[giwook]

Lots of us have noticed that usage limits for Claude have been nerfed in recent weeks/months.

If anything, these new multipliers are more transparent than anything OpenAI or Anthropic have communicated regarding actual costs and give us a more realistic understanding of what it's costing these providers.

The fact that we were able to get such a substantial amount of usage for $20/$100/$200 a month was never meant to last and to think otherwise was perhaps a bit naive.

This feels like a strategy from the ZIRP era of tech growth where companies burned investor capital and gave away their products and services for free (or subsidized them heavily) in order to prioritize user acquisition initially. Then once they'd gained enough traction and stickiness they'd then implement a monetization strategy to capitalize on said user base.

[dualvariable]

However, inference costs for entirely good enough models are likely to keep declining in the future. We're probably hitting diminishing returns on model size and training. The new generations aren't quantum leaps anymore, and newer generations of open source models like DeepSeek are likely to start getting good enough.

There's going to be a limit to how much they can raise prices, because someone can always build out a datacenter and fill it up with open source DeepSeek inference and undercut your prices by 10x while still making a very good ROI--and that's a business model right there. Right now I'm sure there's a lot of people who will protest that they couldn't do their jobs with lesser models, but as time goes on that will get less and less. Already right now the consumers who are using AI for writing presentations, cooking recipe generation and ELI5 answers for common things, aren't going to be missing much from a lesser model. That'll actually only start to get cheaper over time.

Also for business needs, as AI inference costs escalate there comes a point where businesses rediscover human intelligence again, and start hiring/training people to do more work to use lesser models--if that is more productive in the end than shelling out large amounts of cash for inference on the latest models. [Although given how much companies waste on AWS, there's a lot of tolerance for overspending in corporations...]

[geodel]

> because someone can always build out a datacenter and fill it up with open source DeepSeek inference and undercut your prices by 10x while still making a very good ROI-

Not sure how it all works out. Currently trillion dollar companies can't make a native app for platforms. Everything is just JS/Electron because economics does not work for them.

And here companies can make GW data center running very expensive GPUs for 1/10th of current prices. Sound little fanciful to me.

[bootsmann]

The price you pay for anthropic must include the price of training new and better models which is incredibly costly. If you use the models someone else already spend money to develop you don’t need to pay this price.

[croes]

I guess the new models will still be quantum leaps, but literally: "The smallest possible change in a system"

[hansmayer]

They've been like that for a while actually, I think at least since the big hype around ChatGPT 4.5 (or was it 5?) and that underwhelming, lukewarm, oversanitised presentation by Altman and his team.

[ctoth]

Yups... Mythos is the smallest possible leap. Not a standard model generation advance, not even a version point advance. Just the smallest possible quanta of a change. We are absolutely hitting a plateau any day now. Any day. Any time. Any second now. Yup. Right now! Surely!

[hansmayer]

I mean let's be realistic - all that we know about the "mythical" Mythos is the carefully curated and release stuff by the Anthropic's PR team. Is it really a huge leap they are making it to be? I doubt it. In fact I bet if it was indeed that powerful and dangerous, as they imply, they'd find a way to release it immediately, devastate OpenAI and DeepSeek and secure a leading position in the market. Why is it not happening? I suspect because Dario is again at it, peddling his bullshit.

[cubefox]

Yeah. AI progress is insanely fast if you compare it to anything else. Where else is a one year old technology already hopelessly outdated? 10 years ago is basically stone age.

[madamelic]

I am continually tripped out by the fact when I was 16, I didn't have a 'smartphone' beyond a Windows Mobile 6 phone that had no internet on it.

Now, I have this high-resolution shiny object that can near instantaneously get any information I want along with _streaming HD video to it_ *anywhere*.

15 years even feels like a stone age. I can't fathom what it has to feel like people in their 60s and 70s.

[nonameiguess]

I'm not quite 60, but it's always interesting to me that I feel quite the opposite of this. When I was 16, I didn't have a computer, didn't have a phone, had never used the Internet, but when I think of how life has changed, it's frankly not much. I woke up this morning, scooped my cats' litter boxes, took out some trash, made myself breakfast, ate that, read some news while eating, then lifted weights in my garage, had some work meetings, wrote up some instructions per a customer request from Friday, and am about to go drive to the lake to go do a 9 mile longboard loop.

