> I noted that my own token usage comes to about $1,000/month against each of Anthropic and OpenAI - which currently costs me just $100 per provider thanks to their generous subsidized plans for individual subscribers.
Do we know that AI providers are going to keep these per-token prices, or eventually lower them because of competition from China?
Many lower-budget individuals are now moving to China open weight models like DeepSeek. I wonder if China's really subsidising the providers, or if inferencing costs are actually much lower, and Anthropic/OpenAI are just making sure no money's left on the table for their eventual IPOs.
One aspect Paul Kedrosky mentioned recently is the concept of „duration mismatch“. The price per token goes down over time (either because the AI vendor reduces due to competition pressure, or because customers are now incentivized to use older cheaper models). But datacenters are financed through debt, with the assumption their revenue increases over time. Quoting him: „[AI vendors are] paying for a fixed cost with a depreciating commodity“[0].
So you have on one end the token revenue trending down, on the other end the training cost going up for the next frontier models, and you need to pay back your 10y debt.
Per token costs will fall, but the harnesses will get more token hungry. Instead of just centering the div it’ll spin up a battery of agents to architect, critique, advise, code, review, refactor and so on.
I wish I could disable most of these. I already hate all the "oh you're actually right, let me fix that" nonsense. Then it proceeds to burn 50k tokens on the git history instead of copying logic A from a different part of the codebase to logic B, where I want that exact logic without having to write the boilerplate myself...
Makes me think of how my Claude.md files specifies to use the built in framework code-generators (rails). Those generators are deterministically right every time.
I wonder how often the Agent actually follows the guidance. I do see them follow it when I look. But it doesn't seem so every time.
This is tricky since it can and will ignore your md directions. When possible I try to lean on tool call hooks or skills that invoke deterministic scripts. As much as you can remove the "choice" the better though still there's a lot of randomness in how reliably it invokes skills ime.
A lot of the time if you're copying code from one place to another what you actually want to do is abstract it so you can reuse it in both places.
The LLM can easily do this type of stuff, just tell it and it'll happily do it. This is exactly what I mean when I tell people they need to work closer with the AI, tell it how to do things. Don't just tell it what to do and get frustrated when it does it differently than you would.
A good way to achieve this without writing huge prompts is tell it to plan the change first. Just give it some vague low-effort directions. It'll usually get most things right, you tell it what you want different and once you're happy you tell it to go ahead.
Nah the codebase is legacy fucked and I cant be bothered to try and optimize business flows without the fear of other stuff breaking.
Claude 100% of the time even thinks we use laravel despite the project being some old lumen codebase, so most of laravels features are not available. It also gets the PHP version we are using wrong 100% of the time.
> Do we know that AI providers are going to keep these per-token prices, or eventually lower them because of competition from China?
I genuinely do not know how prices can get lower from the current major providers in NA without the whole market collapsing. Everyone is spending copious amounts of money to presumably make more money back.
An inference only platform selling good open weight model inference without the research overhead could capture a-lot of market for lower size model uses (haiky, gemeni flash). Diffusion-transformers and clever cashing can drop inference even lower, which is improving at a high rate.
The biggest reason large models are un-attainable for local applications is the lack hardware with large amount of unified/graphics memory (and the cost of the platforms that do). Once the memory slog goes back to normal and hardware manufacturers adapt to demand, we may see consumer hardware with large memory capacity effectively opening the door for slow but usable frontier model inference (assuming improvements in model efficiency and compute capacity)
At that point, inference becomes a race to the bottom. The large labs hope they can attain a leap in capability (which is increasingly looking bleak, with a average catch-up of just a few months) or market dominance through integration (integration in platforms and OS, exclusive deals with companies or governments).
For coding agents, i suspect no player will manage lock in enough market to enforce pricing much higher than the true inference cost, and catering to programmers becomes an unsustainable proposition. We will instead be further hit with a lot of AI integrated into our other tooling costs, such as GitHub, Microsoft suite, G-suite, forcing in AI functions as a value-ad into the total cost without giving the option to exclude them. (using their market position)
“Any” is a very high bar Unless laws prevent it, I don’t see why a substantial minority wouldn’t buy services from where they can get them at a similar quality and much lower price.
Any IT cost center will send to the lowest bidder. This isn’t intellectual property: it’s annoying shit that is an unwelcome cost of doing business. China might copy our tedious scripts? Will they make a product out of it? Can I buy it and fire my IT staff? Great!
Not everyone using AI is using it to code core value IP.
Why there are so many people that still believe that AI coding is a fad? It's something that started less than two years ago and companies are already paying thousands per seat. I know one that gives you 5k per month. Which other tool went from nothing to this level of acceptance so quickly?
No disagreement on computing 2.0, but companies spending 3-5k per employee for hardware isn't generally a monthly cost. It's a at the time of hire, and then once every 3 to 5 years after that, for a monthly amortized cost of about $50/employee.
I have my concerns with current inference pricing in that there's a non-zero possibility for a rug pull in the future for the subscription plans for organizations and individuals that can still use them. For now, its only companies larger than ~150 users that need to pay per token, but what if that wasn't the case? Not every company can afford over $1k/month/employee to give them access to AI tooling, further making it harder to compete against the behemoths. If we get to a point where an individual can no longer pay $100/month for nearly unlimited usage and instead must pay per token, that's going to be a problem.
Personal computing eventually became an equalizer (until we started centralizing on mainframes again, aka the cloud) because it got cheap. My hope is that inference also gets just as, if not cheaper.
I have high hopes for local AI and open weight models and we will continue the ethos of local, personal computing and not needing to offload everything to OpenAI/Anthropic/Google, etc. to get work done once the hardware and hardware availability catch up.
Hardware's not generally a subscription, monthly cost though.
You update it for them every 3/4 years (if they're lucky).
