CoolestBeans

The increase in token volume corresponds with a price cut. The S-1 revision already reflects what's going on financially. Anthropic is buying volume with margin. There's nothing wrong with that but we shouldn't expect profit to show back up as Anthropic's costs are increasing and have a fixed minimum. Given that OpenRouter users can switch models with the click of a button, what this demonstrates is that OpenRouter users are extremely price sensitive.

It demonstrates why Anthropic has to IPO ASAP. Private investors are beginning to tap out. Anthropic is selling junior debt over 10% where it was about 8.5% last offer. If Anthropic wants to meet its compute purchase obligations (via actual business), it needs to start tapping public markets ASAP as its position is getting weaker. The size of the float and how much is primary vs secondary sale is the data we're missing right now. If the float is small and from secondary sales, the IPO is an exit and investors would be wise to stay away as Anthropic's runway stays where it is and therefore would require a re-acceleration of revenue growth to survive. If it is large and mostly from primary sales, it is a legit fund raise which extends the runway and revenue growth can be allowed to slow because of the pricing pressure you mentioned.

The issue is that investors expected the legacy revenue sources to compensate for the negative cash flows required for the AI build out. The expectation was that the traditional business would provide a risk backstop because they would be uncorrelated to AI related revenue. The Q3 earnings provide some evidence that the old business may not be the backstop everyone hoped and worse AI may be cannibalizing that revenue. Even if AI manages to compensate, the unit economics of AI is far worse than traditional software, meaning Google et al would have effectively traded a money printer business for a commodity utility, a horrible trade by nearly any standard you can imagine. Of course this is just one quarter but Wall Street is naturally going to hate that possibility being live which it is now.

It means the spending spree is on hold. The credit window for the neoclouds remains closed. Now the hyperscalers are feeling the pain. The narrative was that the hyperscalers could lean on their prodigious cash flows if the data center rush didn't pan out. This round of earnings reports adds some doubt to that narrative which is spooking everyone. Hyperscalers are still projecting a further increase in capex in 2027, but one has to wonder how long that will last if conditions stay the same.

The next step is whether data center capex discipline begins filtering to Nvidia and the DRAM semis revenues. At that point pure play companies must drive revenue growth or else.

Unknown. AI usage went up exponentially this year but a lot of that was driven by access to frontier and near-frontier models and a shift to agents. Does that constitute a decrease in prices because $ per unit of intelligence (whatever that could mean) has gone down? Maybe.

One issue is that token usage and revenue seem to be decoupling in real time. For Jevons paradox to be true, prices need to be going down but total spend on AI has to be going up. That's technically true right now but total usage is going way up while total spend is going up slightly. Another problem is that we may be witnessing the end of "tokenmaxxing" with Anthropic's revenue growth decelerating and OpenRouter's token volume plateauing. Two points of data do not make evidence but the possibility that AI usage is an S curve rather than exponential is live.

The rates environment directly dictates what capital is available to OpenAI and at what terms. When potential investors can get x% virtually guaranteed, then they are less willing to give cash for equity to OpenAI.

1. They are starting to do it. Take a look at OpenAI's product portfolio, there's a galaxy of attempts at application.

2. Selling tokens via long running agents is the only thing that's provably working with respect to revenue growth.

3. They need revenue yesterday. Increases in borrowing rates means access to capital isn't what it was in the 10s. Their runway is much shorter than you would imagine.

4. Some of the decision makers at these companies believe AI will exhibit Jevons Paradox and that any decrease in price will result in induced demand.

5. The frontier models are still subject to the limitations of AI. They're not good at everything. Better to have someone else learn that on their dime rather than yours.

Everyone in this thread is being pedantic about what optimization is but missing the point that code is a precise description of some virtual machine that is realized with near perfect accuracy by the compiler. Yeah you don't know exactly what the silicon is doing with a modern compiler but you do know that the final result of the program for any input will be defined by whatever the correct interpretation of the program is, as if the structure the programming language implies actually existed.

LLMs are not this, and they cannot be this because it is their ability to operate on vague instructions that make them as powerful as they are. They can take an imprecise natural language description of a process and realize it as a process in a programming language. The problem is that understanding and manipulating the precise problem description is your job as a software engineer. The LLM can and will make its best attempt based on what exists in its training data, but it cannot know what is the precise behavior that actually fulfills goals and I think, contrary to the parent article, it cannot determine where the edges are to ask you with enough certainty that you will be able to take your hands completely off the wheel.

I think the nature of LLMs also explains why it is good at the verticals it is good at. Porting is an extremely well defined, replicate the behavior on another platform. There are far fewer vagueries in that. Game dev is not extremely well defined, but the consequences of imprecision do not have a lot on the line. Furthermore, games are pretty self-contained, so there's not a lot of guessing about what could happen or what the behavior of a counterparty system could be. With some pretty big exceptions, it is all there for the LLM to see in C++ (or whatever).

I think it's a combination of it not being a GAAP revenue number and the timing. If you have flexibility in the report of the revenue number people assume you will choose the most flattering one especially before an IPO. If this is the most flattering number they can demonstrate, people are going to worry about what the not so flattering data says.

