[Analemma_]

Every week there's at least one post on the HN front page bitching about API errors from Claude because Anthropic doesn't have enough serving capacity. I really don't see any signs they're "spending too much", the actual evidence on the ground seems to be exactly the opposite: constant exasperation that they're not spending enough.

[CuriouslyC]

Unlike OpenAI, a lot of Claude's infra problems are self-inflicted and not completely raw-capacity related.

[newaccountman2]

What he means is the customers realizing they are spending too much on Anthropic.

[claytongulick]

I just finished talking to a dev manager friend of mine at a household name company.

He told me they are massively pulling back on the AI stuff.

Right now the lashback is about cost, because that's the most easily measured pain point.

Soon, we'll start seeing a deeper understanding of the quality issues. At that point, it's likely this whole experiment gets firmly put in a bin of the toolbox where it belongs.

[icedchai]

I know people at medium size companies where they are tracking AI costs very carefully. They are pulling back to levels under $100/week in AI spend per engineer, encouraging use of lower quality, lower cost models, etc.

[vidarh]

You can run models near 24/7 per developers at that price with judicious choice of subscriptions, so that's not really saying much.

Most people don't yet have mature enough setups to fully exploit that level of use.

[icedchai]

With open models, perhaps. But $100/week isn't going to get you 24/7 use of Claude Sonnet.

[vidarh]

$100/week will get you both a higher tier of Claude and significantly more tokens with GLM5.1 or Kimi, both of which are competitive in a decent harness (for Kimi that means their own CLI - it works well in that, but has a lot of quirks that requires special treatment). Just slightly more will get you all three.

[decimalenough]

I don't doubt you, but $100 is approximately the cost to company of one hour of dev time. If companies end up being willing to spend only 2% of their dev budget on AI, this bubble is not going to last long.

[icedchai]

I agree. $100/week is absurdly low if you want to allow for any real experimentation and productive use of these tools.

[fontain]

I mean Anthropic’s customers are spending too much on Claude. Anthropic’s customers are encouraging tokenmaxxing amongst their employees; measuring employees by token usage. That’s great for Anthropic’s short term revenue numbers but terrible long term because at some point companies will realize tokenmaxxing is not good. OpenAI is much less exposed to tokenmaxxing, which is a good thing.

[solenoid0937]

> at some point companies will realize tokenmaxxing is not good

Why? Have we figured out the limits of what agents can do?

> OpenAI is much less exposed to tokenmaxxing

I don't think this is true, from my own experience & chatting with my acquaintances.

[fontain]

Tokenmaxxing is the practice of measuring employees by how many tokens they use, encouraging employees to burn tokens needlessly, it is unrelated to what agents can do.

If a task can be completed with 100k tokens but employees are considered better performers if they complete it with 500k tokens instead… that’s unsustainable and cannot possibly benefit Anthropic in the long term.

At some point, Amazon and Uber and so on and so forth are going to realize that actually, employees using 100k tokens or even 50k tokens is better than 500k and Anthropic’s revenue will fall off a cliff.

[solenoid0937]

Oh. I thought tokenmaxxing was just removing token limits

[fontain]

I think removing limits is fine. There’ll be overspend and at some point adjustments in expectations as we learn more about the value that can be delivered which will likely result in a reduction in spend, but even now, during this period of relative immaturity about measuring the value of output, so long as more tokens = more output, I don’t think the introduction of limits represents much of a risk to Anthropic and OpenAI. Tokenmaxxing is uniquely bad because it is not tied to any additional value (more tokens for the same output).

And I could be wrong about tokenmaxxing being a Claude specific problem but as far as I can tell, all of the major companies encouraging employees to maximize their token usage are Claude Code users. And the music has to stop on that at some point, whether because the companies run out of money or because they learn better ways of measuring productivity in the AI age. And if tokenmaxxing is what is driving Anthropic’s lead in revenue, it could be catastrophic to lose that, because Anthropic are spending billions of dollars per month on the infrastructure to support it.

If tokenmaxxing is evenly distributed between Anthropic and OpenAI then they’ll both hurt but equal hurt shouldn’t disadvantage either much.