It probably makes a bit more sense to compare it to existing software subscriptions like Office, or the old-school 'per-seat' licenses per user for software.
Two things can be true at the same time. It can be true that this is here to stay. It can also be true that companies are grossly overvalued right now and that the market is irrationally exuberant. This would mean we could both have a crash and also see AI coding be the new future.
How many more months do we need to wait, until big companies realize that flash models work just fine if you:
1) Don't ask LLMs for big changes
2) Review everything and point them in the right direction
Large models still suck at big changes, they produce questionable architecture and you still have to review the code, if your project is serious enough.
The codebase quickly become a mess, if you don't pay enough attention. Does not matter which model.
So why bother with big models, when flash models are 10x cheaper and much faster to iterate under guidance? Large models can be used for security and bug audits. Flash models work almost the same for changes under 300 LOC when you dictate how you want your code to look.
The easy decision is to just go with the biggest SOTA model you can afford.
But this overlooks the other critical part of getting the most out of these things: the harness. I run an autonomous plan/design/code/build/test pipeline with agents using my own orchestrator. Different models are better at different stages, and I use LLMs to judge the output between them. Not everything needs Opus 4.8.
The harness provides both the scaffolding to get the right things into the model, and the right things out. But it also lets you dictate which model does which work.
It's the pipeline, not the model, that gets you quality at a given token budget.
I wonder to what extent models should figure out which model to forward a query to. Or perhaps the big models could learn the difference between an easy and a hard question and charge accordingly? Perhaps, if it can measure complexity, even generate a quote?
Small models are fine for small coding tasks but I don't see why big ones can't be broken down most of the time.
> I wonder to what extent models should figure out which model to forward a query to. Or perhaps the big models could learn the difference between an easy and a hard question and charge accordingly?
This sounds like something a harness could do (and might already be doing), with work delegated to subagents running on lower-cost models.
128 GB machines that can run local LLMs are a bargain even if priced $5-8k. Yes, tok/s is not quite there, but that's probably OK since the bottleneck really isn't the code; it's WTF did Uber build with all of that spend? How did it meaningfully impact their revenue in a positive direction?
You’re way better to run your own on premise models. Laptops are depreciating assets, do not benefit from economy of scale, have fixed specs, result in a fragmented fleet where you need to keep models up to date. Without talking about power consumption and cooling issues. I really don’t see why companies would go that direction
How is tok/s not a bottleneck I? I assume most people still use ai agents interactively rather than leaving them to do their own thing during the night.
I find anything below 50 tps or so entirely unusable...
Regardless its Apples to oranges anyway, inference is quite cheap for open weight models its just that Claude and OpenAI can charge very high margins compared to e.g. DeepSeek or various provider on OpenRouter since open models are a commodity.
I startup 4 or so projects then go do other things for 4 hours. I don’t have enough energy to steer overnight, but I’m at least “semi afk” for daytime steering. So throughput is king for me, tokens per hour. Not latency or actual tokens per second.
Running locally is even worse for this, because if you're running 4 jobs at once they just run at 1/4 speed. Not literally, you can make up some of the difference with batching, but you have limited resources instead of spreading your requests out on an API provider's nodes.
Even if companies decided to move away from expensive models from the major labs, it probably much more economical to pay a cloud provider to host some open weights model which could then be amortized across all (internal) users and do inference at a substantial batch size, rather than giving everyone their own hardware -- which means the company would need to provision for peak usage and inference at batch size of one.
I agree on the basic point, but running $1500/mo's worth of SOTA local AI is non-trivial already, and that's a figure for a single seat. That's equivalent to generating at least 20 tok/s on a 24/7 basis, in fact probably quite a bit more than that (because open-weight models are vastly cheaper than proprietary ones even when served from reputable Western providers - reaching the same spend would take around 100 tok/s or more, which is well within datacenter hardware territory).
You could probably reach the former figure on a prosumer platform but only for very special workloads. If you spend a lot of time on prefill (which is common for agentic workloads) the outlook is even worse since that's a significant constraint for any on-prem AI.
I think companies will eventually just buy a local AI server.
Using local hardware is expensive when it's running a complicated software stack that can break in 10,000 different ways.
These eventual local AI servers will just talk some protocol for AI and sit in the corner and nobody will think about them.
I guess they still might need access to various systems, so idk. Eventually I think someone will offer "AI in a box" though, running the latest open model or whatever.
Yep, its already quite easy to do so with tools like opencode/openrouter. Ive used some open source models and they seem … ok? Im not doing foundational math, just refactoring code, understanding existing code etc. I don’t see a future where companies blow 11% of employee compensation on a single tool; the hosted AI server + oss models will 99% win out.
I don’t think companies will do that. Why don’t they just buy local on-premise infrastructure even though it’s cheaper than AWS?
“AI in a box” sounds a heck of a lot like “the box” from the Silicon Valley TV show. Or the Google search appliance. Or name any other on-premise thing that is equally dinosauric.
The real finding of this article is that AI tokens are direct competitors with offshoring. $1,500/month buys you a whole employee in India.
And this is before AI companies inevitably increase pricing after the conclusion of the growth phase.
> I don’t think companies will do that. Why don’t they just buy local on-premise infrastructure even though it’s cheaper than AWS?
For customer facing, production software, its worth paying a cloud tax to get the reliability guarantee. For tools that are used by engineers for code development, there is no need for such bulletproof guarantees.
at their scale they could also just run a large on-premise or rented (basically still cloud, but cheaper) GPU cluster and run through that. fixed costs, even license a SOTA model’s weights if you’d like
> even license a SOTA model’s weights if you’d like
Yeah, I bet all labs releasing SOTA models are more than happy to remove the main way they make money and let you run it locally, especially if you're a big spender like Uber who seems very willing to throw money into the sea as an experiment.