The other part is how big the debt load is downstream. Anthropic is one of the revenue engines that has to feed the rest of the supply chain, clouds, SPVs, semiconductors, etc. They have to be pulling very large numbers to justify everyone else's finances.

I think the growth is real and tremendous and that's probably more important than this blip in revenue growth. But things tend to look peachy until they're not and so everyone is on watch for cracks in the foundation.

But it is only two points of unclear data among a sea of exponential growth stories. It is just that the total debt amount and rates environment mean the industry has to hit an extremely small target which prices everything toward perfection.

I don't think you can draw an accurate narrative from two data points but I'm going to do it anyway. This and the slowing in revenue growth at Anthropic are not terrific signs. This should be the era of exponential growth for the AI industry. Even if the unit economics are sound at these prices, companies really need more revenue to service their current debt obligations and for now that can't come from price it has to come from volume. Furthermore, this is OpenRouter's statistics which partially represents the cost conscious section of the market and this doesn't line up with an induced demand story.

Again, two data points do not make a full story but if you're an AI bull this is not what you want to see.

The traditional pedagogy still works for those students who could be disciplined enough to not use AI on problem sets and projects. But that's not a realistic expectation. You can learn how to use AI on the job. It is very rare someone gains the baseline knowledge of algorithms on the job.

Programs are swapping capstone projects and courses for comprehensive oral examinations and that's partly from input from industry that there's been a skill drop. But that feels like a panic stopgap and we really are going to struggle with this for a while.

The bigger issue is that undergrad students learning in the era of AI is really a terrible environment for actually learning the fundamental theory that gives someone a long and productive career. And the workplace has never been a good place to learn that stuff. The talent pipeline from colleges has always been a steady trickle despite eras when signups into computer science was a flood. With AI, that trickle has become a scant drip.

Following is my experience and constantly pestering my friends across the industry for anecdotes. It isn't even that AI moves the bottleneck, the thing it solves was never the bottleneck in the first place (except for certain narrow verticals). AI both can't and doesn't do your thinking for you and worse yet, it robs you of the environment where you do your best thinking about data structures, algorithms, organization, etc. Code. Of course we can do that in natural language, but it's hardly the right tool for the job.

All this being said, I don't think AI usage will meaningfully go down in the long run cause it's a very powerful dopamine loop and tech companies are very good at making those. How many people report social media making them unhappy and how many people do you know who have actually quit? Or to paraphrase Jim Gaffigan, nobody says they eat at McDonalds, but the sign outside says "billions served". Somebody's lying (and it's not McDonalds).

The only charitable part of this idea is that it will insulate the stock price from the actual balance sheet. It asks investors not if they want to invest in an EV automaker but in "Elon Musk Inc" essentially. He's been doing some variation of this his whole career, this would just give him the ability to dilute the numbers of each component industry further. The problem is the brand name of Elon Musk is worth a lot less than it was in the 10s.

A payments platform is not a bank. Beyond the services being different and the customers being different (consumers vs businesses), both had to partner with... well a bank, to provide the actual financial services.

The moat for audio effects but especially guitar effects has never been just the quality of algorithm. Yes I know you and the teams at companies like Strymon and Universal Audio and whatnot have slaved over getting every detail right. But what makes Strymon the go to choice for so many players isn't just the sound you can get out of it. Because you can get those sounds elsewhere now.

The moat a physical pro audio product is its interface and its branding. And I know that's tough to hear especially if you're an engineer but its true. People buy Strymon pedals because they know exactly what they're getting and they know its going to be in a format where if you want to dial in a tone without getting into the weeds of menus you can (but if you need to advanced features you can also do that). If people wanted the cheapest way to get those sounds, they could always hook up to a live VST host rig. But you get Strymon because you know that brand name and you know you're going to get a box that you can twiddle the dials on the floor and yes you are also getting that as close to perfect algorithm as well. The sound is part of Strymon's branding but it isn't all of their branding.

Look at it this way, pretty much every original circuit design in the analog guitar pedal space has been copied, duplicated, sold for pennies on the dollar, modified, whatever. But those companies still exist and do well because their moat has never been the circuit, its the branding.

Anyway, I hope you find some peace in the next few days. Its hard to argue that your moat hasn't shrunk and maybe you're in the process of building up that customer base. You gotta think about the copycats and it sucks, but you always had to do that. Its just harder now.

Japanese keiretsu also usually have a bank at their center or have both horizontal and vertical integrations. Where they don't have integration, they at least have financial integration. Whatever Musk is doing is far less cohesive in every way.

The more constraining the language, the better. For example, languages with strong static typing are better for agents because they reject entire classes of errors. This is why Rust is a particularly good target for coding agents. Its hard to write data races, draw null pointer exceptions, access freed memory, double free memory, etc. Not only that, but the feedback loop becomes shorter because the error happens in the compiler and not when that code path gets run (which may never happen because its so hard to write comprehensive tests). This helps correct the stochastic nature of agents and bring them back on track when they wander.

