I have the feeling that the age of 'i can't be blamed by AI stuff' will be a "this was the computer guy mistake" for a moment.
PS. I've been using Claude opus 4.8 and it is worse than 4.6 and I will say that even sonnet 4.6 is better. PhD. Level of software and engineering I believe! I know many PhD who never coded or worked anyway
If you're following a bunch of people who are from LLM labs, you're going to be more incentivised to tokenmaxx because it's in the Lab's best interest tonget you to behave that way.
Practically, many companies aren't labs with endless runway. Companies hopefully follow a PnL model. And when you look at things with that lens, many of the times the LLM use case falls apart.
You're seeing a bunch of companies starting to realise that tokenmaxing yields very little ROI.
Even the LLM labs, the guy that spent $1+mil tokens has nothing to show for it in terms of revenue to the company. And you have to keep sinking that much into AI for ... "features".
There are some good use cases for AI. I ended up with a positive ROI on a greenfield project myself, albeit on a small scale.
The way that AI has been making people have totally irrational decisions on executive, pure business and technical standpoints is simply mindblowing. I don't understand how people can't take a step back and see what's actually happening from a macro perspective.
Pretty sure from inception the phrase “tokenmaxxing” was never seen in a positive light…
Assuming the intelligence of a model continuously improves with scale, the token price of the best model will become increasingly expensive.
I know that tokens are currently experiencing rapid price drops, but they will eventually encounter physical limitations.
There are so many useless cases such as people bragging about their token consumption that has no product and no value add, or those with OpenClaw doing useless automation that could be a Python script.
It depends where you buy the tokens from. Jevon's paradox exists in China and not in the US for now.
> In just a few months, companies became obsessed with “tokenmaxxxing,” then turned against it due to the high costs.
Casinos (in the US) telling customers to spend more on tokens, introduces free spins, discounts, resetting limits on peak hours. Then introduces new slot-machine that promises to give better odds to the gamblers, but instead is more expensive to use.
The ones in China did the opposite and made their discount on tokens permanent.
All this 'tokenmaxxing' was an outright scam. Now the AI companies want you 'tokenmaxxing' your agents on loops as the token prices increase.
I knew right there and then that he was a moron. There’s something about American companies where the best and brightest rarely show up in senior management. It seems to be populated by some weird class of golf playing NPCs that figured out how to game the system and bring all their cult members along for the ride.
My own company spent 2+ years enforcing extreme austerity, to the point of firing the very people who built everything, only to run wild with AI spending and seeing little results from it.
Surely, out there in the wilderness, there is a company staffed by intelligent, skilled people. Right?
The corporate side seems to be well... stupid? Execs asking their people to burn tokens do not understand the politics and cadence of business. Corporations do not actually demand more work to be completed in the way we traditionally think. Creating a lot of stuff in a corporation tends to naturally banish most of it to the void because that stuff requires other people to exist and engage with it in order to use it, deploy it, get customers using it, etc. AI does not take up that slack in the way that we are being told because it lacks agency. For most people in corporations the problem is not that they can't do their work, their real jobs are mostly being political nodes in a vast system. There is no solution on the table to change that at all.
Of course the question remains, who is supposed to be buying products through this system if AI systems continue to displace jobs?
as Jensen said, get ready for $1000 per mil token
those for which this price makes sense will push out those for which it doesn't - to lower models or to local models
but those who want to run local models need to compete for hardware with the data centers, which have strong scale effects thus will always be able to out price local hardware allocations - can already be seen now as hardware makers get out of retail business
Here are my concrete predictions
1. Token costs will come down and performance will go up
2. Everyone will spend even more on LLMs not less - the article points at small blips but if anyone thinks it will go down from now, you are mistaken
3. AI Companies will be profitable
If anyone wants to counter bet on me, please go ahead.
- The frontier AI companies have realized they won't be able to count on gaining ground and earning more in the future through sheer moat. They have to start earning right now.
- The playing field on the market got a whole lot more even as a result. Now everyone is competing on cost and quality - while there are still a lot of competition. AI suppliers can't easily get away with subsidizing their own product and enshittify later.
I might be missing something obvious here? It feels to me that if the frontier AI companies thought they could gain a lot more moat they wouldn't raise their prices this much this early? And their current moats/head start doesn't seem insurmountable?
Anecdotal experience - my coworkers will use the "max-think" and the most expensive model on every change they do with Claude, pumping out 100k's of tokens just because they can (and brag about hitting the limits).
I suspect this kind of behaviour will need to change in the very near future.
[0] - https://en.wikipedia.org/wiki/Betteridge%27s_law_of_headline...
kimi-k2.6 can do a pretty damn good job with vision for optimizing ui design workloads in a loop. not cheap but significantly cheaper than anthropic.
mimo 3 is jsut pretty damn good when you need a high end reasoning model - also reletivly affordable.
I was able to run gemma and do some coding locally on a 32 gb machine. it was slow as molasses but the fact that it worked at all on a local machine that wasn't desinged around AI workloads is great.
Its only a tokenpocalypse if you rely on these closed and frankly overpriced american models. is opus better than kimik2.6? arguably yes but not 16 times better from what I've been seeing.
When the interaction is exploratory, the marginal cost feels invisible: ask again, summarize again, try another agent. In a business workflow, the same pattern becomes a metering problem. You have to decide which parts actually need a frontier model, which can use a smaller/local model, and which should not be generated at all.
That probably pushes AI products away from "chat with everything" and toward much narrower tools with explicit ROI: less open-ended generation, more constrained pipelines, caching, evaluation, and human review at the points where mistakes are expensive.