Anthropic and OpenAI license to the public clouds. Google reportedly licenses to Apple. licensing to Fortune 100 companies running on their own infra is an obvious next step
it is a race to the bottom and I’m not sure the labs win that race. we’ll see!
The problem isn't really Uber, Microsoft or Nvidia, it's all the smaller none IT companies that also have developers on staff. They are screwed. $1500 per seat per month is just way to expensive, but they also can't afford to build and maintain their own on-premise solution. If Microsoft can't afford to run CoPilot for their own developer, what chance does any of their customers stand?
If the large, well founded IT companies in the world believes the current AI cost is to high, then Anthropic, OpenAI and CoPilot have no actual customer base. AI is then relegated to very profitable niche business, but that can't fund the R&D for the models.
Why are smaller non-IT companies "screwed" because they can't pay out the nose for their developers' AI usage? They're non-IT companies, developers are presumably not on their critical path, or not their bottleneck. Developers can keep on writing code the old way, or doing it with a more reasonable AI spend. I don't see how this "screws" any company.
You can ask the same for the median 330k salary in the US for Uber Engineering...
and being a bit snarky, attending Uber engineers talks here and there at a few conferences, looks like. they love to (re)invent internal tooling/platforms. That's pretty expensive on its own.
EDIT: I'm not saying that Uber's engineers didn't add value to the company, they absolutely did and handling the scale up they had to handle is not an easy feat. But I do challenge the notion of "what features did they create with that (LLM) spending?" of GP.
> You can ask the same for the median 330k salary in the US for Uber Engineering
People DO.
It's well known that most tech companies are ran incompetently. As you say, it's not the engineers' fault.
But most projects and hiring in these companies exists to juice promotion criteria. And that, depending on perspective, these companies are either massively overstaffed or massively underproductive.
The comparison to AI spending being wasteful holds up pretty well, these are companies that readily piss away billions in pointless spending.
This is a very good answer but there's a flip side too.
The idea of "if you add intelligence you make more money" is contradicted by the fact companies don't just always hire more people. Wy doesn't google just hire everyone?
This is what all "platform engineers" have to do once things are working nicely: you have to keep inventing work.
I don't know; I'm a Ron Popeil "set it and forget it" kind of guy. Make the dumbest, simplest thing that's going to work with some clear path for scaling. Then go do valuable things instead.
But most Platform Engineering teams in smaller companies (and especially non-US) add a layer on top of existing technologies. A layer that usually maps to the specific culture and idiosyncrasies of that company; a bit like the deployment flow which is usually very specifically shaped on how a company is.
But in Uber's case, they tend to reinvent lower level pieces of platform/infra.
Right - the future of LLMs is like ol' windows XP+Dell. Commercialized "things" you run locally offline, co-designed with hardware, with a known productivity suite, and large businesses building the next generation thing and suite with 18mo release cycles (ish).
I don't see it. Leasing equipment and paying per seat license fees makes a lot of accounting and cash flow sense. Maybe when it gets to the point where you can run SOTA LLMs on consumer hardware. But that seems a solid decade and probably much more away.
Even then it makes more sense to rent the bigger GPU and get your answer faster.
There's waayyyy too much money betting on that not happening, to the point I feel there'll be regulations popping up for "safety reasons" etc to ensure the big players control this.
XP? I can see the argument for enterprise support but in that case the latest windows OS is going to be virtually free and I dont know if MS and Dell etc. would even support an XP machine. Might even be required for hardware. If no enterprise support wouldnt Linux make a lot more sense?
I get that if it's offline the security downside of XP doesnt matter, and I assume XP is free, but being free doesnt really seem that valuable compared to alternatives (free linux and virtually free OS if buying wholesale).
"Windows XP+Dell" should have been in quotes. It's similar to the way enterprise productivity software was developed, packaged co-designed with hardware, and sold on an 18mo upgrade cycle assumption. It's not literally windows xp.
I think probably the correct spend is something closer to 10x that if people can figure agent coordination problems out. It's not even really about capability at this point, it's about keeping track of what agents are doing.
Your last question is really important. What did they accomplish with all that spend?
I suspect there’s some mass delusion with respect to actual accomplishments as a result of LLM use. Sure, things are moving faster, but does it matter?
I am wondering more and more if this becomes true as these smaller models take off. I might be old fashioned but I have yet to crack the workflows some of the hype people spout like Claude codes Boris where he and others talk about running hundreds of agents overnight.
I have still found the sweet spot for me is using LLMs but I am still in the drivers seat.
That's because for some of these folks, the cost of the tokens doesn't have to match the value of the output; the hype from the story is all they need.
Normal people have to produce something of value from that spend. So starting 100 agents and then waking up to something cool but useless just means you spent a few thousand dollars and created nothing of value............
If you believe a 128gb machine that is essentially DGX Spark in a laptop chassis can run models comparable to SOTA you either never ran open models on hard tasks, or you aren't scratching the surface of SOTA closed LLM capability in how you're using them.
Can you show me an example of a hard task that can't be achieved using light models? When we don't want the model to work on autopilot without reviewing the code at all. Even SOTA models will produce garbage code, if you don't guide them all the time.
Hard tasks require a lot of guidance and code reviewing, unless you are creating another throw away project where correctness, maintainability and code understanding does not matter.
What's the point of running it locally though? Inference for open models is quite cheap already. They could just selfhost, anyway. The experience of running LLMs locally will be excruciatingly bad in comparison at least for the near future.
WTF did anyone build with all that spend? Despite all the feel-good anecdotes about how productive folks feel using ai coding tools there's a deafening silence when it comes to actual, demonstrated efficacy. How can we be this far entrenched in these workflows and still not know whether they actually do anything useful?
