The other part of that is that while price per token may be going down, tokens per task is going up

For better performance of ~equivalent tasks. That's what all the harness tooling people are using does: (often) increasing output quality by significantly increasing token counts.

Price per M/tokens is also a fuzzy metric when newer models reason longer, and then burn more tokens while doing so.

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...

If Anthropic are then they are making a big mistake, their token hungry Claude code is far too greedy

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? This sounds like something a harness could do (and might already be doing), with work delegated to subagents running on lower-cost models.

> But you can't talk to them about the flow of the code. You can't ask them for their thinking as to why certain things are. There are plenty of valid criticisms or warnings about over-reliance on AI coding, but this is not one of them. Today, I am using a semi-autonomous agentic coding system which has an `interview` functionality built in - when it spits out the PR from the input, if you have questions about the motivation or context for a particular choice, you can start up a clone of the original agent in a sandbox to question it. Now, you might claim that those responses aren't always reliable, accurate, or consistent, and that claim has a little more weight (though, in my experience, decreasingly so) - but it is _certainly_ not the case that you cannot interview an agent about choices made. I'm literally doing it every day.

No, but you do want Opus-tier models to do desktop and office software automation (think about people who intensely use Excel and the like). Actually those might take even more tokens that coding in a lot of cases. Why do you think Claude Cowork is successful, and why do you think Codex is leaning so hard into Computer use?

Because writing huge amounts of code is easy for humans too. Agents already proved that they can do it. But are agents able to maintain it? I do not know and unless I know for sure, I am not fully committing to AI generated code. i.e. I am able to write about 1k lines of code of "acceptable" quality per week. Which means in 1 year, there will be about 5Ok LoC. I am pretty sure, that I would have to spent like 60-80% of time to maintain 1st year code and the rest to make new features in the second year so I would have to hire more people and spent time to onboard them to maintain velocity. All of that are rough estimates, probably overoptimistic and way worse in 3rd year. Good luck doing such estimates with code agents. Even worse if you already have huge amounts of legacy code.

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.

I would expect the overwhelming majority of output tokens would not be the actual code but used for analysis, reasoning, testing and iteration. If you only use the agent for autocomplete then yes, the calculation is probably different.

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 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............

Running hundreds of agents overnight is almost certainly 99 percent waste.

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.

> The point of a refactor is for you to think deeply about the code you are responsibility for, so you can make it better (faster, easier to work on, more tests, whatever). I'm pretty pessimistic on AI and don't have access to good agentic workflows, but refactors are exactly the thing where it seems to me like agents could be really strong - once I've refactored something architecturally, I might have hundreds of instances of a thing that needs to be updated in a predictable way, but is complicated enough that it's going to be faster for me to manually update hundreds of instances rather than writing a generalizable find/replace tool.

I use the $100/mo sub but my 30 day API cost is about $1700/mo. It really depends how you use it, if you're using prompts to generate detailed designs, breaking those into lists of tasks, and then feeding those to multiple agents - it's really easy to burn through many thousands. If you're being more deliberate and using a few agents at a time interactively, having it review PRs/resolve issues, automated clean-ups and performance optimization, etc it could be more like $1500. If you're just throwing it one-off questions like a better stack-overflow that is well under a $100. I've really gotten into /goal, if you can find something verifiable and leave it overnight - it's kinda like christmas morning to see where it landed.

I think the main thing companies should try to understand is avoiding the use of 'claude -p'. I definitely have written a goal file, and then just ran claude in a loop over the goal in order to 'token max'... why not? I'm doing research and have some clear KPIs where research into all kinds of techniques / tuning can improve the results. I can spend my budget on a "experiment with blah blah blah to improve blah blah" or give it a list of things to try that I know will take awhile. Its no problem hitting hundreds of $ of API spend while sitting at a computer with 3 monitors have 6 windows of useful claude code interactive sessions, while working on 2 or 3 projects and using worktrees, and it's a little weird when you hit your limit by 2 o'clock and have to wait for token budgets to reset; god forbid, I manually edit code... which I did do for the first time in months. You can also start to generate a lot of token spend if you do something like "hey make me a stylized slide deck using internal skill / agent XYZ based on commits A through C", which as an engineer, makes presentations building much less painful . This uber limit is not high compared to the big SV companies.

