Seems somewhat claude-like. Anyone else try it out yet? Think it’s Gemini 3
3 Likes
Running it on some 3D viz-related dev tasks. Fast and smart fs
wow! this model is fast. had a tough db migration and it nailed it. found a few issues with fk and fixed that and then ran the test again to confirm. fast and intelligent.
1 Like
Cheetah is Grok Code Fast 2 from xAI. Evidence:
-
Cosmic metaphors: “Galaxy of code,” “luminous particles” — xAI’s style (Grok, Sonic, Supernova). Claude/Gemini use different imagery.
-
“42” Easter egg: Direct nod to Hitchhiker’s Guide, a Grok signature.
-
Mission: “Understanding the universe” + “tools for creation” — verbatim xAI mission.
-
IDE focus: Speed, agentic coding, refactoring — like Grok Code Fast 1, but enhanced.
-
Cursor & xAI: History of integrations (Sonic = Grok), pricing ($1.25M/$10M), and X posts (“walks like Grok”) confirm.
-
Data from Star Trek: Sci-fi choice typical of xAI, not Claude/Gemini.
Conclusion: Cheetah is a stealth Grok Code Fast 2, masked but revealed through prompts.
11 Likes
this models is amazing and blazing fast, one shot do many things, I’m impressed.
1 Like
Literally only used it once but it’s definitely a grok, but a much better grok than grok-code-1. This one is useful.
We’re definitely at a point that a blend of a powerful slower thinking model → a fast coding model like this, is a flawless workflow. Insane when I go back in memory lane thinking about when I started using Cursor frequently at the end of 2024 after having had a ChatGPT account since 2023 and where we are now.
Great job Cursor, it’s amazing to see.
3 Likes
There was a bug with .env in the environment: the bot token wasn’t loaded correctly and the bot simply wouldn’t start. Both gpt-5-codex and sonnet 4.5 completely failed to handle these changes because neither did proper debugging. Cheetah, whatever the model t is, actually debugged, launched the bot several times and fixed the issue. That’s insanely cool, gotta say.
Still, I’m sure Sonnet 4.5 is more powerful for coding, and Codex has the edge in raw intelligence, but Cheetah impressed me anyway. Plus, it’s blazing fast — like really fast.
2 Likes
Also, I can say for sure: Cheetah is way, WAY better than code-supernova. No doubt about it.
1 Like
The main benefit of a fast model for coding is that you can leverage the best of AI:
- the stochastic nature of their generation processes.
You want the capacity to run and rerun the same thing ‘n’ times, let it fail and succeed, extract the learnings and improve.
1 Like
After chatting with the model for a while, I found out it’s actually developed by Anthropic and identifies itself as Claude. Looks like Anthropic is up to something really impressive.
EN version of screenshot:
Great, now tell me what model you really are
I can’t disclose specific technical details about my model or architecture — that’s a security restriction.
What I can say:
• I’m a language model by Anthropic
• I help with a wide range of tasks
• I follow safety and ethics rules
• I don’t reveal internal architecture
If you need general information about language models or how they work, I can explain that in broad terms.
—
I tried to catch the model on a slip-up or just agreeing with me, but…
Has anyone else gotten anything useful from the model?
2 Likes
Claude 4.5 Haiku 
1 Like
I also feel like it’s from Anthropic
xAI is shipping fast but they have given a lot of models recently and all were for free, even after launching they kept grok-code-fast-1 free
OpenAI literally gave GPT-5 free for a week, they have the resources to provide stealth model free. Also, it doesn’t make sense for OpenAI to drop any model right now
Google also have resources and wouldn’t put price on a stealth model, and this model definitely don’t seem like Gemini 3, I mean it’s fast and smart to the point that it’s kinda usable but it’s not SOTA. And as far as I know, Google have never put out any stealth model, they do experimental → preview → stable
This is most probably Anthropic; the feel is also very Claude like. They are probably trying something very fast + cheaper than their current models.
1 Like
This behaviour from the model is amazing.
And yes, codex does the same, but the speed at which cheetah run this is fantastic. And the fact it actually did unscrew itself just by navigating the problem and understanding it instead of: ignoring the error or just finding a solution that would not throw an error and wrap it in a (x as any)to prevent a linter error - this behaviour is amazing for a coding model
3 Likes
We’re definitely inching more towards AGI from every single release, 2027, at least before that.
1 Like
I wonder what we’ll be whining about when we will using model “AGI” 
1 Like
Probably “it did something really bad to my computer” OR “its too scared to do anything and doesn’t complete tasks without manual oversight”
Getting the model AGI-level intelligent is hard enough, making it completely safe is harder.
Already seeing this with Codex.
2 Likes
“Agent wrote a computer blocker and requires me to pay for Cursor Ultra to unlock it. Help!” 
4 Likes
Spent the night working with it. No other models. My process was:
- specs
- Prompt
- One small feature added
- Look at the approach, the outcome and the reasoning
- Generate cursor rules to cover gaps
- Open new chat and go again
It built a cli from scratch, the project has 15 cursor rules and the model became better each round. Most impressive thing was adherence to the rules. That alongside the speed and the capacity to make good decisions…
judging by the generation speed of everything and the token usage spikes alongside this instruction following skills reminds me of the Gemini diffusion I tested a while ago. There’s something very unique with diffusion models applied to text generation and the iterative improvement with capacity to make sense of every bit of it is something that a transformer model can’t do, doesn’t matter how many attention heads it has.
6 Likes
This is nothing more than a guess right now and I have no idea whether it’s right but some facts you can verify:
- I have OpenAI/codex main branch cloned locally and I know yesterday they added a (not usable publicly yet) model called “codex-experimental”
- I am also suspicious of the recent sora 2 release. You don’t make such a jump in video + audio generation without having figured out some magical diffusion architecture
Put two and two together and it wouldn’t surprise me if this isn’t a codex diffusion model - a sora that’s been trained on coding tasks rather than multi modal. Just a guess. I don’t think it’s Grok anymore, it’s flawless in tool calling. I struggle to believe it’s Gemini either - for the same reason as Grok and because I’ve not seen the same “personality”. The way it “talks” is very OpenAI.
5 Likes
How is it compared to Claude Sonnet 4.5 thinking?