I can say at least for me at a small-ish company (~40 FTE) there has been a surge in internal productivity tools. Nothing to improve the end user product directly but a lot of tools to make processes easier and less error prone.
What would previously be janky internal dashboards or excel sheets are now actually nice to use tools. That said of course the maintenance cost of all that has yet to be discovered, and the ROI is questionable.
About the same ~40 FTE team. We're doing the same thing. Smattering of internal tools, but no net gain in external revenue. Who knows which of those tools will have any value or ppl are just doing it because it's cool now to make fancy dashboards.
Yeah this seems to be a pretty widespread story, from what I've heard as well. The thing about those janky dashboards and spreadsheets though is that somebody understood them and built them with intent to solve a particular problem. Despite the rickety appearance, they're trustworthy tools. A polished single page app might look nicer but it's harder to debug than an excel sheet, and much less transparent in its internal workings--especially if nobody actually wrote it...
~70 FTE Engineering team. We are shipping more features, especially features that previously would not have survived the cut to make it on the roadmap. Even though we are shipping more, our total amount of escaped bugs has not increased, so our escape rate has actually lowered. On top of that we are able to triage and fix escaped bugs more quickly now. And then of course there has been an uptick in internal tooling that makes the rest of the company more efficient, and we have been able to address tech debt at a higher rate than before.
I don't think this would have been possible without having solid engineering culture and processes in place before bringing in ai coding tools.
And I don't want to sugarcoat it, this hasn't been easy, requires continued discipline, and took well over a year to get good at. And we still have to continuously learn, experiment and adapt our training, tooling, and processes.
There can be an increase in productivity without a corresponding increase in total output. The gains could be captured by software engineers doing a days work in an hour then fucking off in a variety of ways.
Quite possibly. Doubftul it will happen all at once. If you can get 8 hours of work done in 1 they'd need to ramp up demand 8x. Would be interesting to see that happen over night. Happy monday. Here, take these 30 tickets.
Imo its pretty clear that anyone who is taking the issue at least somewhat seriously knows the amount of value they provide is not non-zero. However, the problems are manifold: firstly, toolchains vary wildly, from fancy autocomplete, to engineers chatting with codebases they're unfamiliar with, to people integrating them into devops and infra, to people doing spec driven development, with a thousand philosophies inbetween. Many people suspect that those above them in the ladder are on the cusp of massive failure due to losing track of the code, and many people higher on the ladder think those below them are overly cautious. I hate to be the guy saying "oh it must be somewhere in the middle", but I will say at the very least I like being able to use it to read docs for me, and to synthesize syntax and simple scripts (give me a join that works across these tables and gives me column x, y and z - give me a python script that parses a file like this example and extracts abc data - given this api spec figure out how I can get this data from this endpoint, go)
as for building actually complex software, the art of that is not in simply chaining together such scripts. Its the art of using architecture and testing to shape uncertainty, and developing requirements (and extrapolating sensibly from incomplete requirements). I don't think llms are great at this, but they arent terrible either. A lot of the more active users in the space are doing stuff where theyve realised they need more detailed specs, which like, yeah, we knew this already - better defined problems lead to better software.
I agree the most interesting use cases I've heard of are about increasing the rigor of software development practices, but there's definitely a lack of coherence in methodology.. I believe that some users and companies are successful in this effort, but the odd (and interesting!) thing is that so far we don't seem to know how to communicate how to do it successfully.
You can't get an edge using local models, these guys may have competitors that will spend on SOTA models. They won't likely ever consider local machines even for some offloading scenarios, the complexity and costs will be even higher.
Consider rewiring your perspective: getting an edge doesn't really matter; the only thing that matters is will customers pay for this? Is this a useful, valuable problem to solve?
Coding faster doesn't really solve that.
Uber makes more money if people buy more rides, order more food, have some breakthrough in autonomous driving. They can save money if they can optimize some ops or spend somewhere. Is there any evidence that with the spend on AI that they achieved any of this? If they did, I'm sure we'd hear about it in some engineering blog.
$18K a year is a fraction of the salary of a junior engineer.
Claude has allowed me to do refactors that would have taken weeks to instead take a couple of days. It has, objectively, increased the velocity of the engineering component of greenfield features by 40% in my org. You can put a number value on that and decide if it gives you favorable ROI.
I think this is the part I struggle with. The code I write makes me money or is a way of teaching me something, both of which are reasons that I would write the code regardless.
I don’t think I have any projects in mind that I’d be willing to spend half of a car on that I also wouldn’t have written myself.
Obviously just a personal take though. I’m glad you get the usage you want out of it.
My "job" is building open source software for data journalism (and anyone else who needs the tools data journalists need, which is pretty much everyone else). I can build more of those tools, and better, in exchange for a fraction of the cost it would take to hire a team to help.
No, that's not what it means at all even if just doing it purely in math terms. Really it is just a reasonable amount to cap at to stop the long tail of super spenders (tokenmaxxers). You could also call it "the amount of AI spend after which Uber has decided there is diminishing returns for the average engineer".
It's not so simple to determine and generalize how much value AI adds. It's going to be different on a per-company basis and a per-engineer basis. It's also affected by the competitive market place and how many other companies are using AI for their engineers.
For example, what if you're a tiny startup and you're considering whether to hire an extra engineer or do all the coding yourself. I would estimate that AI is worth far more than $18,000 a year in that situation where you might reasonably decide to put off hiring an engineer.
I find it really doubtful anyone has managed to quantify that in any meaningful way. Seems like mostly an arbitrary number. Also the article does claim that's its actual several times more than 18k if you are fine with using Codex, Cursor or etc. when you Claude tokens run out.
Not really. There are clearly diminishing marginal returns, so it's likely that the first $2,400/engineer/year adds >>$2,400 of value, even if 18,001st $/engineer/year adds <$1 of value.
