Gemini 2.5 pro context window seems to be much less than 1M

[divyam25]

Type: Bug

While using Gemini 2.5 pro in both Ask/Agent mode in Copilot Chat, I generally dont use #codebase since I realise the codesearch that copilot performs is quite poor and doesnt seem to include all the neccesary files as context. So i tend to manually include all files in the context.

Now, the models like Claude 3.7 and others have context window around 128k, but Gemini 2.5 pro has 1M context window. But when I am trying to attach approximately 10-15 python files as context with Gemini 2.5 pro, I see the "yellow markers" on some file saying "context window limitation. However, when I attach the same files with Gemini 2.5 pro (exp 03-25) from OpenRouter in Copilot Chat, I dont get context window limitations.

Note - I have copilot pro subscription. My VSCode is at latest version 1.99.3

My question is : Is the context window size of Copilot Pro version of Gemini 2.5 pro very less than the original 1M? If so, this is very sad and destroys the purpose of using Gemini over other models in the first place.

If this is just a bug, fix it!

Extension version: 0.26.5
VS Code version: Code 1.99.3 (17baf84, 2025-04-15T23:18:46.076Z)
OS version: Windows_NT x64 10.0.26100
Modes:

Logs

Info: Initializing Git extension service.
Info: Successfully activated the vscode.git extension.
Info: Enablement state of the vscode.git extension: true.
Info: Successfully registered Git commit message provider.
Info: Logged in as divyam25
Trace: CodeSearchRepoTracker.tryInitGitRepos(): started
Trace: CodeSearchRepoTracker.tryInitGitRepos(): Found initial repos: [].
Debug: [context keys] Updating context keys.
Trace: CodeSearchRepoTracker.tryInitGitRepos(): Complete
Info: Got Copilot token for divyam25
Debug: [LanguageModelAccess] UPDATING language models
Trace: Resolving embedding model
Trace: Should request permissive session upgrade: false - already have permissive session
Info: copilot token chat_enabled: true, sku: free_educational
Info: GitHub.vscode-pull-request-github extension is not yet activated.
Info: Registering default platform agent...
Debug: onDidChangeSessionsHandler has token: true
Debug: [context keys] Updating context keys.
Trace: Resolving embedding model
Info: Fetched model metadata in 492ms 22f8e5f0-f94b-4e47-9b0e-48d103968ebe
Trace: Resolved embedding model
Trace: Resolved embedding model
Info: activationBlocker from 'languageModelAccess' took for 1041ms
Trace: Resolving chat model
Trace: Resolved chat model
Debug: [LanguageModelAccess] ADDED language model gpt-3.5-turbo
Debug: [LanguageModelAccess] ADDED language model gpt-4o-mini
Debug: [LanguageModelAccess] ADDED language model gpt-4
Debug: [LanguageModelAccess] ADDED language model gpt-4-0125-preview
Debug: [LanguageModelAccess] ADDED language model gpt-4o
Debug: [LanguageModelAccess] ADDED language model o1
Debug: [LanguageModelAccess] ADDED language model o3-mini
Debug: [LanguageModelAccess] ADDED language model claude-3.5-sonnet
Debug: [LanguageModelAccess] ADDED language model claude-3.7-sonnet
Debug: [LanguageModelAccess] ADDED language model claude-3.7-sonnet-thought
Debug: [LanguageModelAccess] ADDED language model gemini-2.0-flash-001
Debug: [LanguageModelAccess] ADDED language model gemini-2.5-pro
Debug: [LanguageModelAccess] ADDED language model o4-mini
Debug: [LanguageModelAccess] ADDED language model gpt-4.1
Debug: [LanguageModelAccess] DID UPDATE language models
Debug: [LanguageModelAccess] UPDATING language models
Trace: Resolving chat model
Trace: Resolved chat model
Debug: [LanguageModelAccess] DID UPDATE language models
Debug: [context keys] Window state change. Needs offline check: false, active: true, focused: false.
Debug: [context keys] Window state change. Needs offline check: false, active: false, focused: false.
Debug: [context keys] Window state change. Needs offline check: false, active: false, focused: true.
Debug: [context keys] Window state change. Needs offline check: false, active: true, focused: true.
Debug: [context keys] Window state change. Needs offline check: false, active: true, focused: false.
Debug: [context keys] Window state change. Needs offline check: false, active: true, focused: true.

System Info

Item	Value
CPUs	11th Gen Intel(R) Core(TM) i7-11800H @ 2.30GHz (16 x 2304)
GPU Status	2d_canvas: enabled canvas_oop_rasterization: enabled_on direct_rendering_display_compositor: disabled_off_ok gpu_compositing: enabled multiple_raster_threads: enabled_on opengl: enabled_on rasterization: enabled raw_draw: disabled_off_ok skia_graphite: disabled_off video_decode: enabled video_encode: enabled vulkan: disabled_off webgl: enabled webgl2: enabled webgpu: enabled webnn: disabled_off
Load (avg)	undefined
Memory (System)	15.71GB (8.27GB free)
Process Argv	--crash-reporter-id 6eb2a5e6-d4df-42f4-837c-d879b56b2306
Screen Reader	no
VM	0%