Nope I'm a couple levels too far removed from the code at this point for that. Closest I get is during meta-management (modularizing, complexity reduction, etc) with agents

My favorite solution to this is to use the Cline coding agent, which is open and allows you to easily switch between different providers and models.

I use Claude every day. Often for multiple hours a day. Basically doing my job not worrying how many tokens I spend (as in too many or too few). This is a pretty complex code base (database optimizer and related). Just looked at spent for the past 30 day, didn't even come to $600. 95% of my tokens are from cache. If I were to reach even $1500 I have to let claude run unsupervised over night (and with the amount of mistakes it still makes and guidance it needs, I do not believe we are there yet.)

Days ago he said… “I'm finding that coding agents can take me from a vague idea to a working solution, one with tests and documentation and that looks like a carefully considered project evolved over the course of many weeks... in less than an hour. Even if the code is rock solid, there's a limit to how many projects like that I can sensibly care for - and if they're instantly abandoned, what value was there from creating them in the first place?” https://simonwillison.net/2026/May/31/the-solution-might-be-... Here is Simon questioning a fundamental belief held by the pro-LLM lobby. Would a paid shill question that? Simon is, without question, an enthusiastic pro-LLM person. I disagree with what he says often, the product market fit post was a bad take. But I don’t believe he is shying away from sharing his thoughts when they’re not favorable to the industry.

You must not be using coding agents. You can sneeze and spend $1 on Opus in Claude Code.

I'm on a $100 Claude Max plan, my usage is only about 50% of the plan limits, but in the last 30 days my usage was equivalent to API token spend of $1850. If you save all your Claude Code conversations, the saved files include API costs and you can calculate this yourself. One of my most expensive sessions cost me over $100 in token spend in a single evening. I'd just found out that the time tracking & invoicing SaaS I use is increasing their monthly pricing by 2.4x - so I assigned Claude Opus 4.8 to recreate the entire SaaS for myself, and load in 13 years of my historical data. I've only completed a full read-only implementation so far, with adding & editing of records still to come, but I do expect Claude will have fully recreated the entire SaaS for me at an API cost less than a single 1 year seat of continued subscription to their service. And since I'm actually on a Max plan, it didn't actually cost me $200 of tokens at all. coff i would not buy the Bending Spoons IPO coff saaspocalypse I could ramble on about where the other $1750 of usage goes, but I imagine it's similar for most heavy Claude / AI users. Interactive coding sessions, a daily personalized podcast, some automated overnight agentic "proactive" sessions, a daemon that wakes up if I send Claude an email or voicetext to check something when I'm out. I've also noticed that if Claude's tool-use goes haywire & Claude gets confused or lost, sometimes a single email reply session that would normally be just $1 of API might spiral to $12 of API while it bangs its head against trying to run a program that's in a different folder to the one it's currently in. Sometimes a simple 'pwd' would save you a lot of headache, Claude....

Uber engineers reported that loading their workspace and pulling recent commits exhausted that AI limit for Claude Code (4.8 x-high) immediately.

I don't think at $1,500 you're not forced to code on your own at all, in the sense of typing code. You're simply forced to not yolo-max twelve parallel agents at all times.

I'm curious how much of the usage comes from vibe coding vs using agents/harnesses in internal tooling

I would not be surprised if they have engineers vibecoding 2-3 projects each simultaneously, nonstop, on largely un-moderated review-suggest-iterate-test feedback loops. All the code gets summarized and fed into their manager's agent contexts, probably duplicated several times across levels and departments, with some generated back-and-forth emails pinging around the org chart, eventually generating 2-3 long-winded reports that nobody will read chock full of generated visualizations that can all get consolidated into a generated slide deck that they'll show (maybe, at some point) to a handful of humans with more money than a human brain can conceptualize to demonstrate all of the innovation they're doing. I am increasingly convinced that many of these companies are dead trees whose only function is to burn money lest it fall into the hands of the peasantry.