There's definitely a way to use Claude code that is token conscious.
I've tried throwing unsupervised agentic software factory workflows against the wall, and they burned through my tokens like nobody's business but didn't produce much.
Supervised, human-in-the-loop process on the other hand is much more productive but doesn't consume nearly as much. Maybe that's why everyone's pushing agentic approaches so much.
Yeah. Claude does good work but reviewing it all properly takes quite a bit of time. It got to the point I started having trouble maxing out my weekly allocation.
Dealt with that by going all out and making an agentic parallel code review skill. Basically an infinite TODO list generator. Now I'm definitely getting 100% of the usage I paid for. It really burns tokens like nobody's business, and catches a lot of issues while at it. I've been looping this review/fix process every week. It's dramatically reduced the amount of stuff I need to pay attention to during my human review sessions.
I really don't like how the payment plans work with the providers right now. I feel this pressure to use all my tokens for the week, often just "wasting" them. But also, I want to take advantaged of the subsidized tokens in Claude Code and Codex for as long as I can.
There is this real danger that our thinking, and the things we make, become bloated without constraints.
IMO software has gone to shit since both mobile phones and laptops mostly have massive amounts of compute. We always seem to use it to the limit, just because it's there.
It's the gacha of software development. We've got periodically resetting timers. Prompting is like booster packs: we have a finite number of dice rolls before the timer resets. We might even get a Legendary Ultra Rare pull whenever Claude happens to be feeling extra motivated. Before you even know what's happening, it's hijacked the brain's reward circuits to the point you're waking up at 3 AM because that's when the timer resets. Gotta saturate those timers with pulls and minmax everything in sight.
At least it's doing something productive instead of just sinking money into literal gambling simulators. Mercifully, unlike video games, automation is not "cheating".
Any corp (> 150 seats) has to use API pricing, so e.g. I don't have pressure or weekly limits, just a set budget I can use each month.
That will be true in March 2027, but isn’t necessarily true today. I work in a large organization that is grandfathered into the Enterprise Subscription plan until then. We have thousands of seats using Claude Code.
Oh, I didn't know about that. We got into Claude Code in AFAIR December, and we were on API pricing from the beginning.
I’m interested in how this works in practise - I guess you’ve written a skill to do code review, then your Claude.md file tells it to use it after every change as a bg task? So does this work as a background task while Claude is working on the next ‘feature’?
I just committed the skill to my dotfiles repository.
https://github.com/matheusmoreira/.files/tree/master/~/.clau...
There are many "critics", one for each quality I want reviewed. Correctness, consistency, maintainability, security, testing... Everything I could think of, and I keep adding more.
https://github.com/matheusmoreira/.files/tree/master/~/.clau...
The scrutinize skill is the entry point. The Opus I'm talking to becomes an agent coordinator. He explores and autodiscovers the project's structure, subdivides it into logical sections.
Then he runs a truly absurd critic x section matrix against the entire project. Literally hundreds of these agents running in parallel, each focusing on one area. Ten minutes of this is enough to exhaust my Max 5x five hour window and put a serious dent in the weekly usage numbers.
It literally takes days to run a full agent sweep. I designed it around the rate limiting. The agents do file system style journaling in order to resume cleanly. They commit all of their findings as they go into an orphan branch in the repository. Further review runs can build on it and avoid searching for known issues.
The way it works in practice is I just run /scrutinize sweep and then go work on something else, or just go do my actual job, live my life, play video games, write an article for my blog or something. Come back five hours later to either resume the process or check the literally hundreds of issues that have been found by all the agents. Then Claude and myself will go in and evaluate and fix all of those issues one by one. Then review again. Then evaluate/fix again. I'm just gonna keep looping this over and over until zero issues are found. For all of my projects.
Going from solo hobbyist programmer to this was pretty insane. I can only imagine what these corporations with infinite money must be doing.
Those critic skills are great. I see a real business opportunity for someone who can bundle everything you describe into a turnkey solution for a programmer like me who doesn't want to take months coming up with their own system and extensive .md files.
I'm currently, very painfully, removing a tiny bit of tech debt at a time from a massively complex project that we inherited from a 3rd-party vendor. Some of the tech debt is AI-related, some because it's a vendor who rarely has to maintain anything they create, some because when we first inherited it we had no grasp on the entire codebase and were just trying to change the plane wheels while flying (we still are).
