The model is (like Composer 2) based on Kimi K2.5 and they claim SOTA performance for 1/10th of the cost. The tweet also mentions that they've started a new model from scratch on Colossus 2 (xAI/SpaceX Cluster). Really impressive how they've made this jump from being called the vscode fork with no moat just a couple of months ago.
I am not sure we should dismiss what they have today. Nobody has yet to come close with a full package ide that works well for coding. Is that not a moat? It is easy for my to in my head discount it, thinking that I could build something myself but between autocomplete and their workflow for agent use, it feels like they have some tangible moat emerging.
If we ignore cost (which is kinda hard to ignore), I feel Codex kinda' does it for me. Sure it's not really an editor but I find I don't need that _that much_ and it's easy to launch an external editor (they actually have the feature).
The ironic thing is that half a year ago, after trying factory.ai I thought chat-first interface was a stupid idea that will never work.
A few years ago I tried Zed when it was still pretty early, but eventually settled on Cursor. I gave Zed another shot a few days ago because Cursor’s worktree support still feels pretty weak.
In my setup I use multiple agents like Claude Code and Codex, and Zed’s ACP support makes it pretty nice to manage them all as “threads” in one place. Worktree switching also feels much smoother.
Overall the experience was pretty good, but the way the agent and editor are integrated still feels a bit lacking, and tab completion is the big one for me. Cursor’s tab completion is still the best I’ve used.
So now I’m using both. For work that needs a lot of focus and careful iteration, I use Cursor. For things that are easy to split into worktrees and hand off to agents, I use Zed with Claude/Codex.
The tab completion is "faster than vim" from a long-time vimmer. It's at the point where a lot of times i'll lead with the comment instead of the code:
# now take the list and sort by x.lastName
<tab>
...and it'll "do the thing" (w/ type hints, its own comments, etc). Obviously in this very simple, understandable, completely contrived example, it's "trivial" (but 3 years ago would have seemed like magic), but it'll also pick up on "continuation / more of the same" type edits. A comment like `# use random_utility to call the api and only accept matches which supplement addresses that have already been found` will (usually) autocomplete all the gobbledy-gook w.r.t. tokens, URL's, function names, etc. so it's effectively an "automatic omni-complete with simplistic post-processing"
Example #2: I was just fixing some vibe-coded slop, where it was taking `click.echo( some_api.whatever_endpoint() )` and the "slop" portion was literally emitting: `str('{ "A": 1, "B": 2 }')` and that function call was emitting it directly.
On the command line, I was doing `blah whatever-endpoint --something | jq '.'` and got tired of the JQ thing, so I'm like: "I'll just use `json.dumps(...,indent=2)`", but lo and behold, I'm getting a dumb JSON string literal, not a pretty printed object shape.
I start typing `json.loads(` to move from "str()" to "dict()" ... and it autocompletes the whole scenario (on that line), then I move to `def some_other_endpoint` and it basically has that same edit queued up. (ie: it "knows" what i'm about to do).
...so overall, "faster than vim", even with high skill bar for repetition, motion, macros, sed-style edits, etc. You can't beat: "<tab>", especially when it's lightly intelligent (ie: knows when/what/str/int, adapts do different function calls, etc).
I've been using cursor for over a year for my personal projects. At work, I use Claude Code, and so I've been wondering if I'm missing something in the other agents.
Over the last week, I tried out two other agents on my personal projects: dirac and forgecode, after seeing impressive results from both of them on terminal bench.
After a good amount of testing, and over $100 in open router spend, I'm back to cursor.
I really liked forgecode the best, and it feels better than claude code, but cursor definitely feels best to me. Composer 2.5 is fast and effective, and it makes a huge difference. I was running `forge` with Opus, and it was taking dozens of minutes to do things, and the feedback loop was so slow.
The previous version of composer was also much faster, and it makes a difference. Maybe people like context switching, but I prefer to stay focussed on the task in front of me, and I'm reviewing the code carefully.
I think that's a pretty good moat. I was ready to end my subscription a week ago, and now I'm back after learning the grass is not necessarily greener on the other side of the fence.
They don't have the same quality and kind of data. For example, Claude Code might have general conversation flow data for implementing feature X, but Cursor has users individual editing actions AND the chat flow. Which line did the user manually edit after the agent did it's thing? What's the commit message (if done manually)? Stuff like that is worth it's weight in gold.
Honestly the data itself is probably worth heaps even in the company itself collapses. Early attention engineering when humans were still in the loop!!!
> Early attention engineering when humans were still in the loop
Exactly. Cursor was the first product used by tons of devs on real codebases. Just the signal "acceptance rate" is huge and can't be easily captured w/ synthetic data.
How much the RL they are doing really improves Kimi K2.5 is to be seen. So, right now, the ground truth is that they combined what they had with a strong open weights model. The RL improvement may be both marginal (since may folks report strong results with vanilla K2.6) and may mostly bias the model towards coding tasks: when a model like this is trained to be generalist, there is a tension between being good at one thing and the other, in terms of SFT and RL. You can see this in the DeepSeek v4 Flash training report for instance but it is a known fact. So if you have the GPUs and a decent RL pipeline that does not run the model you can indeed specialize it a bit more for a given task at the expenses of tasks people will not do inside Cursor. But, so far, the measurable reality is that Cursor uses an open weight model like most could do, and the RL story could be partilly a marketing move to call to Composer 2.5 more than a real strong gain, given that there is no way to verify and K2.5 was already strong. And we also know that they had to partner to do the training, which is also not a good news.
They are still a vscode fork with no moat? Like they lost about 70% of users in half a year which goes to show how there is not even the tiniest of moat.
I think it's going to be brutal for them to compete with OpenAI and Anthropic.
I switched to claude code because of usage. For $200 a month, I would run out of usage halfway through the month. Then be forced to use their composer model or whatever slow, dumb model they served up in their "auto" mode.
For that same $200 a month, I could use claude code and basically never hit usage limits.
I don't understand what people are doing who run into the limits on that max x20 plan. I NEVER have.
Since the frontier is only 8-month ahead of DeepSeek, it is hard to see how model training can be a moat as all the tricks are available from open labs in China. You really just need <100m to bootstrap at this point.
>& now they're still losing all of their users to Claude Code and Codex.
Why pay for Cursor when I can use GLM 5.1, Kimi K2.6, MiniMax M2.7, Xiaomi MiMo V2.5 Pro and Deepseek v4 for cheap and use whatever harness I want, including Claude Code.
It's not like Cursor harness is the best out there.
And even if I want to edit the code, I don't need to run the agent harness in an IDE.
these are in the trillion parameters range, not sure it's actually that cheap to have at a reasonable speed without quality degradation & without like.. your own DGX B200
It's still a VsCode fork just now with a Kimi fine tune and still no moat...
I won't debate that it turns out none of this mattered when it came to being as successful company though and kinda makes anyone who tried to roll their own instead of fork look a little silly.
How I see this is that its so important to bundle the model with the right tooling.
Like a racecar, having the best engine doesn't help if the rest of the car lacks other winning properties (reliability, aerodynics etc).
So for Cursor, which IMO, they put themself in a strong position by having both a solid IDE __and__ a solid+cost efficient model. Those two working great in combination for the task they are designed to solve (coding) is more important than benchmarks
