Comment by adamoshadjivas
21 hours ago
Agreed on everything. Just to add, not only anthropic is offering CC at like a 500% loss, they restricted sonnet/opus 4 access to windsurf, and jacked up their enterprise deal to Cursor. The increase in price was so big that it forced cursor to make that disastrous downgrade to their plans.
I think only way Cursor and other UX wrappers still win is if on device models or at least open source models catch up in the next 2 years. Then i can see a big push for UX if models are truly a commodity. But as long as claude is much better then yes they hold all the cards. (And don't have a bigger company to have a civil war with like openai)
Not sure this is true. Inference margins are substantial and if you look at your claude code usage it's very clever at caching
as an example here's my usage. Massive daily usage for the past two months.
> CC at like a 500% loss
Do you have a citation for this?
It might be at a loss, but I don’t think it is that extravagant.
The way I am doing the math with my Max subscription and assuming DeepSeek API prices, it is still x5 times cheaper. So either DeepSeek is losing money (unlikely) or Anthropic is losing lots of money (more likely). Grok kinda confirms my suspicions. Assuming DeepSeek prices, I've probably spent north of $100 of Grok compute. I didn't pay Grok or Twitter a single cent. $100 is a lot of loss for a single user.
Comparison should be with Claude API pricing. It doesn't matter what other models cost.
1 reply →
what?? sonnet/opus is way better than deepseek, how can you compare that to deepseek
also you probably talking about distilled deepseek model
4 replies →
I'm also curious about this. Claude Code feels very expensive to me, but at the same time I don't have much perspective (nothing to compare it to, really, other than Codex or other agent editors I guess. And CC is way better so likely worth the extra money anyway)
I think GP is talking about Claude Code Max 100 & 200 plans. They are very reasonable compared to anything else that has per-use token usage.
I am on Max and I can work 5 hrs+ a day easily. It does fall back to Sonnet pretty fast, but I don't seem to notice any big differece.
64 replies →
Before they announced the Max plans, I could easily hit 10-15$ of API usage per day (without even being a heavy user).
Since they announced that you can use the Pro subscription with Claude Code, I've been using it much more and I've never ever been rate limited.
This is what I don’t get about the cost being reported by Claude code. At work I use it against our AWS Bedrock instance, and most sessions will say 15/20 dollars and I’ll have multiple agents running. So I can easily spend 60 bucks a day in reported cost. Our AWS Bedrock bill is only a small fraction of that? Why would you over charge on direct usage of your API?
1 reply →
API prices are way higher than actual inference cost.
You can spend $200 worth of tokens in a single day using the Max $200/mo fixed cost plan.
Where is a citation on Anthropic increasing cost to cursor? I had not seen that news, but it would make sense.
Seems like the survival strategy for cursor would be to develop their own frontier coding model. Maybe they can leverage the data from their still somewhat significant lead in the space to make a solid effort.
I don’t think that’s a viable strategy. It is very very hard and not many people can do it. Just look at how much Meta is paying to poach the few people in the world capable of training a next gen frontier model.
Why are there actually only a few people in the world able to do this?
The basic concept is out there.
Lots of smart people studying hard to catch up to also be poached. No shortage of those I assume.
Good trainingsdata still seems the most important to me.
(and lots of hardware)
Or does the specific training still involves lots of smart decisions all the time?
And those small or big decisions make all the difference?
13 replies →
interestingly windsurf have done this (I'm not sure how frontier this model is...but it's their own model) https://windsurf.com/blog/windsurf-wave-9-swe-1 but AFAIK cursor have not.
> to develop their own frontier coding model
Uh, the irony is that this is exactly what Windsurf tried.
Why did they fail?
1 reply →
Which is interesting because Sonnet is cheap and Opus is not on par with o3 for tasks where you want to deploy it.
But Cursor is also offering OpenAI and Google models.
If open models become big, open coding agents would be bigger at that point. Even more motivation as well.
It probably doesn’t cost them all that much? Maybe they were offering the API at a 500% markup, and code is just breaking even.