What I'm doing now is the hardest kind of programming imo. I spend hours/week just meditating on how to chip away at this out-of-control codebase, figuring out how I can surgically remove some leaky abstraction that's spawned 5 cousins w/o disrupting the whole project. I'd be fascinated to see if the latest frontier model with a system like yours can actually help me. But I don't have the time or desire to invest the months of trial and error that I'm sure it took you to get to that point.
I used Claude Code's /insights function. It gets Claude himself to go over your sessions and usage patterns. It'll produce an HTML report that you can view.
In my case Claude saw that code review was my main activity and that I was manually and repeatedly asking claude to "review X, Y, Z..." so he suggested turning it into a skill. So I fired up the superpowers:brainstorming skill and bikeshedded it until I ended up with this heavy duty massively parallel super reviewing super claude. Refined it a bit after a couple weeks of use and the result is what you see in my repository.
Thank you for sharing!
Thanks for sharing that, I am looking to improve my agentic use, and it will be useful as I develop my own path.
You're welcome. Email me when you discover your own path? I could learn a thing or two as well. I'm pretty much a beginner when it comes to this stuff. Subscribed like two months ago.
I would love to see the codebase once you reach the zero issues point.
I would advise against it, depending on the project.
My lone lisp project gets the most love. I spend weeks reading, reviewing, restructuring and rewriting everything. It's the project where I'm concentrating all my efforts. Everything I push to master is absolutely my own work and I do want everyone to read it.
I had no trouble letting Claude take over maintenance of my static site generator and virtual machine orchestration scripts though. I wanted to care but... I didn't. I did glance over the finished product just to ensure it wasn't going to nuke my laptop the second it ran, but that's pretty much the extent of it.
I did the same thing - task oriented work, each task a md file. I have a harness based on it: https://github.com/horiacristescu/claude-playbook-plugin
The current thinking is automated agents is what turns this from an industry in the tens of billions to a multi trillion dollar one. So yes you are right on the money, agents stimulate demand for this thing they've built.
"The bureaucracy is expanding to meet the needs of the expanding bureaucracy"
AI is expanding to meet the needs of expanding AI. Why worry about jobs? AI will provide plenty of work. If anything, I worry we'll be working more, not less. All that AI will need someone to vouch for it and to scapegoat when it makes mistakes.
I didn't know that one. Loosely said to be Oscar Wilde.
Delivered in the voice of Lenard Lemoy to millions of GenX during their formative years.
Is that a civ 4 reference? You sir, have my upvote.
[flagged]
There is always a quantity of lubricant that can get any machine moving. Just add so much that you create an all consuming river of lube and watch your thing sail away.
Good then that Amazon sells it by the 55 gal drum then.
https://www.amazon.com/Passion-Lubes-Natural-Water-Based-Lub...
> This product is out of stock
Ah, shoot, there go my weekend plans. Bummer.
I think it's great. People at a broad scale are getting first hand experience with resource management. It's a fairly cheap way of doing it too (in contrast to: learning this by managing humans) and we can all benefit from the skill transfer.
I find myself observing how my lead manages meetings ... "ah, this is like when I do that with Claude", "this is where he wants to understand what happened, like when I ask Claude" ... it's funny.
At the enterprise level though, its going to be hard to want to use a service in which costs are not predictable, and keeping those costs under control requires employee training.
>...use a service in which costs are not predictable, and keeping those costs under control requires employee training.
Isn't this a (mildly exaggerated) description of AWS, which is a very successful service?
Mmm… but for AWS its pay for external use right?
So your costs scale with the number of users you have.
Thats an op ex that you can explain.
For tokens for developers its maybe closer, cost/outcome wise, to hiring an external consulting company to write your code; money paid scales with work done, no promise of delivery, arbitrary unpredictable external price changes.
Its not quite the same; though, similarly lucrative for consultants.
>Mmm… but for AWS its pay for external use right?
Not if you're using it for running builds, running research jobs, model training, etc.
You can put a limit on token spend and provide training (and even pre-configured workflows) on how to limit token spend.
Like the other commenter said: cloud spend can also spin out of control if you don't pay attention, yet we've found ways to keep it under control (training, guardrails, limits, transparancy).
The problem that I see is what you do if someone runs out of tokens. It doesn't very well work to say "well I guess you just get fired because you can't work at full speed for the rest of the month".
Personally, this feels like its just trying to push the work of managers in allocating resources onto developers so that they have more work to do and can be blamed if anything goes wrong.
