Copilot Token Awareness

Stop being surprised by your AI credit bill.
This extension gives you a live, transparent estimate of how many tokens GitHub Copilot will consume - and what that costs in AI credits - before you send a single message. Use it to understand where your context budget goes, choose the right model, and get more out of every credit.

Screenshot

The Breakdown panel shows Ask/Agent mode, every context source contributing tokens, a step-by-step cost calculation, and a link to official GitHub pricing.

Why This Exists

GitHub Copilot Chat and Agent mode bill in AI credits (1 credit = $0.01 USD) Source. Each request sends a large bundle of context to the model - your active file, snippets from open tabs, a system prompt, custom instructions, and workspace-index retrievals - all before you type a single word of your question.

Most developers have no visibility into this until they see a usage report. This extension surfaces those numbers in real time so you can:

• Understand what is driving your token count
• Choose the most cost-effective model for the task
• Optimise by closing unused tabs or splitting large files
• Budget AI credit usage across a team or project

Important: Code completions and next-edit suggestions are not billed in AI credits and are excluded from all estimates in this extension. Only Ask mode and Agent mode chat interactions are in scope.
Source: GitHub Copilot billing - Code completions

Features

How the Estimate Is Built

Copilot does not expose what it sends to the model. This extension reconstructs a realistic estimate using Copilot's published architecture and empirical measurement. Every assumption is shown in the Estimation Assumptions section of the Breakdown panel.

Context sources (in order)

What is NOT included (by design)

Ask mode vs Agent mode

Switch modes using the Ask / Agent toggle buttons at the top of the Breakdown panel, or via the copilotTokenAwareness.chatMode setting.

Cost Calculation

Input cost = total_tokens × input_multiplier × token_unit_price_usd

Where:

• total_tokens = sum of all context sources above
• input_multiplier = model-specific value from the GitHub Copilot pricing table ($/M tokens ÷ 10)
• token_unit_price_usd = 0.00001 (i.e. 1,000 token units = $0.01 = 1 AI credit)

The Breakdown panel also shows a worst-case output ceiling (expandable) that assumes Copilot fills the entire remaining context window - this almost never happens but shows the absolute maximum exposure.

Supported Models

All pricing is sourced directly from the official GitHub Copilot models and pricing page.

Anthropic

Anthropic models include a cache-write cost in addition to cached-input pricing. The extension uses the non-cached rate (accurate for Turn 1).

Google

OpenAI

⭐ Included models (GPT-4.1, GPT-5 mini) do not consume AI credits within your plan's monthly allowance. The extension flags these with a green banner and notes that the cost shown is the overage rate only.

Fine-tuned (GitHub) & Microsoft

Settings Reference

Keeping pricing current

GitHub may update multipliers or add new models at any time. You don't need to wait for an extension update.

Override a built-in model's multiplier:

// settings.json
"copilotTokenAwareness.modelOverrides": {
  "claude-sonnet-4-6": { "inputMultiplier": 0.35 }
}

Add a brand-new model:

"copilotTokenAwareness.customModels": [
  {
    "id": "my-new-model",
    "displayName": "My New Model",
    "inputMultiplier": 0.2,
    "outputMultiplier": 1.0,
    "contextWindow": 200000
  }
]

Update the base token unit price:

"copilotTokenAwareness.tokenUnitPriceUsd": 0.000012

Commands

Token Counting

Uses tiktoken (cl100k_base encoding) running entirely in WebAssembly inside VS Code - no data is sent to any external service. Falls back to a character-based heuristic (~4 chars/token) if the WASM module fails to load; the status bar shows a warning in that case.

Tokenizer note: cl100k_base is OpenAI's tokenizer. Anthropic (Claude) and Google (Gemini) use their own tokenizers. For typical English and source code the counts are within ±5%, which is within the stated estimate variance. The variance percentages shown in the UI (±10% Ask, ±15% Agent) account for this approximation.

Accuracy & Disclaimer

This extension is an awareness tool, not a billing meter. Estimates are based on:

• Copilot's published architecture and empirically observed behaviour
• Turn 1 only (conversation history is not pre-knowable)
• Non-cached token rates (cached turns are cheaper; the disclaimer in the panel notes this)
• The assumption that Copilot uses the entire active file and snippets from tabs (actual selection may vary by feature and version)

Expected accuracy: ±10% for Ask mode, ±15% for Agent mode.

For official billing information, plan allowances, and current per-token rates, always refer to:
📄 GitHub Copilot - Models and Pricing

License

MIT © Raj Uppadhyay