Ricky Doar

@threepointone Would love to hear your feedback on it! 🫡

my love

@SucciMas @mitsuhiko We are very proud of our editor, but our goal is to help builders ship incredible software, and the editor is only a part of that story (and needs to evolve drastically as AI evolves). We have a lot of other things we're proud of that is not our editor 😉

@mitsuhiko Everyone asking this is the best signal that their moat isn't the editor 😂

Does anyone know if there is a way to use composer 2 via API on a non cursor harness?

@JakeHaines_ @mark_k You are able to do this using the ACP client! x.com/leerob/status/…

@mark_k I would like to pay for Composer 2 access, but use it through OpenCode.

Are you going to give Cursor another chance, now that Composer 2 is here?

@tejashaveridev @leerob @jonathan_wilke I like what you're doing here 😉

@leerob @jonathan_wilke 👀 its clear whats coming

time to go back to cursor cursor.com/blog/composer-2

@LiorGal785825 @cursor_ai I hear you. It's how we judge models internally, and we find it more representative of reflecting real world coding problems. But we understand that there is an inherent trust issue to benchmarks in general.

@rdoar4 @cursor_ai Yes i can see trminal and SWE but comparing other models with CursorBench still doesnt fill right no one can actually validate it.

Composer 2 is now available in Cursor.

@abijahkaj @cursor_ai Cursorbench is how we judge models internally and we've found it more accurately reflects models in real world use cases, but you can take a look at Terminal Bench too! x.com/cursor_ai/stat…

@cursor_ai nice but "cursorbench score" sounds like a "trust me bro" score

@LiorGal785825 @cursor_ai It's there! x.com/cursor_ai/stat…

@cursor_ai Looks promising. Can you share the industry-standard benchmarks, such as SWE-bench Verified, Terminal-Bench, OSWorld, etc.?

@KYDNO_ @cursor_ai It's a subtweet, but the Terminal Bench scores are there too! Take another look 👀

@cursor_ai please release other benchmarks exept for your own please

Should we drop cursor x jensen merch? @nvidia

@dillon_mulroy Why outside of Cursor?

i think i’m back to wanting a really good tab model - any progress here outside of cursor (i don’t have access to supermaven) and for nvim?

@AlexGonchX @cursor_ai There is absolutely nothing better than a clean context window for new tasks, but there are still cases where you are working on the same task across multiple steps, and smarter compaction makes a HUGE difference.

@cursor_ai nice but cmd+N still works better than any amount of RL training on this

We trained Composer to self-summarize through RL instead of a prompt. This reduces the error from compaction by 50% and allows Composer to succeed on challenging coding tasks requiring hundreds of actions.

Matt Maher tested frontier models in Cursor v. other harnesses. Cursor boosted model performance by 11% on average: Gemini: 52% → 57% GPT-5.4: 82% → 88% Opus: 77% → 93% His benchmark measures how well models implement a 100-feature PRD. @cursor_ai consistently outperformed.

@jxnlco cc @jediahkatz

Who do we work with to make Cursor as 5.4 optimized as opus.

@tombuilds @edwinarbus @cursor_ai Building a great agent harness goes well beyond a set of system prompts

@edwinarbus @cursor_ai Surely the labs would just copy the prompts?

@cramforce genuine passion >>> perfection 🫡

In hindsight I should have STFU. Still in it for the love of the game 😘

@milichab @SpaceX @xai @JasonBud One of the most impressive builders I've had the pleasure to watch work. Congrats Andrew!

I’m joining @SpaceX and @xai with @JasonBud. X is the company realizing science fiction - reusable rockets, humanoid robots, data centers in space, and more. Almost 10 years ago, I joined SpaceX as an intern on Dragon 2 crew displays. This was in the era of the first rocket landings on barges, long before the Dragon 2 restored human spaceflight to America or Starlink delivered internet from space. Every day since then, I’ve thought about the next steps to land on the Moon - and to build a city on Mars, data centers in space, the brains behind robots, and beyond. There is no better place to build teams and products from the ground up with planetary scale resources. If you’re looking to work on the hardest problems that lay a foundation for humanity’s future to the Moon, Mars, and beyond - DM me.

@aakashgupta Fair points but that’s not how you calculate NDR. It’s inclusive of churn. If you exclude it you’re cooking the books

$100M ARR in 8 months. $200M in 12 months. $300M in 14 months. $400M in 15 months. On the surface, this is the most absurd revenue ramp in European software history. But there's a number Lovable has never disclosed, and it tells you more than the ARR figure ever could: churn. Sifted asked directly. Lovable declined to share churn rates or the split between monthly and annual subscribers. When a company growing this fast won't tell you how many customers leave, you're not looking at a revenue number. You're looking at a gross bookings number that gets re-annualized every month. Bolt's CEO said it publicly: "The churn rate for everyone is really high. You have to build a retentive business." He was talking about the entire vibe coding category. Then Bolt immediately changed its subscription model to try to keep people from leaving. That tells you everything about the structural retention problem in this space. Barclays flagged it too. Traffic to Lovable dropped 40% from peak as of September. Vercel's v0 dropped 64%. Bolt dropped 27%. The analysts wrote that these companies could have "questionable economics" because the revenue comes from month-to-month subscribers who won't stick around. Lovable's response? Osika says net dollar retention is above 100%. But net dollar retention only measures customers who stay. If 50% of your customers churn and the remaining 50% spend 2x more, your NDR looks incredible while your business is a revolving door. The unit economics are even spicier. Lovable pays Anthropic and OpenAI per inference call. Every app a user builds costs Lovable real money. A source told Sifted margins might have actually gotten worse after switching to agentic mode. At 45 employees generating $400M ARR, the revenue-per-employee ratio looks legendary until you realize most of that revenue flows straight to model providers. Meanwhile, Cursor just hit $2B ARR with 60% coming from enterprise contracts. That's the difference. Enterprise locks in annual commitments. Vibe coding's user base is overwhelmingly individual creators who build one app, ship it, and cancel. Lovable is real. The product works. The growth is genuinely unprecedented. But a $6.6B valuation on $400M ARR requires that revenue to stick. And the entire vibe coding sector has a structural retention problem that no one has solved yet. The fastest company to reach $400M ARR could also be the fastest to find out what happens when the denominator in your LTV/CAC ratio collapses.

@naderman Now try Bugbot! :)

Review 2 PRs for $37 this morning. Took 9 minutes on a 8 line template change, removing some unnecessary info, got a tiny nitpick out of it. Then 30 mins on a 15 file (+331/-153) PR where it missed the actual bug introduced, which manual review found quickly. Turning off for now.

@GradyGaugler @cursor_ai @leerob Yes, included in your subscriptions!

@cursor_ai @leerob Dang, Cursor has it all! Are automations included in the subscription? Really pushing me to Ultra…

We're introducing Cursor Automations to build always-on agents.