[dead]
Am I losing my mind, aren't there multiple headlines each day about companies penalizing employees for not using AI enough?
That was roughly 3 weeks ago, with the reprising of Claude 4.7 and GPT 5.5, things have become more spicy.
2 months ago: no limits. 1 month ago we had a leaderboard for whoever had the highest token spend not taking into account what was actually produced. This week: “everyone is using opus too much, just use it for planning.”
use AI, don't use AI, this whole thing is getting really hard to follow
i've worked at so many places where the propaganda/marketing and reality on the ground is so disorienting/shocking i don't really expect this to be any different...
It is allmost as if humans ain't of a single mind.
since those headlines started ive felt it just encouraged inefficiency. "say as much as you can without saying anything." if you were accomplishing your task the need for more would end, thus there is incentive to never succeed.
To be fair, the cost of software development has always been fairly unpredictable. What may be different is that the cost used to be roughly proportional to man-hours spent, while now the number of agents running in parallel may be less predictable.
The cost per month is 100% known and always has been. What has been variable is the rate of delivery. AI is different and can be substantial in countries with lower wages.
> To be fair, the cost of software development has always been fairly unpredictable.
Yes, but in a "oops this is gonna take another two months to finish" kind of way, not the "oops this is the 12th time this month 8 developers have burned $2K in tokens in a single day and no one really knows how it happened" kind of way.
We’re all being given belt-loaded machine guns and tossed on to Planet K. We used to pay for the salaries of soldiers, now we have an Ammo Budget.
A belt loaded spinwheel machine gun, where there are some chances the next bullet is a dummy round, or goes in the wrong direction. And everytime you reload a new soldier is in charge of the gun
You don't need that analogy as the normal use of a automatic gun in war is not to kill, it is to suppress - stop the enemy from moving. If you are hit by a gun in automatic mode it is your own stupid fault. When you want to kill someone you switch to one shot or maybe 3 round bursts.
TIL. I know literally nothing about automatic guns
The only important thing to know is they're loud and thus impossible to conceal. Maybe someone out in the open would be hit if one starts firing unexpectedly. However, the vast majority of cases, the purpose of an automatic gun in war is just to put oatness stream of bullets so that everyone on the other side knows it's stupid to show your head and this allows your own people to move in relative safety because they know where your automatic weapon isn't going to be firing and the enemy, of course, isn't going to be able to respond.
There's no fucking training to mitigate a slot machine.
that analogy is so boring now with so many real world examples of actual LLM work.
people still can't get over the unreasonable effectiveness of algorithms.
There have also been winners of a slot machine gamba, so the analogy quite holds. I would even argue that there are considerably more slot machine gamba winners than the real world examples of actual LLM work.
nondeterminism will always be anathema to the engineering mind
Odd, I train teams (at large companies) to use harnesses effectively. So some training does exist.
I get the anti/skeptic sentiment. I've been called a lot of horrible things by a vocal contingent when they hear that I help train folks to learn software engineering best practices and then apply AI to that.
There’s actually been a ton of research on how to optimize “slot machines,” at least in a generalized sense. For more reading, check out the literature on multi armed bandits.
Games like Diablo are basically a whole bunch of slot machines, and there are strategies you can follow to optimize your run.
Yes, because in video games there is always a chance to win so you can optimize your strategy around that chance. If you have a 1% chance to drop a legendary weapon, the question becomes how do I manufacture 100 chances for a weapon drop in the shortest possible time. With agentic coding there is no such guaranteed chance - in a way it's worse than a slot machine that is guaranteed to pay out eventually. You could spend hundreds of millions of tokens and still not get what you asked for.
You’re right, the arpg analogy isnt great, it’s too simplistic. I was trying to come up with something heavily stochastic where people are coming up with strategies to get the odds in their favor. Maybe closer to speculating on the real estate market? But even that feels too simplistic compared to LLMs. Even the definition of a win isn’t well defined.
Actually it’s really its own thing, I don’t think the slot machine analogy works too well, you also have fixed odds (and you know they aren’t in your favor), and a binary output
The analogy to slot machine is that you're spending your own resources in hope of a reward. So you're ultimately bound by your resources and your strategy doesn't count for much in the grand scheme of things.
With employees, there's a lot of punishments in place for people to not want to mess up. Loss of wages and reputation, prison time,... Startup do not fail because they have a bug-ridden product, they fail because of the market.
With AI, all bets are off. They're not aligned with your goals and it's very hard to discern when they go off unless you're an expert. And if you are one, at best it's just a slight boost in typing especially with all the works involved in software development.
