GitHub Community
Why can't we fully utilize context_window? (Data from Copilot's own API) #186340
Replies: 6 comments 17 replies
Evidence: Raw API ResponseHere's how to reproduce this data using # 1. Get your Copilot token (requires authentication)
TOKEN="<YOUR_COPILOT_TOKEN>"
# 2. Call the official API
curl -s "https://api.githubcopilot.com/models" \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-H "Copilot-Integration-Id: vscode-chat" \
-H "Editor-Version: vscode/1.96.0" \
-H "Editor-Plugin-Version: copilot-chat/0.35.0" \
-H "User-Agent: GitHubCopilotChat/0.35.0" \
-H "X-GitHub-Api-Version: 2025-10-01"Key Model Limits (from API response on 2026-02-04)
Utilization = (max_prompt + max_output) / context_window Raw JSON Sample (gpt-5.1-codex-max){
"id": "gpt-5.1-codex-max",
"name": "GPT-5.1-Codex-Max",
"capabilities": {
"limits": {
"max_context_window_tokens": 400000,
"max_output_tokens": 128000,
"max_prompt_tokens": 128000
}
}
}This confirms that while |
|
max_prompt doesn't seem to calculate the full input, possibly excluding thinking. Previously, when using the claude sonnet model via chat completions, I've used over 200K of context. They probably have their own internal implementation; with gpt-5.2, I often use over 200K. You'll notice it once you turn off automatic compaction. @cuipengfei |
|
yes, but still, would be nicer if we can use all 400k 😄 |
I have Pro+ Plan. I'm using Claude Opus 4.6, which supports a 1M context length. Limiting it to 128K reduces its ability to understand the full picture. In this case, the model needs access to the entire context to make precise changes without wasting time iterating over and over. When I use the OpenRouter API with GitHub Copilot, it correctly uses the full context length. That's exactly how I want it to behave. but the current limitation is preventing the model from operating at full capacity. Therefore is better for me to keep Copilot Pro not plus, mainly for autocomplete and occasional questions, and rely on the OpenRouter API for any large context requests. |
|
@cuipengfei Yes I have 39$ Pro+ plan and yes same limits. That's why I said I will go back to 10$ Plan and use copilot with Openrouter API for long important things. 
|
@SoftwareLogico You can use opencode and disable compaction in opencode. Normally, models can be used up to 200k, including the Claude model, but this area is a gray zone. Avoid frequently exceeding 200k; within 200k there shouldn't be any issues. Currently, VSCode Copilot has also integrated the Claude code agent, which should support a context of 200k. |
|
@SoftwareLogico @cuipengfei I tested with gpt-5-mini, which can be used up to 200k+, but if the first user message exceeds 128k, it will be detected and an error will be prompted: prompt tokens exceed 128k. However, if the conversation reaches 128k after many rounds of dialogue, no error will occur. |
|
|
I did noticed that, however what it seems to do it to either truncate or summarize previous message to keep it under copilot restriction. |
|
What's the rationale behind this, Github? Seems like a very, very stupid limitation. |
|
Having the same issue here. Does Github team not have context limits on their "insider developer" plan - so they can using AI to write code, and forgot about updating the base 128K context limit? |
|
I found a website (https://github.blog/changelog/2024-12-06-copilot-chat-now-has-a-64k-context-window-with-openai-gpt-4o/) from 2024 that says regular users can use 64K context, while the insider version can use 128K context. Maybe now only the insider version has a longer context. |
|
Just a reminder... the actual limits of sonnet 4.6 is 1 million tokens! That's exactly what you get when you use other providers (like AWS Bedrock or Anthropic). 128k is LAUGHABLE. |
aha, that is good to know, i will try vs code insider to find out |
|
I tested this. Queried the Copilot backend API directly ( Result: Almost no difference. Only 2 legacy models changed:
All other 41 models — Claude, GPT-5.x, Gemini, Grok, etc. — have identical So the 2024 blog post about insider getting bigger context was specific to the old gpt-4o era. For current models, insider doesn't unlock anything extra. |
|
@kangkannnng so, only for 4o |
|
For now I will stop paying Github and switch a better AI provider. Thanks Microslop |
Uh oh!
There was an error while loading. Please reload this page.
Why can't we fully utilize context_window? (Data from Copilot's own API)
TL;DR
The Copilot API returns large
context_windowvalues (e.g., 400K tokens), but the actual usable capacity is limited bymax_promptwhich is often much smaller (e.g., 128K). This means users cannot fully utilize the reported context window.Data Source
This data comes directly from GitHub Copilot's own API:
The response includes three key fields for each model:
limits.context_window— Reported total context capacitylimits.max_prompt— Actual maximum input tokens allowedlimits.max_output— Maximum output tokens per responseThe Problem: Utilization Rate
Utilization = (max_prompt + max_output) / context_window
Key Observations
gpt-5.1-codex-max has "max" in its name but only has 64% utilization — the same as regular gpt-5.1-codex
gpt-5.2-codex is the only GPT-5 series model that can fully utilize its 400K context window
Users hitting
prompt token count exceeds the limit of 128000errors are confused because they see 400K context_window in the API responseQuestions for GitHub
Why is there such a large gap between
context_windowandmax_promptfor most models?Is this a deliberate limitation imposed by GitHub Copilot, or inherited from the underlying models?
Will this be improved? Many users are hitting the 128K limit and don't understand why they can't use the full 400K.
Can you document this clearly? Currently there's no official documentation explaining these three limits and their relationship.
Related Issues
Environment
https://api.githubcopilot.com/modelsAll reactions