This is the antithesis of the idea that more powerful languages are better but it makes sense. For humans, it costs a lot for the language to get in the way all the time. Yes, a constraining language prevents you from writing errors but it also interrupts flow, forces you to describe processes that are meaningless at the application level, and usually is more verbose. But for an agent, the marginal cost of an additional line of code is less. Both the literal cost but also the opportunity cost.

Its not that Rust is a good or bad language per se, its just that agents don't care about ergonomics while humans do. And when you don't have to care about ergonomics this lets you push the tradeoffs way outside the range that would be even remotely acceptable for us. The flip side is that so far, we are in the mix so you can't go all the way to constraints... not that I can imagine what that would look like.

I remember this being one of the first things I had to learn in my first "real" job. Like someone literally had to tell me there's not much of a difference between coding six hours in a day and one. Later on I learned there's just a class of tasks that just take time for your unconscious brain to break it down to understand it. Algorithms are like this. A lot of things that have unintuitive physical components are like this, e.g. bending on harmonica. Some things you just have to show up every day, give it your best effort and then something will click. Programming is like that.

There's just no need to unify the look and feel. Convergence didn't happen, its time to give up on it.

Man I hate the squircles. I understand for iOS you want apps to be consistent uniform touch targets. But for the desktop let the UI breathe a little, take off the jacket and tie. Such an absence of negative space with the squircles. There's glimpses of creativity like the Notes icon being the top of a legal pad but god so many icons are just white symbol in square.

I know the Leopard era design is having a bit of a retro cool moment right now (frutiger aero) but for me its all about that pre-Leopard Aqua. Something very bright and friendly about that design language where the Leopard era feels a bit dark, moody, and dusky. Bordering on chiaroscuro. Especially when you compare classic Aqua to like the classic Windows look which while coming off as corporate felt like it had a precision to it (at least to my kid brain).

I have to say my hot take is that I actually kinda like Liquid Glass. It feels like a bit of a throwback to that era and dragging objects behind transparency to see the refraction was the first time in a long time I've ever just been transfixed with just playing with a UI detail. Especially since I use the Dock as bootleg taskbar (undock everything so the only things in it are the open applications and autohide), when it does pop up its fun to see everything bend around.

Man even a cursory look of this site which is very much in the Silicon Valley / Big Tech world and very AI forward will tell you that the usage of AI is still all over the place.

Agreed but the math results don't seem out of reach for the publicly available models in 2026. Certainly, they have an unrefined but superior internal model. But the math results don't prove said internal model would be as big of a leap if released today as Mythos was when it was announced.

That's more a function of the amount of compute they can throw at it. They can afford to let an agent run in a loop for longer than you and I can. Its not proof of a vastly superior model.

You sure wouldn't want to rely on one at an international diplomatic meeting even if the accuracy is phenomenal.

I agree and I think the question is then, how much economically valuable work exists in a closed world? My hunch is less than people think. Finance can't rely on LLMs because its adversarial we can't all be listening to the same oracle and evidence of LLMs being any good at trading are nil. And then lawyers, as per usual, will simply regulate their job security into existence. Of course, AI feels amazing to Silicon Valley / Big Tech, software is the product. And often its software you sell to other people making software. But outside of that? Useful but lots of landmines.

People have to come to grips with the possibility that a technology can be extremely impressive but be a bust in utility. I'm not saying that's what AI is, I'm saying that nobody actually knows.

Agreed. You would expect this to be a task that AI has a massive advantage over humans as it has both the testing suite to keep it on track and has the source implementation to reference. The classic failure case for agents is that as the task increases in complexity, it requires more steps to complete. With each step there's a chance for error. As errors accumulate, they become more likely. With verification functions, like test suites or comparison to some source, the agent is more likely to catch errors and recover.

So porting should be perfect conditions for an AI agent. It has the testing suite. It has the original implementation. Hell, it can A/B the original with its working copy and attach the original to a debugger.

So yeah I dunno. I guess despite some people's rhetoric on this site we shouldn't underestimate how difficult these tasks are "in the real world" (as opposed to like a game port which doesn't have real money on the line so far). And it is kind of a miracle these tools can even get in the ballpark and fool a lot of people.

If AI does anything, it amplifies asymmetries. Hacking, scamming, reducing code quality... its really good at making defending against those hard problems even harder.

I appreciate a more nuanced take than the typical "it's all slop" vs "you're an idiot for handcoding anything anymore" crowds on this site. Especially one backed by an experiment.

My experience is agents really need cheap easily runnable third party ways to validate their work and stay on track. This usually means compilers and test suites but it can also mean visual checks (like all those vibecoded game ports flooding feeds right now). Of course, these things are useful for humans too. Tightening the feedback loop is always good. But for agents, the lack of feedback can be catastrophic. Better models can ward off catastrophe for longer but eventually without enough feedback, the errors not only accumulate but increase the chance of future errors.

To the points in the conclusion, humans still have the edge on dealing with vagueness and lack of clarity in goals. And we fill in the details of our goals in real time. In other words, the act of writing code feeds back into the other parts of building a program. Agents aren't like this, if they can see where they're going they can get there with remarkable tenacity but without it they'll get somewhere goal shaped fast. You can use agents to take big steps for you, but you're going to lose some fine touch.