It's among a wave of fresh "non-insane" takes on AI in the enterprise. Maybe we can reel things in to a sustainable level before a giant bubble bursts.
The $1500 number is less interesting than the fact that they hit a ceiling at all. Most engineering teams I've talked to have no idea what their AI spend is per developer because it's buried in a consolidated cloud bill. Having a hard cap forces two useful conversations: what workflows actually justify API calls vs local inference, and whether the output is being measured against any real productivity metric. Without that feedback loop it's just a race to see who can burn tokens fastest.
In my experience, this is far below the cost the average dev will incur per month so this seems very reasonable to me. And, no doubt there are exceptions for heavy users so they can get some extra token usage when they need it.
unless they changed something in the like 2 months (edit: besides implementing a cap for claude code specifically, since other tools already had caps) since ive left my job there im pretty sure 1500$ is the very max you can use after maxing out free calls, initial budget, then 2 extensions individually reviewed by your manager
higher ups pushed for these last 2 years to be AI focused so I don't think this restriction is a measure of "don't use too much AI" as much as it is a measure of "don't use only 'manual' AI tooling" since we had a dozen more specialized tools in-house running locally or otherwise that didn't count towards the budget
I wonder what they are doing with $1500 per month. I'm on Claude Pro $20 plan and I'm doing well. That's 3 days per week. On the other 2 days I'm using a customer's Claude Max, I don't know if it's the $100 or the $200 plan, but I'm sharing it with some of its other developers.
Your other plans are fixed price with rate limits where you get more tokens than the dollar equivalent you pay monthly. These plans are economical only if majority of users spend less tokens in $ than the plan's costs. This subsidizes the gap vs. power users who spend multiple k$ monthly in API tokens.
Yea, I’m sure the personal plans are subsidized. I have $200 Claude Max at home and straight API pricing at work and equivalent work would easily cost me 5x if not more on the API.
These are still at currently subsidized prices. We'll see if they think they're getting $1500/month of value when that buys significantly fewer tokens.
yes, and theres no evidence that they arent (or can't) use profitable inference to subsidise those other expenses. Some companies will keep spending massively to train better models, and some other companies will not, and offer good api prices. Which will end up being used? That depends on whether the spending turns into better value models
The fact that Anthropic models are offered at the same API pricing by not just themselves but AWS, Azure and Vertex despite Anthropic taking a major slice on licensing along with the cost an open weight 1T parameter model like K2.6 costs to run on any third-party provider, make it unlikely that API inference cost are subsidized by the labs.
Openrouter? i.e. Even excluding Deep Seek inference for very large open models is way cheaper. Maybe these providers are not very profitable but its highly unlikely that they are losing $4 for every $1 they make since selling inference is their only product...
This is market introductory pricing that hasn't factored in cost recovery. Most of it has been run on early investment with the assumption they will recover costs in the long run. The prices are subsidized across the board and they will need to go up signficantly to recover them.
Yeah, that's not going to work if you can get e.g. 80% of value by using 10-20x or more cheaper open models. At some point it would just make sense for large companies to rent compute and deploy their version of DeepSeek or whatever (if they don't trust Chinese providers)
Assuming this were accurate, then presumably the AI companies would be betting that inference costs come down before the bill is due - I don't see enterprises being willing to absorb another ~10x price increase for tokens (as they've just done going from subscription prices to per-token pricing)
For claude shops this was a huge hit. But lets back this up. There are some companies that haven't even built a break-even model at this price because they are funded by investment. As soon as those investors lose patience the first dominos will fall. For those who have somewhat of a business model, will it survive a price increase? The bigger question is do the base model providers have enough runway and have a way to keep going as they need to recover costs.
Aren’t the Chinese labs quickly turning them into a commodity?
The open-weight models will have a steady race to the bottom on inference costs just by dint of competition between providers. They aren’t at the frontier yet, but they are rapidly eating the flash market.
True but they will raise prices slowly so people will optimize their workflow so they aren't just throwing as much inference as fast as possible like the current state. Right now you should do everything you wanted to try out because it is cheap (as long as you don't become dependent ... the risk).
Uber is in the business of experimenting with robotaxis and automated food delivery.
They can't say that $0 per employee is the appropriate amount for AI spending. So they capped it, perhaps in order to "send a signal" that is eagerly picked up by the AI boosters.
There is no signal. Uber does not work any better since AI. They still want to promote AI, so they chose the highest number that doesn't bankrupt them so the press and AI promoters pick it up as the new price anchor.
Probably they'll quietly reduce the number more soon.
no....the fact that you could buy a reasonably prices MAC or AMD395+ thats AI tool pricing; it loads a big enough model and spits out tokens just fast enough that you can read what it's doing and comprehend it instead of magic.
That's the most useful signal. Pre OpenAI mafia RAM pricing, that comes out to $250/month.
1) This happened because they fundementally misunderstand how to use AI and how AI is priced
2) Most organizations are throwing everything in for analyses and not limiting the answer they want. You need to be specific of about what you analyze and what answers you want
3) People undervalue prompting or templated responses. I will have written. validated and sanity checked a prompt several times and run it across several models before I say its ready for use. But when it is, I know what it will give me and that the scope of its research and answer is as close to what I want as it can be. As little excess as I can. This all saves tokens
If you estimate 10k salary per engineer that means the moment it’s cheaper for them to hire another engineer but that doesn’t mean it’s improving productivity 15% but if 15% is the moment it stopped being better than another human we can assume 7.5%?
Probably even less because you would spend those 1500 extra per employee also if you just save 10% so 150 per employee that’s 1.5% on salary.
This is imho one of the best ranges we can assume for now how much would that be on the whole swe market?
Seems odd limit, especially since it highly dependant on Token provider used, with Opus this is not much and could easily be burnt in a week or less, but with something like deepseek the 1500 can literarily be an annual budget.