> If you have a 1% chance to drop a legendary weapon, the question becomes how do I manufacture 100 chances for a weapon drop in the shortest possible time.
Sidenote but I hope everyone realizes that 100 is kind of arbitrary here and does not mean the total chance to to get something is 100%.
you don't have to do the math unless it's on the exam, lol.
LOL, that's a sophisticated and sometimes slightly unpredictable multitool.
If this is the "analogy" you go for, you don't seem to be suited to make that comparison.
> There's definitely a way to use Claude code that is token conscious.
Colleague used Sonnet 4.6 on some pretty normal agentic coding tasks through AWS Bedrock to keep the data in the EU, 100 EUR usage in a single day. In comparison, the Mistral subscription costs about 20 EUR per month and we tested that for similar tasks it was okay, the usage got to around 10% of that monthly limit in a single day. Or Anthropic's own Max (5x) plan where you get way, way more tokens to do with as you please.
I feel like the sweet spot is having a monthly subscription with any of the providers (you're subsidized a bunch), but if you have to pay per tokens, now I'd just look in the direction of what tasks DeepSeek would be okay for, sadly probably not in the situation above. For a startup, though...
On the other hand, this feels a bit hypocritical:
> It was part of an effort to get project managers, designers, and other employees to experiment with coding for the first time, and sources tell me that Claude Code has proved very popular inside Microsoft over the past six months.
They're gonna say that the future is all AI... until they get the bill.
I was a Mistral Le Chat Pro subscriber (the €20/month plan). Yesterday I hit my monthly limit. Switching to PAYG I burned through another €40 in one evening, working on the same project, with the same tasks.
I upgraded my plan last night to Mistral Le Chat Teams. This now costs me €60 per month for two users. Limits have been reset, but I have no idea now if my per seat limit is higher than the Pro plan, or if the limit is shared between the seats, it’s really not clear. I guess I will find out next month. The limits reset on the first of the month and I really hope I don’t hit them in the next seven days.
I use Mistral Vibe CLI and I’ve written and implemented a couple of new skills[1]. Caveman, based on an idea I found online somewhere, this skill removes all extraneous response text, including articles. Makes for some fun reading, but supposedly reduces output tokens significantly. Hash-anchors, this one is based on a concept from Dirac[2], reduces search failures and also includes multi-file dispatch. It will be hard to measure, but Vibe tells me these two should result in roughly a 40% reduction in token burn.
[1] https://codeberg.org/MimosaDev/skills
[2] https://dirac.run/
I was trying to get a better sense of the time cost quality matrix of these, so I threw together a quick eval of Sonnet 4.6, Mistral's dev model, and Opus 4.7 (figuring it's what you'd use if you were on Max).
The results for a function implementation and test of levenshtein distance in js are pretty similar but Mistral is 30x cheaper than Opus 4.7 and 4x faster than Sonnet 4.6.
https://5m6qnuhyde.evvl.io/
But that's not very informative.
Levenshtein distance is not only a well-understood problem, it's small, self-contained, and extremely well-represented in the training data. The kind of problem where even small/bad models can excel. The golden standard for those tasks is just "use a library" so no wonder the beefy models are expensive: you're chartering a commercial airplane to go grocery shopping.
My personal benchmarks are software engineering tasks (ideally spanning multiple packages in a monorepo) composed of many small decisions that, compounded, make or break the implementation and long-term maintainability.
There's where even frontier models struggle, which makes comparisons meaningful.
>> many small decisions
It’s making guesses not decisions, framing as decisions will lead you astray to wasted time and tokens.
It’s vaguely productive to tell them a ton of relevant info upfront attempting to minimise their need for load bearing guesses. I say vaguely because obedience is generally only around the level where it's good enough to lull you into a false sense of security, not to actually be obedient.
It’s a bit more productive to use the various loop mechanisms (hooks, /goal etc) to evaluate each end of turn against guard rails and reject with clear instruction on whats unacceptable. Obviously if you only do this without the front load of info then you’re likely to spend more tokens to reach a satisfactory end of iteration.
If I perfectly know all the guardrails I need, I don't need an LLM, only Prolog.
While you are correct that something like Antigravity 2 + Opus 4.6 can handle large scale software engineering tasks, I would argue that it is usually (but not always) better "coding agent hygiene" to work on smaller code modules and as the human in the loop be a partner, not someone who prompts and then disengages.
