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 consistently outperformed.

@cursor_ai wow. opus actually shined there! these are very interesting results. i'd love to use cursor, its DX is really really good but currently with claude code and codex inference pricing discounts it's just seems too expensive in comparison.

@cursor_ai DM me

@cursor_ai dm'd :)

Full vid:

@cursor_ai I have noticed cursor with opus seems to give me better results than using claude code directly!

@cursor_ai Harness matters more than people want to admit

@cursor_ai damn are you guys doing better than anthropic directly RLing with the cc harness is it the same cursor harness for diff models?

@cursor_ai Cursor quietly becoming the benchmark itself.

@cursor_ai

@cursor_ai Pretty impressive.

@cursor_ai lol opus has always worked better in cursor than claude code for me. codex on the other hand is already amazing

@soleio @cursor_ai I believe this

@cursor_ai Cursor really is a great harness. too bad you can't use a provider subscription with it API pricing points do not justify the benefits

@cursor_ai The Opus jump from 77% to 93% is wild. That delta is bigger than most model upgrades give you. Feels like harness quality is becoming its own axis of improvement, maybe as important as model capability itself.

@cursor_ai What a brilliant result! I’ve been using Cursor for almost a year now, I had a Max plan before, but was tempted to switch to Claude Code because of the more affordable price.. I have high hopes for Cursor’s future

@cursor_ai I've noticed similar boosts in performance when using Cursor with my finance app. When I optimized my model for a specific use case, I saw a 7% reduction in errors, which translated to a 5% increase in user retention. Those small wins add up quickly.

@cursor_ai Ok, I stand corrected, Cursor seems to be a strong harness for the frontiers, but the problem still stands on pricing. Rn, it makes more sense for me to remain a max subscriber on claude and codex simultaneouslly. Still hitting limits on both this week. 🫠

@cursor_ai Midi is not talked about enough in how inanely fast, good, and cheap it is in cursors harness

@cursor_ai the boost is context. cursor can see your entire codebase, your project structure, your dependencies. it's not that the models got smarter. it's that they can finally see what they're shipping into.

@cursor_ai Native Violation... But this has been my feeling as well as using the tools. I think Claude and Codex just have a lot of marketing going through right now.

@cursor_ai opus going from 77 to 93 in cursor means the harness is doing more work than the model itself. the wrapper became the product

@cursor_ai Love this, actually explains why working on cursor always feels much smarter than Codex or even Antigravity, but yeah, using Claude on a $20 plan is just not sustainable

I used to use Cursor and was very happy with it. It did a significantly better job compared to other CLI tools and IDEs I tried. Every new update made it better. I had even started porting tools made for other CLIs to Cursor because Cursor did a better job at directing the models than other IDEs. Once Cursor moved away from the request model, it was financially inefficient to use Cursor anymore. Composer 1.5 had more leeway in terms of usage but it was unable to build as effectively as Opus. Now, I use Claude Code with the $200 plan and I don't have to worry about usage anymore. If Cursor can improve the usage limits on external models, I'd switch back to Cursor today.

@cursor_ai Yeah I believe this. If frontier labs were not selling at lost that would be a big deal.

@cursor_ai I am the co-founder of an AI startup. Our team of 8 was spending $10,000 a month on Cursor and now we are spending $1,000 per month on Claude code with the same usage.

@cursor_ai context: gpt-5.4 xhigh in codex cli got 95% making those gains less important

@cursor_ai opus jumping to 93% in cursor is insane. wonder how much of that is the system prompt vs the tab context

@cursor_ai I was using $2500/mo with cursor with Sonnet 4.5/4.6, switched to vs code with claude code extension, have stayed within my max plan no problem. Not acceptable for 11% increase in performance.

@cursor_ai opus at 93% is wild. the harness matters more than people admit

@cursor_ai The Opus 93% number is the real headline. Means the harness matters more than the model for complex PRD tasks. At some point the gap between raw capability and harness-engineered output becomes the actual product moat. We're already there.

@cursor_ai This is awesome! Something that other IDEs should copy.

@cursor_ai He's one of the best youtubers

Can you please try out @FactoryAI / @droid and let me know what results you get? I'd be interested to see how you go because it's been the best harness that has an observable marked improvement across all models I've used when compared with any other harness (including Cursor at the time) hands down.

@cursor_ai is it the same cursor harness in the cli? like exact same?

@cursor_ai super cool!!

used cursor about a year ago when i was 15, the experience was genuinely great. but that was before claude code and codex dropped, once those came out with way better pricing, hard to justify paying for cursor on top of model costs. would honestly go back because of the DX in cursor, but paying for cursor on top of everything else is just too much rn

@cursor_ai prompt engineering to IPO run

@cursor_ai Cursor harness is great, especially for Gemini. I’ve found that Gemini just spirals and loops forever a lot in other harnesses. Only thing is Claude code is just way cheaper for me