That being said, I do have to wonder why someone as bug as say Uber, simply not rollout OSS model in the cloud for their team, I'd imagine that would be cheapest & most flexible option, while also keeping all the data shared with LLM private.
It’s not just about the model but also setting up the system to create and share compute (GPUs) which is quite complicated on its own. Ubers primary business focus isn’t infrastructure.
I have strong conviction that companies will now choose tech stack/programming languages based on 'tokenomics'. I am vibe coding using Clojure, a language I can read but cannot write and I never hit the usage limits even when using the latest model on Claude. I have similar experience with F#, which is a bit more verbose than clojure but absolutely beats every OOP language, Python, Typescript etc.
The reason, I use F# & Clojure is they hit JVM and CLR, two popular enterprise stacks.
In my not so humble opinion Lisp(Clojure) still remains the language of AI.
It's interesting to me how ineffective LLMs are at refactoring, but when you think closely about how they work, it makes sense.
They are good at searching for things that have been done 10,000 times before, and slightly changing them. This is the majority of all "new" features.
Almost nothing is "new"...
Refactors are not this. If you can't just write a gsub to do the work, they need to essentially break it up into N problems to solve, each of them pretty slow and expensive. Sure, none of these problems individually are "new" - which is why they can do it. But they can't do it as effectively as you'd think.
Exactly my experience. I always refactor first myself then delegate boring tasks to AI. It saves me energy, time and also tokens. If code is not prepared for easy implementation agents always fail.
Good point about the unit of consumption shifting from prompts to agent loops. That makes pricing even trickier for vertical-specific AI tools.
We see this firsthand building AI Workdeck (open-source AI workspace for legal teams). A single due diligence review might chain 20+ agent calls: OCR -> text extraction -> clause classification -> risk scoring -> evidence chain assembly. The user sees one action, but the backend burns through significant inference.
The interesting thing about vertical tools is the pricing model can be fundamentally different. Horizontal tools charge per seat or per token. But in legal, the value is in the document, not the seat. A lawyer reviewing a 500-page M&A file gets way more value than one reviewing a 2-page NDA.
Self-hosting changes the calculus too. Our users run on their own infra, so the AI cost is whatever their GPU costs. That makes $1,500/month caps less relevant and throughput optimization more important.
Do we know that AI providers are going to keep these per-token prices, or eventually lower them because of competition from China?
Many lower-budget individuals are now moving to China open weight models like DeepSeek. I wonder if China's really subsidising the providers, or if inferencing costs are actually much lower, and Anthropic/OpenAI are just making sure no money's left on the table for their eventual IPOs.
So you have on one end the token revenue trending down, on the other end the training cost going up for the next frontier models, and you need to pay back your 10y debt.
0: https://youtu.be/wGZboZcSGDY?is=64GuKyqBh_4aSjTE
I think its only accounting depreciation.
I have been using my laptop for a decade, what is stopping datacenters from using the purchased GPU chips for a decade?
I wonder how often the Agent actually follows the guidance. I do see them follow it when I look. But it doesn't seem so every time.
The LLM can easily do this type of stuff, just tell it and it'll happily do it. This is exactly what I mean when I tell people they need to work closer with the AI, tell it how to do things. Don't just tell it what to do and get frustrated when it does it differently than you would.
A good way to achieve this without writing huge prompts is tell it to plan the change first. Just give it some vague low-effort directions. It'll usually get most things right, you tell it what you want different and once you're happy you tell it to go ahead.
Claude 100% of the time even thinks we use laravel despite the project being some old lumen codebase, so most of laravels features are not available. It also gets the PHP version we are using wrong 100% of the time.
I genuinely do not know how prices can get lower from the current major providers in NA without the whole market collapsing. Everyone is spending copious amounts of money to presumably make more money back.
The biggest reason large models are un-attainable for local applications is the lack hardware with large amount of unified/graphics memory (and the cost of the platforms that do). Once the memory slog goes back to normal and hardware manufacturers adapt to demand, we may see consumer hardware with large memory capacity effectively opening the door for slow but usable frontier model inference (assuming improvements in model efficiency and compute capacity)
At that point, inference becomes a race to the bottom. The large labs hope they can attain a leap in capability (which is increasingly looking bleak, with a average catch-up of just a few months) or market dominance through integration (integration in platforms and OS, exclusive deals with companies or governments).
For coding agents, i suspect no player will manage lock in enough market to enforce pricing much higher than the true inference cost, and catering to programmers becomes an unsustainable proposition. We will instead be further hit with a lot of AI integrated into our other tooling costs, such as GitHub, Microsoft suite, G-suite, forcing in AI functions as a value-ad into the total cost without giving the option to exclude them. (using their market position)
I'm not sure about OpenRouter but I wouldn't be surprised if they offer a US-based provider of DeepSeek.
For reference, Cursor has their first own light fork of Kimi that they use as their baseline coding and review model.
Not everyone using AI is using it to code core value IP.
everyone making comparisons to the dotcom bubble seems misguided. this is clearly computing 2.0 imo
I have my concerns with current inference pricing in that there's a non-zero possibility for a rug pull in the future for the subscription plans for organizations and individuals that can still use them. For now, its only companies larger than ~150 users that need to pay per token, but what if that wasn't the case? Not every company can afford over $1k/month/employee to give them access to AI tooling, further making it harder to compete against the behemoths. If we get to a point where an individual can no longer pay $100/month for nearly unlimited usage and instead must pay per token, that's going to be a problem.
Personal computing eventually became an equalizer (until we started centralizing on mainframes again, aka the cloud) because it got cheap. My hope is that inference also gets just as, if not cheaper.
I have high hopes for local AI and open weight models and we will continue the ethos of local, personal computing and not needing to offload everything to OpenAI/Anthropic/Google, etc. to get work done once the hardware and hardware availability catch up.
