@nrehiew_ says Opus 4.7 and GPT 5.5 effective costs converge because developers reject more GPT 5.5 code

the economics of intelligence

> Cost per agent request varies by nearly 9x across model families, showing that the same workflow can have very different cost profiles depending on the model behind it.

some higher cost models are cheaper in the long run due to increased intelligence, but for p50 requests a model like composer 2.5 will do the job both faster and cheaper

a lot of interesting data in this report, recommend a read

Despite GPT 5.5 having a higher output price than Opus 4.7, Cursor found that Opus4.7 Agent requests are almost double the cost of GPT 5.5's.

But, normalized for % of lines accepted, both models end up in similar cost territory, which indicates more lines from 5.5 are rejected

@ericzakariasson the 10X engineer is real

@nrehiew_ they should control for task difficulty though, even a vibe-based analysis would suffice

@stalmico not necessarily. highest cost models are usually more intelligent, meaning that they more often make the correct change on the first try. if they don't, it'll be both slower and more expensive

@ericzakariasson Can you guys improve support for cursor windows app?

@ericzakariasson so the expensive models finish tasks faster?

@ericzakariasson this is insane. i dont see opus 8x better than composer 2.5. The composer price is a big deal!

@Lay_Bacc literally

@ericzakariasson That 9x cost variation is a good reminder. We've seen similar stuff internally; the 'cheaper' model often needs more human help or re-runs, which quickly eats up any initial savings.

@nrehiew_ "output more, accept less" territory is wild

5.5 really out here costing the same for way more words nobody keeps

@nrehiew_ 2 different strategies, 2 similar results

Interesting. Thanks for the analysis

@ericzakariasson I can’t scroll with swipe feature when using RDP using windows app - swipe to scroll works for every other app except cursor

When initiating new chat, it shows a weird pop up saying - open “session” (or something like that) with - and gives list of apps like notepad cursor etc.

@ericzakariasson the 9x variance is real but i think it grows further in agentic loops specifically. a model that fails 20% more doesnt cost 20% more in multi-step workflows. failures cascade into retries, context refills, and downstream rework. the cheap model premium compounds nonlinearly

@nrehiew_ this says more about the average claude enjoyer than on the models

@ericzakariasson using this type of flow at work - we default to lighter models now and only bump up when the task actually needs it

turns out most requests don't need the nuclear option lol and sometimes even auto option in cursor works well

@darkstalwart yes! how can we improve it?

@ericzakariasson Isnt it why they sometimes have a quick/fast version tho

@nrehiew_ So the real metric isn’t raw token cost, but *accepted* lines per dollar — and on that, 5.5’s higher rejection rate pulls it level with Opus 4.7

@ericzakariasson Efficient agents are not just about cheaper models. They are about smarter memory decisions ✅