That's very close to a normal day in 1996. The biggest difference is I read the news on my phone instead of a physical newspaper. The news was not any more interesting or informative because of that. I guess I can also still do the loop reasonably well, but I'm a lot slower than I was in 1996 when I was a cross-country state champion.

My parents are closing in on 70 and I guess I can't speak for them, but I'm at least aware of the daily routines of their lives, too. Walk the dog, do housework, DIY building projects, visit kids and grankids. Seems much the same, too, with the biggest difference being they're now teaching my sister's sons to play baseball rather than me, but shit, one of her sons even looks like exactly the same way I looked when I was 7! The more things change, the more they stay the same.

[giwook]

I think so too.

And at some point even frontier model costs will hopefully come down (if there is still a meaningful difference between closed and open source models at that point) as all of the compute that's being built out right now comes online.

[Fire-Dragon-DoL]

I hope it's true, but right now hardware prices are insane

[hirako2000]

It does feel like the music is about to stop.

It has been years now, of cash injections, investors can't keep feeding the beast forever.

[Gigachad]

This is the best AI programming will be. From here on the enshitification starts and the prices go up.

[ochronus]

As predicted by many. The math is, as usual, mathing.

[ctoth]

It has been years now of reading this same comment... Surely people can't keep typing it forever.

[stuartq]

But the prices haven't been going up by multiples of 6 for the past few years. Things are actually changing now. I don't think it's over, but in the short term, it's going to be considerably more expensive.

[hirako2000]

They will smooth up the spike. Or be subtle and transform the existing quota so that they run out more quickly. Calling it caching, compression, optimisation, of course for the sacred benefit of the users.

That would be, even is, the smart thing to do.

[Applejinx]

And it didn't really get flawless, did it? All the same objections stand, but the cost is inevitably blowing up for the same kinda jank product.

[soraisdead]

The difference is we're now in a world where Disney has pulled out of OpenAI without comming, and Sora was dropped off a ditch.

In other words.

The bubble has burst. You're just in denial.

[hirako2000]

My read is that the bubble as burst internally (angels, seeds, VCs, and even corporate got a grasp of the inflated promise). It will take while for the actual bubble to implode.

[tclancy]

I’m not willing too, but I can set up a cron job to Claude -p the task.

[stefan_]

Dunno, if in this day and age you are making inference more expensive, more scarce, you are honestly moving in the wrong direction and DeepSeek and others will gladly take your lunch.

[Gigachad]

The hardware to run deepseek is still incredibly expensive.

[cheema33]

> The hardware to run deepseek is still incredibly expensive.

Deepseek API pricing is very low compared to Anthropic/OpenAI API pricing.

For many, the 300% difference in pricing may be difficult to justify, if the quality difference is very small. And there will be many tasks where the most expensive/the best model, is not needed. Currently many people end up using Opus 4.7/GPT 5.5 for many tasks without thinking about it.

[Gigachad]

Is deepseek still on subsidized pricing though.

[2ndorderthought]

Near zero probability of that. The model is more efficient and the company who trained it did not blunder trillions of dollars to do so. China has better electricity infrastructure than the US too, so the likelihood they can scale out before the US ever could is high. Long term deepseek, Alibaba, etc hold the most cards for sustainable AI even despite the attempted Nvidia embargo

I am not shilling China, this is just what is happening right now.

[ConSeannery]

Lol what? You seriously think that the #1 Chinese AI company is not being subsidized by the Chinese government?

[nl]

Judging by the multiple providers selling it for around the same price (including non-VC funded competitors): no, it isn't subsidized.

[johndough]

Is there somewhere I can look up whether a certain provider is VC-funded?

[2ndorderthought]

Have you seen the news about qwen3.6? People are running it on sub 1000 euro hardware. Apparently it's about as good as Claude sonnet.

[sergiotapia]

That is folly because there is very minimal cost to switching providers, let alone models.

[bluescrn]

Did anyone really expect AI to be cheap?

If/when it gets to the point where it can replace a skilled worker, the service can be sold for close to the same price as that skilled labour. But the AI can run 24/7, reliably, and scale up/down at a moments notice.

