Source?
Itās crucial to check the actual data before making such definitive claims. Your numbers are quite off compared to the current benchmarks.
According to Context Arena, Gemini 3.0 Pro Preview is actually performing at:
-
128k: 99.0% (Ranked #1) ā vs your 77%
-
1M: 72.2% (Ranked #1) ā vs your 16%
The model isnāt getting āstupiderā; it is retaining high coherence even at 1M context.
Can you give us an ETA for when Gemini 3.0 Pro High will be available?
I have noticed with /summarize command, that it does summarize BUT then also right after that the model will start performing like it was prompted to do something else too.
Thatās normal with all models if you hit /summarize when the model is in the middle of processing a previous command or an instruction.
Check the thinking process. Itāll say something like āI was in the middle of doingā¦.ā
Um, sorry, what? The session is open of course, but it was not performing anything before the command?
Iām just telling you my experience. I use /summarize all the time, in different contexts, to save tokens.
Hmm, I donāt remember it ever happening with sonnet 4.5 for me, but I must confess that I have been quite lazy to use it and more relied on the automatic summarize 
Try using it when your token count usage hits ~50%. Itāll help you save tokens while keeping the agent sharp.
Now that Iām bound to use my own Gemini API due high demand issues, the 50% would mean 500k tokens in the Max mode 
Okay, cannot use this with own api key either anymore. Constant āmodel overloaded, try againā messages. I think google made this model too good, and it is not going to be useful in nearby future.
Either there is something wrong with the model, or there is something wrong with the integration. Gemini 3 seems very weak and stupid in the areas of planning, programming, and coding. In addition to constantly crashing, it does not follow instructions, responds very poorly, and produces poor, unverified solutions.
Looked very promising after going through a few turns, unfortunately now I seem to constantly get āOur model provider is experiencing high demand right now. Please switch to another model, or try again in a few moments.ā - hoping this can be resolved soon, excited to see how it can do on some larger context-heavy refactorings
Yes. I have the same problem.
I already emailed the Cursor team, like, 5 times.
All they say is that this is the way their token usage system works.
But I know for sure that two days ago the token usage was completely different, much, much lower.
The system is broken. In my humble opinionā¦
Cursor has been dropping the ball a lot lately. Extremely unhappy with this.
I cannot even connect my own api key ,i get āAgent and Edit rely on custom models that require a Pro or Business subscription. Please use a Pro or Business subscription and/or disable API keys. Ask should still work.ā
Iām sorry that I got the second number of MRCR v2 mixed up.
Interesting ā I always kind of felt that GPT-5 starts getting dumber at around half of its context window, and in the benchmark its score at 128k was exactly 61.6%.
Gemini 2.5 Pro also clearly got weaker when I shoved a few hundred thousand tokens into it in AI Studio. So Iād trust MRCR v2 more than Context Arena, even though Iām hearing about both benchmarks for the first time.
I have the INVALID_ARGUMENT issue again with Gemini 3 Pro.
- Version: 2.1.7
- VSCode Version: 1.105.1
- Commit: 3d2e45538bcc4fd7ed28cc113c2110b26a824a00
- Date: 2025-11-20T18:51:07.967Z
- Electron: 37.7.0
- Chromium: 138.0.7204.251
- Node.js: 22.20.0
- V8: 13.8.258.32-electron.0
- OS: Darwin arm64 24.6.0
Request ID: adcd162a-f38b-4241-9088-a48b8d66364c
{"error":"ERROR_OPENAI","details":{"title":"Unable to reach the model provider","detail":"We encountered an issue when using your API key: Provider was unable to process your request\n\nAPI Error:\n\n```\nRequest failed with status code 400: [{\n \"error\": {\n \"code\": 400,\n \"message\": \"Invalid JSON payload received. Unknown name \\\"top_k\\\": Cannot find field.\",\n \"status\": \"INVALID_ARGUMENT\",\n \"details\": [\n {\n \"@type\": \"type.googleapis.com/google.rpc.BadRequest\",\n \"fieldViolations\": [\n {\n \"description\": \"Invalid JSON payload received. Unknown name \\\"top_k\\\": Cannot find field.\"\n }\n ]\n }\n ]\n }\n}\n]\n```","additionalInfo":{},"buttons":[],"planChoices":[]},"isExpected":true}
ConnectError: [invalid_argument] Error
at DQc.$endAiConnectTransportReportError (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:6330:408409)
at cPo._doInvokeHandler (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:6946:21873)
at cPo._invokeHandler (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:6946:21615)
at cPo._receiveRequest (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:6946:20377)
at cPo._receiveOneMessage (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:6946:19194)
at fPt.value (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:6946:17286)
at Ce._deliver (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:49:2962)
at Ce.fire (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:49:3283)
at qmt.fire (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:6315:12156)
at MessagePort.<anonymous> (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:8969:18439)