After spending some time testing Cheetah, I’ll say it’s one of the fastest models I’ve worked with. It keeps pace better than most and produces stable, executable code with solid reasoning. I’m not sure who’s behind the architecture, but it’s clearly optimized for speed and practical use in coding workflows.
From my experience, Opus still feels like the most balanced overall, with GPT-5 following close behind. Cheetah has now joined that rotation as my third pick. It excels at technical debugging and quick analysis. I had a bug that ate up close to six hours of manual digging—Cheetah isolated and fixed it in under four minutes.
It’s strongest when working within established codebases—reading, diagnosing, and extending existing logic. Where it struggles is with open-ended system creation or tasks that require external web context. It doesn’t pull from web sources as effectively and seems tuned more for local reasoning than research-oriented generation.
Overall, Cheetah is efficient, lightweight, and reliable for engineering use. It’s not the model I’d use for greenfield builds yet, but for iterative debugging and refinement, it’s an impressive tool that’s earning its place in my daily workflow.
It seems that different users are given different versions of Cheetah, with varying levels of intelligence, but the only thing they all have in common is speed.
When I first used Cheetah, it was blazing fast but not very capable. The model seemed to stop working at one point, however a couple days ago it seemed to start working again. This second round, it seems much more capable than it was the first time around, and is still blazing fast. I think Claude Sonnet 4.5 still has the edge in a number of areas, but Cheetah in its current state, seems a very solid runner up and my second most preferred model now. I don’t know if there was an update there at some point, but the model is certainly better than it was on its initial stealthy release.
Yeah, I noticed that too — it does pause at certain points, but in my experience those breaks usually happen when the model’s waiting for direction on how we actually want to proceed. It’s almost like it’s prompting us to clarify the next step rather than continuing down an uncertain path or making silent mistakes.
When it stalls, it’s often because the executed prompt wasn’t clearly structured or lacked follow-through instructions. Once that’s ironed out, it picks up again smoothly. I’ve seen it get “stuck” a few times, but usually that’s tied to unclear continuation logic rather than the model itself failing.
It feels like they’ve definitely tuned it recently — the current iteration is way more adaptive and context-aware than it was on release. Still some gaps, but the responsiveness and technical precision have come a long way.
Yeah, agreed. From a cost-to-performance standpoint, Cheetah’s hitting a really good balance right now. The speed and consistency you get for the price is hard to beat. Claude 4.5 and 4 definitely have their strengths in reasoning and creative generation, but for hands-on engineering tasks, Cheetah’s output quality relative to what it costs feels like a better return overall.
It’s efficient, practical, and scales well across iterative dev tasks without burning through quota. For day-to-day development cycles, that price-to-performance ratio makes a big difference.
They both have their strengths and weaknesses depending on what you’re trying to do. I still think Anthropic holds the crown overall — Sonnet is my reliable go-to when it comes to new implementations or logic-heavy builds that I want done right the first time.
That said, I’ve been throwing Cheetah at those same types of problems just to see how it handles them, and it’s been keeping up surprisingly well. I have noticed, though, it seems to perform better with web-based projects than with mobile app workflows. The reasoning and output just feel more aligned with web architecture so far.
Yeah, same here — the speed is unreal. I haven’t hit a real bottleneck yet either, but I’ve started to notice some quirks once the codebase gets large or when I’m working on mobile builds. It doesn’t always maintain full context or continuity once the scope expands beyond a few layers of dependency.
For web projects though, it’s a complete game changer. The turnaround time is ridiculous — I can iterate, debug, and deploy way faster than before. It’s like the model was tuned specifically for rapid web workflows.
Still, even with the occasional hiccup on larger builds, the raw speed and consistency make it one of the most efficient models I’ve used so far.
Forsure, it’s only been about 72 hours since I started using it, so maybe I caught the updated version you mentioned. It definitely feels more refined than what others described early on.
Claude is still my main go-to — especially for critical logic or implementation details where I need things spot-on the first time — but Cheetah has definitely earned a spot in my top 3 rotation now. The speed and responsiveness are hard to ignore, especially when you’re in flow and just trying to move fast.
Yesterday I made it work on multipage WordPress live project (SSH Access and wp cli installed) and this is so unreal - I got it to reorder Categories for me, list outdated articles, update the Categories, edit theme templates and improve existing articles just by using wp cli command and browser option in Cursor.
I had to review the changes of course and tell it to adjust some details but overall it hit the task on point - a task if done manually would have took me days to complete - done in 2 hours.
Its very odd. Yesterday Cheetah was awesome. This morning, its like it has an IQ of 50… The thing is an unmitigated disaster zone. This is something I really cannot account for, such wild and dramatic swings in capability. Thing was AWESOME yesterday. Today, its nuking my local DB without asking, won’t apply a single rule file, doesn’t understand how to run commands it was running perfectly yesterday (notably, it keeps putting command parameters in the wrong order), etc.
This is what forced me to ditch the model the last time. It just went crazy…or nuts…or just retarded… If anything, this model needs consistency. It should behave the same yesterday, today, and tomorrow…
Something is very wrong with Cheetah right now. It is having TREMENDOUS difficulty using, finding and following rules right now. This may be related to more extensive issues with rules in general, as it no longer seems as though Cursor is actually applying the full hierarchy of .cursor/* directories. I don’t know if this is Cheetah specific, its really tough to tell.
However, today, Cheetah seems completely incapable of finding or following rules. Even when I explicitly reference them, it completely ignores very explicit, clearly stated as MANDATORY, and clearly delineated WRONG vs. CORRECT ways of doing things. So it has been struggling massively, trying to do something very mundane: Testing a REST API using HTTPie. This was fine yesterday, today the darn thing cannot even put the command parameters in the right order, and it is failing to run the command 9/10 times.
I’m really quite baffled at how great this model seemed yesterday, and how utterly retarded it seems today. What in the world has happened?
Haha yeah, feels like the devs at Cursor heard us hyping it up and decided to dial it back a notch
But seriously, I noticed that too. The recent behavior lines up with what I was seeing when testing mobile builds — it just couldn’t get anything right on that side. Structuring commands, following the rules, even basic path logic — all over the place. Web projects, on the other hand, still seem mostly fine. It hits a few bumps, but it powers through and usually finishes correctly.
The speed has definitely toned down though — closer to Claude 4.5 levels now. It’s still solid, just not that crazy turbo mode it had earlier in the week.
Its really quite baffling. I had this theory about regions of an LLM. I don’t know how big a model Cheetah is, but I wonder if there is something like that…where today, the model is operating through the retarded region of its model, whereas yesterday it was operating through the genius region?
I honestly don’t know how to account for such a radical shift in a single model’s behavior, just over night…