There's not going to be much competition to drive prices down, the barriers to entry are already huge. There'll likely to be one clear winner, becoming a near-monopoly, or maybe we'll get a duopoly at best.

[hansmayer]

> Did anyone really expect AI to be cheap?

Yes, a lot of people (not me). Why? Well because that was the whole value proposition of these companies, relentlessly pushed by their PR and most of the media- rememmber it was something something Pocket PhDs, massive unemployment etc?

[rwyinuse]

"There's not going to be much competition to drive prices down, the barriers to entry are already huge. There'll likely to be one clear winner, becoming a near-monopoly, or maybe we'll get a duopoly at best."

Based on what exactly? So far every time OpenAI, Anthropic or whatever has released a new top performing model, competitors have caught up quickly. Open source models have greatly improved as well.

I expect AI to be just like cloud computing in general - AWS, Azure, GCP being the main providers, with dozens of smaller competitors offering similar services as well.

[2ndorderthought]

Right now China is flexing the future in my opinion. Smaller, widely available, frontier models for pennies on the dollar.

I think the future of ai will be breakthroughs that let it run on commodity hardware, and the average person will not be paying for it from the cloud unless they want to be surveilled or are stuck on older hardware.

Right now I am running about what was a frontier model 1-2 years ago on a junk machine. Some people are running what was a frontier model 4 months ago on PCs and laptops that cost 5,000. In a year I think the landscape will be even better.

[flir]

I do. "Commoditize your complement". Want to sell lots of silicon? Give away good local models to run on that silicon.

Even if SOTA models in the cloud are a few percentage points better, most work can be routed to local models most of the time. That leaves the cloud providers fighting over the most computationally intensive tasks. In the long term, I think models are going to be local-first.

(Unless providers can figure out a network effect that local models can't replicate).

[bluescrn]

> I think models are going to be local-first.

Why on earth would that happen when everything else is moving into the cloud to tie it to ever-escalating subscription fees and prevent piracy?

Even with gaming, where running high-end 3D games in the cloud seems like madness and inevitably degrades the quality of the experience, they won't stop trying.

[vanviegen]

> In the long term, I think models are going to be local-first.

Why? There's an inherent efficiency advantage to scale, while the only real advantage for local models (privacy/secrecy) hasn't proven convincing for broader IT either.

[solid_fuel]

Local first models aren't just more private than the API vendors, they also have the advantages of fixed cost, lower latency, and better stability - local models don't get nerfed/"updated" in the background like chatgpt does.

Maybe in a world where these AI companies behaved with some semblance of ethics and user-friendliness they would be on even ground, but for anyone paying attention local models are obviously the future.

[still_grokking]

> the only real advantage for local models (privacy/secrecy) hasn't proven convincing for broader IT either

Because of nonexistent regulation. Just wait for it…

The legal situation in for example the EU is crystal clear, only that it will take some time to go though all court instances.

[2ndorderthought]

It's foolish not to care about privacy especially as a company. You know how it prevents you from emailing yourself your tax documents? Meanwhile thousands of employees are sending literal design docs, software, product goals, etc to several ai third partys. Not only is that insane, the companies they are sending it too intend too and openly admit to scanning the data, make software products themselves, and intend to create models that can produce their products automatically.

The reason local models hasn't caught on is several fold. It's marketing to say your company follows the latest trend, and there's an inherent pressure to keep AI companies afloat so the economy doesn't entirely collapse. The other is, it wasn't until the last month that these models have caught up to frontier models. They just did, and they are more efficient and don't require a team of 500 to deploy.

[LtWorf]

To not depend on an external company that can decide the price.

[Dylan16807]

That's a silly reason. For non-agent use cases what kind of utilization are you going to average on your own GPU, 5-10%? And that's without batching.

Even with overhead and scaling for peak use and a large profit margin, any company with an ounce of competition will be vastly cheaper than self-hosting. And for models you can run yourself, there will be plenty of competition.

[LtWorf]

I think you are calculating with current prices. Try to extrapolate the price in one year, seeing the current trends instead.

[soraisdead]

> Did anyone really expect AI to be cheap?

Considering most of the cost of producing a model is the upfront cost rather than the running one, I kinda still do.

The point was never to produce 4 frontier models per company a year.