[memoryleakgame]

If these benches from their site hold up (they likely wont)

Wouldn't this compress ai revenue like 15x quickly

If they really have a 4.7 opus high equivalent at 1/16 the cost wouldn't this significantly effect all the current capex and planing

Maybe they are getting elon to cover cost

[vessenes]

It's worth being specific:

"Will this decrease Revenue?" -- only if demand for high quality tokens is inelastic. If demand is instead elastic (grows with cheaper pricing) then revenue will likely increase.

"Will this lower earnings?" -- they have a current inference margin for their old models, and with the Elon deal in place, they have a new inference margin. It might be better or worse than their old one. If it's worse, then they'd need to see a concomitant increase in usage. If they don't, then yes it might lower earnings.

"Will this lower corporate value?" -- no - not least because this company is going to be owned by SpaceX approximately 90 days after IPO -- so all the new owner will care about is being benchmark competitive with Anthropic and oAI for the first n quarters. If they can do that, it will massively increase the corporate value of SX; it's hard to build a frontier lab.

[infecto]

The way I have read their benchmark results is that they trained a model to work insanely well in their coding workflow. It’s not a general purpose model.

One of the surprisingly hardest problems to solve is to get a model to use the tools you give it access to.

[romanovcode]

The problem with this is that we do not know the actual cost. For all we know they might be pulling an Anthropic. Subsidizing costs to get users, then increasing them later on.

[yorwba]

They're offering a model based on Kimi K2.5 for $0.50/M input and $2.50/M output while the cheapest third-party provider on OpenRouter charges $0.40/M input and $1.90/M output https://openrouter.ai/moonshotai/kimi-k2.5 Those third-party providers have little incentive to subsidize their customers, so Cursor probably has a margin >20% on their inference cost.

The real money furnace is the training, not just of models that get released, but also experimental training runs that fail to move benchmarks and are quietly thrown away. E.g. Cursor claim that 85% of the compute for Composer 2.5 comes from additional training on top of Kimi K2.5, where I'm not sure how they determined that, but it can't have been cheap. Then they say "Together with SpaceXAI, we're training a significantly larger model from scratch, using 10x more total compute."

So yes, they're probably attempting to replicate the Anthropic playbook of paying a large upfront cost for a very good model, and then rapidly acquiring paying customers, hoping that the inference margin will be enough to cover the training cost.

[zackify]

this thing is so awesome on fast mode, so far i am impressed, some of its observations feel similar to opus.

i use gpt 5.5 and opus 4.7 a lot every day, if i can get good results at this speed, hopefully the usage level holds up on my team plan haha

[2001zhaozhao]

> compress ai revenue like 15x

that roughly just puts it on par with OpenAI and Anthropic subscriptions in terms of pricing per token

[smallnamespace]

AI revenue has been going up while the cost per token has been rapidly falling. The Jevons paradox applies here. The cheaper software is, the more software is written. There is not a finite demand for software.

[rafaelmn]

> AI revenue has been going up while the cost per token has been rapidly falling

Every model release now has been straight price increases since what GPT 4 ? When was the last time a new flagship model decreased prices compared to the previous one ?

[jstummbillig]

1. GPT 4 has gotten 6x cheaper over it's evolution (from initial release to Turbo to 4o). Maybe you meant "Only since 4o and only since its final release". Alas.

2. We are not interested in how different model naming schemes relate to prices, we are interested in the capabilities. So if you want to learn something about price development you need comparative levels of capabilities, and then look at the prices. 4o is not comparable to 5.5 in the first regard. It is (according to the benchmarks) maybe more comparable to current 5 nano - which is 98% cheaper.

[dktp]

Opus 4.5 became significantly cheaper directly per token

[rafaelmn]

You are right I forgot about that ! I think my point still stands - price per token is not decreasing for frontier capabilities, in fact it's increasing.

[radu_floricica]

This only means the frontier is growing faster than the price is decreasing. It's just the sum of two separate tendencies, and has little predictive value. TBH, I'm ok with this tradeoff - higher capability at slightly higher cost is perfectly fine.

[baq]

token efficiency

[chillfox]

Not seeing that either, tried really using Opus 4.7 today, and it ended up at $50 for the same kida thing that came out to $25 last week with Opus 4.6.

[baq]

each model is different and nothing should be taken for granted, run your evals for your use cases. I'm not using Opus 4.7 for almost anything. I've seen very good improvements in GPTs since 5.2 and Opus 4.5 to 4.6 was quite an upgrade.

[wesammikhail]

Models consume more tokens than ever for the same tasks.

[vb-8448]

I, and I guess basically everyone here, don't have access to OAI or Anthropic books, and it's really difficult to disprove your statements but:

- AI revenue going up & cost/token are not related metrics, at least not in the way you are assuming - basically all players (except OAI for the moment) struggling with capacity and/or reducing-dismissing subscription based solutions in favour of pay-per-use. If token cost/token was falling, we would see quite the opposite.

[lompad]

This is conjecture. There is a reason both openai and anthropic refuse to comment on inference costs. If it were falling so much, they would use it to brag. I really don't understand why so many people keep repeating it without any actual data for the frontier models.

Apart from that, I'm not sure if focusing on tokens is even a good idea, because they are so different from model to model. I'd almost consider them a red herring now.

We could look at tasks instead. Is there anything even remotely suggesting that your typical task you give an LLM now costs less in inference than before?

[epolanski]

I'm not sure that to be the case, it seems like bringing capabilities up and costs down merely serves to induce more demand.