Comment by fschuett | Hacker News Reader

Comment by fschuett | Hacker News Reader

Parent Full thread

I think that next year the prices for tokens will rise a lot, Anthropic made a loss of 42 billion in 2025, which is not sustainable even with tons of VC money. I thought last year that maybe hardware-based models (Taalas, Cerebras) would reduce this cost in the future, so the expected "token price hike" wouldn't come (VC investors filling in the gap until the hardware-models are ready).

But I think that's still years away and Anthropic has severely reduce the tokens on even the Max 20x Plans, OpenAI starts to introduce $500 plans, every new model suddenly "costs X times more tokens" and I noticed that even downgrading to a previous model doesn't do that much on API pricing (virtually no difference between Opus 5.5 and 4.6 in API costs).

So I think that my personal strategy of tokenmaxxing and doing "hard stuff" while the tokens are still cheap was a decent strategy, only with the problem that I'll have to re-learn programming by hand next year, probably. API pricing is too expensive for what it does, yes, it may be the super-intelligence or whatever but at least right now its nowhere near economic.

> I think that next year the prices for tokens will rise a lot

Given the leaps that Chinese models have made, I don't think they can raise prices that much. Most of our devs can already do most of their work using DeepSeek 4.1. It's not frontier, but it's very good, and it's a fraction of the cost. They keep their Claude subscriptions because it's currently good value, but if Anthropic increased the price significantly, those subscriptions would be cancelled.

The more difficult question is around data sovereignty. Companies do not like sending a lot of sensitive data to Anthropic, OpenAI, or any of the Chinese companies. Right now local LLMs are really expensive to set up competently, and they're not as good.

I think the future for OpenAI and Anthropic is the Ferrari model. Very expensive but very intelligent models which people pay to access for important work. There's a market for that. It's very small, but they have big pockets.

I work at a company that holds very sensitive data and I don't think there will ever be a non-US-made open weight model in our walls.

Even if it's hosted by a trusted 3rd party rather than by China?

Yes, the theory is that they can bake in coding behavior in the weights that will introduce vulnerabilities...

Even cheap models like GPT 6 Luna do decent job in programming these days. Token prices would need to rise A LOT to make companies return to writing code by hand, instead of these models.

My company has largely switched to Luna after Microsoft raised prices for Copilot.

I love luna but unfortunately it needs to be much, much faster in tokens per second to compensate for lack of intelligence.

Even with pay-as-you-go pricing I find the models very economical currently (e.g. Opus 5.5). I regularly check in Copilot what I just spent vs. how long it would have taken me alone. Economics do now work out in my favor mostly.

> Economics do now work out in my favor mostly.

That means, it is mostly cheaper to do it with Copilot etc. and not alone, right?

Yes

I don't know if we'll see price increases. My belief based on current market understanding is we'll see efficiency increases.

With models getting to the quality level they can perform a specific set of work at an acceptable quality level we will now begin to cost optimize that work.

We just invested in Lamb Labs from YC Summer 2026 as example company pioneering this next wave of ASICs to run a model.

We're going to see massively lower power usage and much better tokens/sec per dollar of capital expense as well. Think Moore's law but for tokens and output quality.

i predict the price of intelligence will continue to drop (normalise capability per $) eventually hardware production will meet demand. analog and inference chips with baked models have the potential to make inference distinct from training and substantially make the cost comparison asymmetric.

Reply on news.ycombinator.com