You update it for them every 3/4 years (if they're lucky).
It probably makes a bit more sense to compare it to existing software subscriptions like Office, or the old-school 'per-seat' licenses per user for software.
1) Don't ask LLMs for big changes
2) Review everything and point them in the right direction
Large models still suck at big changes, they produce questionable architecture and you still have to review the code, if your project is serious enough.
The codebase quickly become a mess, if you don't pay enough attention. Does not matter which model.
So why bother with big models, when flash models are 10x cheaper and much faster to iterate under guidance? Large models can be used for security and bug audits. Flash models work almost the same for changes under 300 LOC when you dictate how you want your code to look.
But this overlooks the other critical part of getting the most out of these things: the harness. I run an autonomous plan/design/code/build/test pipeline with agents using my own orchestrator. Different models are better at different stages, and I use LLMs to judge the output between them. Not everything needs Opus 4.8.
The harness provides both the scaffolding to get the right things into the model, and the right things out. But it also lets you dictate which model does which work.
It's the pipeline, not the model, that gets you quality at a given token budget.
Small models are fine for small coding tasks but I don't see why big ones can't be broken down most of the time.
This sounds like something a harness could do (and might already be doing), with work delegated to subagents running on lower-cost models.
> Review everything and point them in the right direction
Sorry upper management doesn't care. That's an engineering problem that you need to solve.
Maybe Microsoft and Nvidia are on to something.
128 GB machines that can run local LLMs are a bargain even if priced $5-8k. Yes, tok/s is not quite there, but that's probably OK since the bottleneck really isn't the code; it's WTF did Uber build with all of that spend? How did it meaningfully impact their revenue in a positive direction?
I find anything below 50 tps or so entirely unusable...
Regardless its Apples to oranges anyway, inference is quite cheap for open weight models its just that Claude and OpenAI can charge very high margins compared to e.g. DeepSeek or various provider on OpenRouter since open models are a commodity.
You could probably reach the former figure on a prosumer platform but only for very special workloads. If you spend a lot of time on prefill (which is common for agentic workloads) the outlook is even worse since that's a significant constraint for any on-prem AI.
Using local hardware is expensive when it's running a complicated software stack that can break in 10,000 different ways.
These eventual local AI servers will just talk some protocol for AI and sit in the corner and nobody will think about them.
I guess they still might need access to various systems, so idk. Eventually I think someone will offer "AI in a box" though, running the latest open model or whatever.
“AI in a box” sounds a heck of a lot like “the box” from the Silicon Valley TV show. Or the Google search appliance. Or name any other on-premise thing that is equally dinosauric.
The real finding of this article is that AI tokens are direct competitors with offshoring. $1,500/month buys you a whole employee in India.
And this is before AI companies inevitably increase pricing after the conclusion of the growth phase.
For customer facing, production software, its worth paying a cloud tax to get the reliability guarantee. For tools that are used by engineers for code development, there is no need for such bulletproof guarantees.
Yeah, I bet all labs releasing SOTA models are more than happy to remove the main way they make money and let you run it locally, especially if you're a big spender like Uber who seems very willing to throw money into the sea as an experiment.
Anthropic and OpenAI license to the public clouds. Google reportedly licenses to Apple. licensing to Fortune 100 companies running on their own infra is an obvious next step
it is a race to the bottom and I’m not sure the labs win that race. we’ll see!
If the large, well founded IT companies in the world believes the current AI cost is to high, then Anthropic, OpenAI and CoPilot have no actual customer base. AI is then relegated to very profitable niche business, but that can't fund the R&D for the models.
Also, I don't believe you need to spend $1500 a month on a coding agent if you optimize usage at all.
You can ask the same for the median 330k salary in the US for Uber Engineering... and being a bit snarky, attending Uber engineers talks here and there at a few conferences, looks like. they love to (re)invent internal tooling/platforms. That's pretty expensive on its own.
EDIT: I'm not saying that Uber's engineers didn't add value to the company, they absolutely did and handling the scale up they had to handle is not an easy feat. But I do challenge the notion of "what features did they create with that (LLM) spending?" of GP.
People DO.
It's well known that most tech companies are ran incompetently. As you say, it's not the engineers' fault.
But most projects and hiring in these companies exists to juice promotion criteria. And that, depending on perspective, these companies are either massively overstaffed or massively underproductive.
The comparison to AI spending being wasteful holds up pretty well, these are companies that readily piss away billions in pointless spending.
The idea of "if you add intelligence you make more money" is contradicted by the fact companies don't just always hire more people. Wy doesn't google just hire everyone?
I don't know; I'm a Ron Popeil "set it and forget it" kind of guy. Make the dumbest, simplest thing that's going to work with some clear path for scaling. Then go do valuable things instead.
But in Uber's case, they tend to reinvent lower level pieces of platform/infra.
Even then it makes more sense to rent the bigger GPU and get your answer faster.
I get that if it's offline the security downside of XP doesnt matter, and I assume XP is free, but being free doesnt really seem that valuable compared to alternatives (free linux and virtually free OS if buying wholesale).
I suspect there’s some mass delusion with respect to actual accomplishments as a result of LLM use. Sure, things are moving faster, but does it matter?
I have still found the sweet spot for me is using LLMs but I am still in the drivers seat.
Normal people have to produce something of value from that spend. So starting 100 agents and then waking up to something cool but useless just means you spent a few thousand dollars and created nothing of value............
Hard tasks require a lot of guidance and code reviewing, unless you are creating another throw away project where correctness, maintainability and code understanding does not matter.
WTF did anyone build with all that spend? Despite all the feel-good anecdotes about how productive folks feel using ai coding tools there's a deafening silence when it comes to actual, demonstrated efficacy. How can we be this far entrenched in these workflows and still not know whether they actually do anything useful?
