Cheetah is good, but how long until they announce who it’s from?
A new model will come out and then the hype will go away. already has. Good model. Hype has gone down with supernova, cheetah is better so not really using it anymore.
2 Likes
Best new model I’ve used. Good work whomever…
3 Likes
Cheetah is very mono task - you need to prompt just ONE thing. But it seems to rip it then, very interesting.
1 Like
Having a hard time putting my finger on this one. I kind of suspect Grok Code, because it is so blazing fast, and I am not sure who else currently can field that kind of broad scale speed, except Musk with his Collossus data center. However, it gives off Claude vibes, specifically Claude Sonnet 4.5, with its light weight usage of verbal feedback. The COST, at $1.25/mtok input and $10/mtok output, also seems much more Anthropic level, but, I honestly have a hard time believing they have the money to make something like this so fast…unless its just a bait and switch, where speed tanks once it gets real-world usage upon full release…
So, Glauk Code? Some kind of chimeric blending of the two beasts here…
How does this seem like Anthropic level when GPT-5 is literally $1.25/mtok input and $10/mtok output lol. But seriously, this cheetah model is something else. I love how it’s just blunt, to the point, no BS. It’s a refreshing change.
1 Like
Just tested Cheetah. Yes, it’s fast but it makes too many mistakes for me. When using GPT5, I almost never have to do follow up requests to fix stupid errors. Cheetah generates code that doesn’t build - missing methods, pass wrong parameters, don’t complete my initial requests. So many follow ups are required to make things work. So, it’s fast to complete a requests, but the number of requests required to complete my tasks is times more than with GPT5. So I’m not sure the overall performance gets higher with Cheetah. Just my first observations.
1 Like
Whatever the model is, it is incredibly fast. That alone is reason to make it part of my workflow. I am tired of sitting around for models to spend many minutes thinking. The only thing that keeps me from always using gpt-5-fast is how expensive it is.
1 Like
This model is definitely not Claude. It seems very much like Grok Code. Reason I don’t think its Claude, is the darn thing DOES NOT follow any rules. That was something that Grok Code Fast 1 started doing, that ended up shifting me back to Sonnet 4.5. Sonnet does seem to just have a better, deeper integration with Cursor.
GPT-5, Grok, these models don’t seem to follow the rules nearly as consistently, and sometimes just blatantly ignore them despite being explicitly referenced. This Cheetah model, feels very much like Grok Code, between not following the rules well, its speed, and its direct, fix the code/write the code approach.
1 Like
Well, I am not sure what model this is, but my honest assessment after a few days of using it here, is that it is not particularly effective. I’ve been working on resolving a naming issue, that ended up being quite extensive. This model has NOT been handling this issue well. It keeps flip flopping, between fixing things in one pass, but the pass is incomplete…types are not updated while code that uses the types is, stuff like that. When I then tell it to fix the outstanding code errors, it REVERTS its prior changes to the previous names, rather than fixing the types. When I tell it to fix all the type errors, it will fix a few, but also revert some of its name changes…
So far, its not been a particularly pleasant model to use. It seems to be “incomplete” a lot of the time, not fully evaluating all that it needs to do and fully completing ALL of the required changes. Its a bit too “short” and a bit TOO “direct”. With the original Grok Code (not saying this is Grok Code 1.x) it just did the work, and then summarized. That was kind of nice, although a bit of feedback like Sonnet 4.5 gives would have been nice. However, this model’s feedback is…well, not exactly useful?
Dunno. Not sure who it is…feels a lot like Grok Code but today, its speed, has been quite slow, the blazing fast nature it had when I first started using it seems to have vanished. Its code-generation rate seems like Claude Sonnet 4.5, but I don’t really believe that’s what it is…it just doesn’t have Sonnet’s mannerisms, summaries are quite different (no checkmark and red x or rocket emojis that Sonnet makes wonderful use of.)
In any case…I’m moving back to Sonnet 4.5. It was interesting trying to figure out which model this is, but, its just not a particularly pleasant model to use, and it takes a lot more work, to get this thing to the finish line, as it is frequently incomplete in its efforts, and requires lots of additional prompts to get it to fully complete a whole task…and even then, it doesn’t quite seem to have the understanding it needs, to do things properly (i.e. fix the type names and the code usage of those type names all at once, don’t flip flop things around, don’t revert valid changes to correct a syntax error driven by an incorrect type, etc. etc.)
1 Like
Did something happen with this model? I am now getting errors stating I don’t have a proper key to use cursor or the agent…
1 Like
The same here
That is an account issue many of us are having today. See `Unknown plan` issue means Cursor does not work - #31 by RafeSacks
There is a high probability that cheetah is a Cursor model. Since it’s so freaking fast, and shows promise in quality, I’m first time excited for Cursor since a looooong time (last time was when gpt5-fast-high was able to debug complex stuff suddenly).
The model seems to run a very thin rail, very mono task. I guess with some thinking (like grok code fast amount of reasoning), it could be really good as a code implementation slave (that needs some planning). It seems you have the data for it, looking forward @condor
Edit: It reasons already.
2 Likes
Completely disagree. You have to know how to steer it.
It’s my default for most tasks.
Perhaps. However, I stand by the experience difference. I did not enjoy using cheetah really. That was the main turnoff, it was just unpleasant to use.
Additionally, I finally switched to sonnet 4.5, and it handled the problem with ease, and there weren’t any “steerage” issues using it for the same task. Perhaps that IS my problem, but, why is it a problem? I guess, IMO, there shouldn’t be so much friction using a model to effect daily tasks. 
1 Like
That’s a fair comment. I feel the same when dealing with Gemini models. The amount of times I’ve wanted it to have a face so I have a punchable target… I would not punch a cheetah though, tread carefully.
1 Like
Hah! Yeah, I stopped using Gemini entirely for coding. It seems amazing when it comes to visual stuff, video, images, even audio processing. We use gemini-2.5 a lot for that stuff in our product. But when it comes to code, its a bulldozer right out of nitemares.
Grok Code, even Cheetah, and GPT-5 are all better than Gemini for coding, IMO.
1 Like
Let’s collect how to steer it!
- ony one task
- no decisions(?)
Edit: The more I use it, the more small model smell - it’s quite dumb (or shall we say specialized?), actually. Feels a bit like small sister of gpt5 auto on world knowledge. Could nevertheless be useful.
Cheetah got nerf? its look so slow now
Honestly - I’m impressed (takes a lot) - it just oneshotted an feature that Claude 4.5 has been thrashing around with for 2 days now for a total of $2.87.
Assuming this doesnt get nerfed and assuming it performs well this looks to be great.
This was a fairly self-contained task, so I have not tested this on deep codebase issues and tasks yet but so far I am happy!
If this does perform well on wider code issues - this would bring me back to Cursor, right now after my previous $161 bill for a single conversation on Claude 4.5 (admittedly they refunded about half due to the recent billing / caching charging error).
Only complaint is that the context window is far too small to keep longer conversations going without a constant summarisation, once it summarises it loses a lot of context, if it cant basically one-shot something then it very quickly loses its effectiveness - I dont think this model will be good for longer context requirements
2 Likes