Breaking code up into composable chunks has worked well for me over 50+ years as a professional software developer, and I can't get away from the idea that it is still usually the way to go using agentic coding tools.
The one detail I did forget to mention is that if anyone goes with the Mistral subscription (instead of paying per-token), then the Mistral Vibe tool gives you their Medium 3.5 model by default, with a 200k token context. It will probably be enough for plenty of tasks, though there's also a noticeable difference between that and up to 1M.
> They're gonna say that the future is all AI... until they get the bill.
I mean, the will continue to say so, they just want to be the ones being paid for the service, not anthropic :)
My experience as well... I've only hit Antrhopic's 5hr threshold a few times, and two of them was within a half hour of the window. Also, all three times I'd already accomplished a LOT.
I tend to work with the agent, and observe what's going on as well as review/test and work through results/changes. I spend a lot more time planning tasks/features than the execution, even using the agent as part of planning and pre-documentation. It works really well. I don't think people burning through the 5hr allotment in under an hour are actually reviewing/QC/QA the results of what they're doing in any meaningful way, and likely producing as much garbage as good (slop).
I'm really curious as to HOW the MS employees were using the agents as much as what they were doing.
I suspect subscription limits are quite a bit higher than the equivalent tokens their dollar cost could purchase. I similarly feel like I can get a lot done with a $20/mo Claude Pro subscriptions, but also can easily spend $10-20/day at API pricing with similar usage.
Yep. I get $6k - $8k worth of tokens (at api rates) using the $200 max subscription.
Can verify that I've gotten about $400 worth of tokens from my $20 sub.
Now that sounds like a business I’d like to invest in! When’s that Anthropic IPO anyway?
I don't understand why people are using the API pricing instead of the Pro/Max subscriptions? What am I missing?
Enterprise customers don't get that option. But also if you want a fully custom harness, you also don't get that option.
Personally I prefer the API pricing because I feel like I'm not going to get rug pulled on my work. When it comes to personal stuff, I use the shit out of my sub, but it's not making me money.
I’ve made the same argument On Here. Paying the full price (should!) make you consider you usage, pick the right model, delegate to cheaper/local providers, …. It makes you use the models the way they’re going to be used after the subsidy ends.
Depends on what you're optimizing for. I'd hope that "after the subsidy ends", the "cheaper/local providers" will be at the level of at least current SOTA models. If not, then there's hardly a point using them anyway; if yes, then by sticking to subscription workflow you'll be learning the very workflow you'll be using "after subsidy ends".
Either way, I don't see much point of intentional austerity in times of extreme growth. There will be time for austerity once the growth ends.
Because with Max subscriptions, you have to use the Claude Agent SDK, which is basically running Claude Code underneath. You don't get to use the chat/Messages APIs with personal subscriptions, for that you need the API pricing.
Anthropic is forcing large enterprises onto api billing instead of subscriptions.
Terms of service prohibit subscriptions for employees of companies bigger than X people. I suppose they could all sign up as individuals and try to get away with it but presumably that would look pretty obvious with a tiny bit of analytics.
> There's definitely a way to use Claude code that is token conscious.
By buying a subscription and dealing with the limits, using claude code and paying per token seems like the fast lane to the poor house.
I get 98.6% cache hits on Claude code. Short of drastic arch changes it’s hard to imagine it getting much better.
98.6% cache hits doesn't distinguish an efficient workflow from an overly chatty linear agent repeatedly reusing the same context. Plus, it says nothing directly that the process has good useful progress per token.
We are all going to be graded by (tickets closed / tokens burned) soon enough.
Sweet. I can get that up to infinity, assuming they're using IEEE-754 division.
I doubt it, the difference between someone slightly inefficient and someone extremely efficient isn't big enough to matter compared to how much they cost in salary.
You pay for cache hits on every turn and even with the newest architectures longer context is slower/more energy intensive. Constructing concise turns that reuse prefix and stop when the new context is no longer useful help, as does pushing generation down into cheaper models while using stronger models for verification.
yeah, by using codex
---- Before it was:
Me: We need to do this this that.
Claude: <random stuff that approximates human outout>
Me: Are you sure?
Claude: Well actually there is a bug <more random stuff that looks right this time>
----- Now it is:
Me: We need to do this this that.
Claude: <random stuff that approximates human outout>
Claude: Let me consult the advisor on that.
Claude: advisor came up with some advice, adjusting according to that. <more random stuff that looks right this time>