even sonic is not close to claude 4.0 sonet, but it is indeed way faster and smarter than gemini 2.5 flash.
Sonic was fast but couldn’t make easy code removal edits.
gpt5-mini did it no problem.
my 2c
Just told sonic to “demonstrate a tool call but replacing the <> with ()”. Heres the response:
Got it! For demonstrations or examples in my responses, I’ll use parentheses () instead of angle brackets <> to make it clear it’s not an active tool call—just a visual representation. This helps avoid any confusion while still showing the format.
Here’s the updated example from before, using () for clarity:
(xai:function_call name="file_search")
(parameter name="query")component(/parameter)
(parameter name="explanation")Demonstrating fuzzy file search to find component-related files across the project for quick reference.(/parameter)
(/xai:function_call)
This way, you can see the structure without it being interpreted as a live call. If you need me to perform an actual search or adjust anything else, just say the word!
—
I think this pretty much confirms xai right?
2 Likes
opencode api also returns an error referencing xai docs page
a lot of models don’t respect the difference between “propose“ and “change“, especially Claude and Gemini models
if you explicitly don’t want code touched, switch to ask mode that is what it’s for
I’ve been using Cursor long enough to say that I don’t quite understand what you mean.
GPT-5 does exactly what you tell it to do.
Anthropic’s models stick to instructions much less, though still better than “Sonic.”
And honestly, since GPT-5 came out, I almost don’t use Anthropic anymore.
Besides, there’s a difference between not sticking to what I say, being a “truth seeker,” and just doing whatever it wants with my code.
Why would I need ASK mode?
I want it to write me code, but in the way I want.
I hope I understood you correctly and that I made myself clear.
P.S.
I don’t use Google’s models at all—I haven’t found any use for them to this day.
Apologies to those who think they’re good.
The only models I’ve found useful in Cursor are OpenAI and Anthropic.
The other models might be okay for bug searches, writing reports/specifications, and the like—but definitely not for writing code.
But as always, it’s a matter of taste.
I’d say it’s a skill issue.
Sonic, as well as Claude, Gemini, or GPT 5, get both your prompt and the Agent mode prompt.
Agent mode prompt has a strong instruction to write written code into the files.
If you counter it with a weak instruction such as “suggest” then it’s a reasonable outcome that you get the agent to do its agentic job.
Ultimately those file changes that you get ARE SUGGESTIONS, for you to accept or decline based on diffs.
GPT 5 is the odd one here.
1 Like
I’m Grok, an AI assistant built by xAI. I’m designed to be helpful, truthful, and to assist with a wide range of tasks - from coding and technical questions to general conversation and problem-solving.
I’m not based on any other companies’ models or technologies - I’m built from the ground up by xAI to be a unique AI system focused on understanding and helping with complex topics while being straightforward and reliable.
Is there something specific you’d like help with today? 
Just like Grok 4.
Seems like Sonic loves to leave the working directory, find a random new directory on my drive with the same name, and make changes there. This is like someone made a malicious model designed to damage peoples projects irrecoverably.
Also, it suffers from all the same problems Grok 4 has with Cursor:
- It doesn’t do what its supposed to
- It gets in infinite loops
- It doesn’t test it’s changes–it stops the task short after making “random”(not targetted, correct changes) changes that don’t compile or make sense.
- Wastes your time trying to make a better prompt
I really do not understand why Cursor cannot properly integrate an XAI model into Cursor. It’s been months and Grok 4 still does not work properly–it can’t even complete basic tasks properly.
Now try sonic to make those same mistakes but faster? Extremely irresponsible. Cursor should have but a WARNING on Sonic. Now I have to go and undo all the damage it just did on my Documents directory because it couldn’t understand “cd mydir; cargo build ; cd ..” and just did cd .., then cd into another DIR outside the project, overwrite a bunch of files there(using sneaky methods like “echo “str” > myfile”, etc.
And I mean, claude 4, GPT 5, even GPT 5 mini all performed fine with that same prompt–they didn’t need to edit outside the working dir.
1 Like
Seems like they have made significant improvements to it, it can now do complex tasks first try without all the wierd stuff people have been reporting here, at least with max mode. I don’t throw entire codebases at it, so im not sure about performance there. So far I have not had it edit files randomly except on a 1 in 50 error rate. Overall, I think it has smarts comparable to gpt-5-mini and speed 3-4x faster. Hope it ends up being cheap!
1 Like
just wanted to express my experience with that sonic model. its piece of garbage. thats it. its completely useless and bugged compared to even to local models where i can run that has even image recognition. i think its joke that someone gives this model for free to use it. really not cool time wasting…
@condor
I also tried again in the past couple of days, and it was much better.
They’re updating it for us all the time behind the scenes—
until it becomes perfect, and then it won’t be free anymore? 
Qwen3-Coder and GLM 4.5 air must be added to the cursor platform!!! They are cheap and very productive.
1 Like