A/B Experiments

vsliv368:30146709
vspor879:30202332
vspor708:30202333
vspor363:30204092
vscod805cf:30301675
binariesv615:30325510
c4g48928:30535728
azure-dev_surveyone:30548225
2i9eh265:30646982
962ge761:30959799
h48ei257:31000450
pythontbext0:30879054
cppperfnew:31000557
dwnewjupytercf:31046870
pythonrstrctxt:31112756
nativeloc2:31192216
5fd0e150:31155592
dwcopilot:31170013
6074i472:31201624
dwoutputs:31242946
customenabled:31248079
9064b325:31222308
copilot_t_ci:31222730
e5gg6876:31282496
pythoneinst12:31285622
bgtreat:31268568
4gafe986:31271826
31787653:31262186
3e8i5726:31271747
996jf627:31283433
usemplatestapi:31288272
7bj51361:31289155
747dc170:31275177
aj496949:31278748
aj953862:31281341
convertfstringc:31290957
gendocc:31290959

[isidorn]

Thanks for you feedback. Yes the context window size is smaller than the original 1M.
We are continuously increasing rate limits as we get more capacity.

fyi Logan Ramos (@lramos15)

[Xzensi]

Isidor Nikolic (@isidorn) you said "We are continuously increasing rate limits as we get more capacity." but OP asked about context window size, not rate limits. Although you did confirm the intentional reduction in context window size, did you mean context window size will get increased to the advertised full 1 million when there's more capacity for it?

[isidorn]

There is actually more work than just getting more capacity to support the full 1 million context.
E.g. engineering work on client side to make sure caching works well for large contexts, and work on service side to support large context windows.

We have this on the plan. And I do expect it to land later this year.

[Qualzz]

Can we get at least a estimate of what are the context length for each model ? It's kinda weird to pay for a service without even knowing what limitation this service has. This is really opaque nor logical.

Why would 3k line of code not fitting in the context windows of a "1M" model, but fit for other models that natively handle way less contexte ?

The only purpose of Gemini is to understand bigger chunk of code and it's one of the model with the less working context as of now. Which was not disclosed when I have upgraded to pro.

I'll be honest, I feel copilot is getting worse every day.

• gpt 4o are basically useless
• autocomplete is a mess
• next autocomplete suggestion is really annoying to use
• Sonnet 3.7 outputs are broken most of the time (yes) ( several duplicates edits of the same file in the same run, making the whole file a broken mess )
• Gemini context length crippled to death ( come on, I can't put 3k line of code ( 30k token on ai studio, in a 1M token model, that mean I can use less THAN 32 TIMES the model context length )
• o3 and 04 mini are totally stupid and change things you didn't asked and hallucinate a lot ( you can NEVER trust it's outputs, you ask for a little addition, and you end up with +3 -126 )

The current stage of github copilot is a very frustrating product that i'm often asking myself if I should continue to pay or not, just because ONCE in a while it does a perfect job.

Exemple of sonnet edit (I didn't have to try hard, I just asked to add a single filter, which should be 4 lines of code) :

Edit: I've tried the exact same prompt on cursor with Gemini: It one shooted the task, no issue, perfect code, no context length limitation. None of the models on copilot managed to do it, which means the issue are not the model, but the vs code code implementation/system prompts or whatever.

[isidorn]

Thank you for your feedback.

You can checkout the context window by writing a VS Code extension, using the lm API, list the models and see the context window. In this case, the limit is currently at 64K. I do agree, that making this transparent to the users makes sense (maybe in dropdown) Logan Ramos (@lramos15)

Your sonnet edit looks like a bug that we are investigating. I expect us to fix this week.

[Qualzz]

In this case, the limit is currently at 64K.

This should be for sonnet and gpt models.
Is this including 34k line of internal vs code system prompt then ? Because using the AI studio, I can see that my file is 30k tokens, yet I have the "Part of this file was not sent due to context length blablablabla" yellow warning on the file, when using Gemini. ( I don't have any other included file, or custom instructions etc, i'm talking about only adding this single file as context )

[Xzensi]

Qualzz Do note that "Part of this file was not sent due to context length" is not necessarily due to the hardcoded limit of input tokens length but general optimizations such as where some (perhaps naïve) assumptions were made on what "sufficient" amount of context is to accomplish most tasks without by default always sending the full files.

[Qualzz]

Thank you for your answers.

Then it would be nice to be able to disable that because I don't want any assumptions about the context ( at least not in edit mode, in agent mode why not, that's part of the agentic behavior ), but when i'm giving myself the context, I don't want it to be altered invisibly. There isn't a single case where this "feature" helped more than doing harm. I still cancelled my pro subscription, but i'll keep an eye on the progress, but right now this is adding more headache than solving ones.

Not just only all the points I mentioned before, but when you add very frequents "no response were returned" and similar responses, it's just too unreliable.

[JohnDotOwl]

Isidor Nikolic (@isidorn) Could you guys consider increasing the context window, charge more credits if you need to , since GitHub copilot is a paid software now and there are limitations in place, please increase the context window, i'm on enterprise plan and we're wiling to pay accordingly

[isidorn]

Jonathan T. (@JohnDotOwl) we are working on this and I expect the gemini context window to increase soon

[JohnDotOwl]

@JohnDotOwl we are working on this and I expect the gemini context window to increase soon

Thanks for your quick reply. What about sonnet 4.0 will it also be increased.

128k token window for VS Code Insiders is only for GPT-4o right?