The fact that it doesn’t accept images is a significant drawback—
because even when I don’t need an image, I can’t switch to it in the middle of a conversation!
If there was an image somewhere earlier in the chat, I can’t let it do anything anymore.
Also, it still does way too many things—far beyond what I ask for—
and it messes up a lot.
1 Like
Good catch! Overall I do not recommend switching models in middle of conversation as that requires sending the whole chat thread to the new provider as context and it can not take advantage of prompt caching at AI provider until the next request or tool call.
3 Likes
the reason I brought up ask mode is because you said
that’s the use-case for ask mode
This model acts very suspiciously like an anthropic model, the copius use of emojis, the same language , even using the same words. The way it says ‘you are absolutely right!’ all the time.. it’s uncanny, the similarities ..
1 Like
Let me put it this way from another angle:
I need to make a change that only GPT-5 knows how to do.
But GPT-5 costs me a lot of money, and I only need it for the actual change. It’s a waste to spend money on it just to locate the right place, the related component, etc.
So, I take the AUTO model, which right now is free, and I ask it to find all the details where, what, and so on.
Only after everything is clear, I switch to GPT-5 to make the change.
Doesn’t that make sense?
Option B: Bug fixing.
The location of the bug is already identified. Now I try model after model to see which one manages to fix it.
Especially for UI-related bugs models still don’t know how to handle UI changes well and fail repeatedly. Sometimes I switch through several models until one succeeds.
1 Like
Hi! That doesn’t really make sense. It won’t save you much money and it will reduce the quality.
It only makes sense if you first ask for a summary in a .md file and then start a new chat using that summary as the starting point.
Why?
What do you know that I don’t, that makes it clear to you this is a mistake?
Condor replied to you. Without caching, your first request to a new model will be very expensive (VERY expensive). At the same time, the conversation context will be of lower quality, since it was built by a less capable model without intermediate processing or reasoning.
A very useful feature has recently appeared: automatic summarization and context compression, which improves the chat’s “sliding window.” This can work well and may be helpful when switching models, effectively replacing the .md file. However, depending on your goals, it can also get in the way—for example, if you want to first read the file code with a cheaper model and then hand over the chat context to a more advanced AI.
So this is a complex and non-obvious issue that heavily depends on the situation and task.
1 Like
Grok code fast 1 is quite fast and good for writing documentation. However, when encountering issues, it can repeatedly call the same problems as Sonnet 3.5. Sometimes, it may correct in the middle but continue to make mistakes. Moreover, it always executes commands first, even if rules are added to make it interpret first and then execute, it doesn’t take effect.
Alright, thank you for your explanation.
I can’t speak for anyone else…but, I’ve been having a BLAST with Grok Code. Sucker is a fiery bat out of hell FAST, seems to be super precise, its just shredding it up! It makes code changes instantly (GPT-5, Sonnet take a while.)
Thing that I’m absolutely LOVING: SHORT reasoning cycles! It “thinks” for 1-2 seconds at a time, and with a 480rps input rate and 2m tpm output rate, this sucker DOES NOT WASTE ANY TIME! THAT, I absolutely LOOOVE! 
So far, its been NAILING every request I give it, breaks nothing, does a perfect job following my rules (can’t say the same for GPT-5 or Sonnet…sonnet was my previous !) This thing does not over think, it clearly does not underthink, it has DEEP understanding and does not seem to skip a beat.
Now, I’ve built up some very stringent, strict, precise, and detailed rules over the last couple months. It may be, that all the effort that went into making those, which was largely to corral GPT-5 when it first arrived (and recently to a degree Sonnet, which seems to be ignoring rules a lot), has had a very beneficial impact on Grok Code.
Whatever it is…this thing is just blowing my mind! 
ABSOLUTELY KILLING IT!
1 Like
How did you use Sonic (small code, large code, refactoring, bug hunting, etc), and what’s your take so far?
BE HONEST
large code, very good, makes small changes as i want, does not create bazillion files or useless things. follows my instructions. good behavior in large codebase
I have tried it twice. Both times it start almost immediately complaining that their API rate limit has been reached and I should wait little bit before trying again.
Aug 30
Grok code fast 1 thinking, it’s absolutely terrible. Can’t code at all. All it wants to do is create demos and styling (that mostly don’t even launch), doesn’t do anything near what you ask.
@Rakly3 Could you post a full separate Bug Report with more info Create Bug Report
Also add a Request ID with privacy disabled so we can look into the details? Cursor – Getting a Request ID