

ai_credits_charged: "..."
estimated_cost_usd: "..."

2. Pre-send cost estimation

Before sending a prompt, Copilot should show an estimate such as:

Estimated cost: 80–150 AI credits

Included context:
- Current prompt: yes
- Current file: yes/no
- Repository context: yes/no
- Conversation history: yes/no
- Agent/tool calls: possible/not possible

3. Explicit context controls

Users should be able to choose:

• use only this prompt;
• use only current file;
• use selected files only;
• do not use repository context;
• do not use conversation history;
• do not use cached context;
• use full repository context;
• allow/disallow agentic sub-calls.

4. Hard per-prompt credit caps

Users should be able to set a maximum spend per interaction, for example:

Do not spend more than 25 AI credits on this prompt.

If the prompt would exceed that cap, Copilot should stop before execution and ask for confirmation.

Final point

Calling this “usage-based billing” is only fair if the user can see and control the usage.

Right now, I cannot.

I am paying for a professional developer tool, but by day 11 of the month I am already effectively blocked from continuing my work, despite using Copilot less than in previous months.

That is not a sustainable or transparent model.

Please make Copilot AI Credits inspectable, predictable, and controllable per interaction.

[chris-rickman]

I am also shocked and disappointed at billing updates. My credits were also drained inside of one 24 hour cycle. This led me to explore other options such as OpenAI Codex. I am pleased to share that the speed and quality of Codex greatly exceeds GHCP. I really like the "steering" feature. In the end, I couldn't be happier. Things are better than ever.