What would previously be janky internal dashboards or excel sheets are now actually nice to use tools. That said of course the maintenance cost of all that has yet to be discovered, and the ROI is questionable.
OK. I guess that's good, too.
I don't think this would have been possible without having solid engineering culture and processes in place before bringing in ai coding tools.
And I don't want to sugarcoat it, this hasn't been easy, requires continued discipline, and took well over a year to get good at. And we still have to continuously learn, experiment and adapt our training, tooling, and processes.
Software engineer quality of life.
There can be an increase in productivity without a corresponding increase in total output. The gains could be captured by software engineers doing a days work in an hour then fucking off in a variety of ways.
Until companies start hiring 5x less engineers than they did before and well.. we are clearly moving towards that direction
as for building actually complex software, the art of that is not in simply chaining together such scripts. Its the art of using architecture and testing to shape uncertainty, and developing requirements (and extrapolating sensibly from incomplete requirements). I don't think llms are great at this, but they arent terrible either. A lot of the more active users in the space are doing stuff where theyve realised they need more detailed specs, which like, yeah, we knew this already - better defined problems lead to better software.
Coding faster doesn't really solve that.
Uber makes more money if people buy more rides, order more food, have some breakthrough in autonomous driving. They can save money if they can optimize some ops or spend somewhere. Is there any evidence that with the spend on AI that they achieved any of this? If they did, I'm sure we'd hear about it in some engineering blog.
Uber engineers do not define their revenue stream; the product leadership team does.
$1500/mo of AI spend by engineers does not equate to revenue. They need to figure out revenue first before zeroing in on AI spend.
Claude has allowed me to do refactors that would have taken weeks to instead take a couple of days. It has, objectively, increased the velocity of the engineering component of greenfield features by 40% in my org. You can put a number value on that and decide if it gives you favorable ROI.
I effectively get to operate at the rate of a small team of engineers - I know that because I've managed small teams of engineers in the past.
I think this is the part I struggle with. The code I write makes me money or is a way of teaching me something, both of which are reasons that I would write the code regardless.
I don’t think I have any projects in mind that I’d be willing to spend half of a car on that I also wouldn’t have written myself.
Obviously just a personal take though. I’m glad you get the usage you want out of it.
For example, what if you're a tiny startup and you're considering whether to hire an extra engineer or do all the coding yourself. I would estimate that AI is worth far more than $18,000 a year in that situation where you might reasonably decide to put off hiring an engineer.
Anthropic: https://support.claude.com/en/articles/12883420-view-usage-a...
OpenAI: https://help.openai.com/en/articles/10875114-workspace-analy...
higher ups pushed for these last 2 years to be AI focused so I don't think this restriction is a measure of "don't use too much AI" as much as it is a measure of "don't use only 'manual' AI tooling" since we had a dozen more specialized tools in-house running locally or otherwise that didn't count towards the budget
Your other plans are fixed price with rate limits where you get more tokens than the dollar equivalent you pay monthly. These plans are economical only if majority of users spend less tokens in $ than the plan's costs. This subsidizes the gap vs. power users who spend multiple k$ monthly in API tokens.
as far as we know there's no evidence that they can produce any profits at all
The open-weight models will have a steady race to the bottom on inference costs just by dint of competition between providers. They aren’t at the frontier yet, but they are rapidly eating the flash market.
They can't say that $0 per employee is the appropriate amount for AI spending. So they capped it, perhaps in order to "send a signal" that is eagerly picked up by the AI boosters.
There is no signal. Uber does not work any better since AI. They still want to promote AI, so they chose the highest number that doesn't bankrupt them so the press and AI promoters pick it up as the new price anchor.
Probably they'll quietly reduce the number more soon.
That's the most useful signal. Pre OpenAI mafia RAM pricing, that comes out to $250/month.
Probably even less because you would spend those 1500 extra per employee also if you just save 10% so 150 per employee that’s 1.5% on salary.
This is imho one of the best ranges we can assume for now how much would that be on the whole swe market?
That being said, I do have to wonder why someone as bug as say Uber, simply not rollout OSS model in the cloud for their team, I'd imagine that would be cheapest & most flexible option, while also keeping all the data shared with LLM private.
Uber’s COO says it’s getting harder to justify money spent on tokenmaxxing
https://news.ycombinator.com/item?id=48268871
Uber torches 2026 AI budget on Claude Code in four months
https://news.ycombinator.com/item?id=47976415
Corporate America Is Starting to Ration AI as Cost Skyrockets
https://news.ycombinator.com/item?id=48335388
The reason, I use F# & Clojure is they hit JVM and CLR, two popular enterprise stacks.
In my not so humble opinion Lisp(Clojure) still remains the language of AI.
They are good at searching for things that have been done 10,000 times before, and slightly changing them. This is the majority of all "new" features.
Almost nothing is "new"...
Refactors are not this. If you can't just write a gsub to do the work, they need to essentially break it up into N problems to solve, each of them pretty slow and expensive. Sure, none of these problems individually are "new" - which is why they can do it. But they can't do it as effectively as you'd think.
We see this firsthand building AI Workdeck (open-source AI workspace for legal teams). A single due diligence review might chain 20+ agent calls: OCR -> text extraction -> clause classification -> risk scoring -> evidence chain assembly. The user sees one action, but the backend burns through significant inference.
The interesting thing about vertical tools is the pricing model can be fundamentally different. Horizontal tools charge per seat or per token. But in legal, the value is in the document, not the seat. A lawyer reviewing a 500-page M&A file gets way more value than one reviewing a 2-page NDA.
Self-hosting changes the calculus too. Our users run on their own infra, so the AI cost is whatever their GPU costs. That makes $1,500/month caps less relevant and throughput optimization more important.