I recently switched off Max flat rate to Enterprise API pricing and I went from 200/mo to 10k/mo with the same usage pattern on Opus. They don’t offer flat rate to enterprises.
So Fable would cost me 20k/mo at Enterprise rates. That’s around the average cost of a loaded SWE in the USA. “But I’m >2x more productive” doesn’t justify doubling the opex of the Software/IT department for most companies when revenue isn’t even up 10%.
I switched to DeepSeek v4 Pro with OpenCode and am on track for a few hundred dollars of spend this month.
Rewriting your stack from Ruby to Go in 2 days where it would’ve taken 6 months is impressive and fun. But that isn’t upping revenue.
Iterating on net new business features and ideas that are niche that the LLM isn’t trained for are much harder. Is 20x the token cost worth it there?
I don't live in USA. I'm getting paid around $2500/month and that's good salary for developers here, plenty of folks are getting below that number.
So this pricing is just completely outside of our economics and nobody I know would pay that, no company will justify spending $20k/month when they can hire 10 more developers instead.
It is very interesting unfolding of events. Can't wrap my head around it completely.
I'll add a concrete example from a not-too-cheap-anymore EU country: Estonia.
* Average software dev salary in Q12026: 4945€ / month [1]
* Total cost for the employer: 6616.41€ [2]
For $20k/month, you'd get 2 x full time mid-level developers + 1x junior dev or QA.
So the calculation becomes: which option can produce better results for your specific use-case, "you + Fable" or "you + 2x mid-level developers + 1x QA". (and from personal experience, mid-level in Estonia = senior dev in the US, in terms of skillset and experience.. but YMMV)
(Of course that's simplified. Your full time devs need _some_ level of AI subscription as well + hardware so add a couple of hundred to their salary per month etc so you might only be able to afford 2x mid level devs, instead of 2.5)
This is a good start, but the calculation doesn't include office space and overhead (for every 100 developers there is maybe 5-10 support staff to cover the additional legal / administrative, and don't forget the extra cost in supervisor time to manage them)
Exactly, that's why I wrote that it's simplified and the actual full cost to the company depends on your company size and setup (fully remote vs in office, management heavy vs lean-flat etc). One point though, from personal experience, I'm spending an order (or two) of magnitude more time in "managing" an agent than I spend in managing employees - so that part might come out cheaper in the end for having actual employees ;)
I'm currently working for an Estonian startup and we pay quite a bit more than that. We hire remote (primarily across Europe) and our biggest issue is finding the right people. You need to consider AI can be "hired" or "fired" instantly too, so it's better to compare it to contractor rates, which start at around €350/day or €7000/mo (20 working days) in Europe.
(Our team spend on AI devtools comes out to around $1500/person/mo)
Well you can just scale your AI employees up and down as much as you want. Companies already pay a large premium for freelancers just to be able to fire them on a whim, so spending 5-10k a month on something that more than doubles the productivity of a senior developer might be well worth it as you can just adapt spending based on your business needs. If you can deliver a feature that lets you write a 100k invoice with 10-20k of tokens within a month or have a senior dev crunch that out in 6 months instead I think it's clear who wins. It's all about money and the AI companies know that, they have their pricing down exactly to sit in the sweetspot where it hurts just enough that companies can still afford it but not enough that they would look for cheaper alternatives.
In Sweden I always heard the figure to double the income of the person to get what the company actually pays, including taxes and "employers fee". I know this has gone down a lot in recent years, also not sure if it was ever exactly true, but likely very close anyway.
Hitting the first calculator I found gave me 50 kSEK costs 69 kSEK. So far from double nowadays.
In the UK, a £45k/yr employee pays their own tax and gets a take-home of £35k.
The employer pays £6k for National Insurance (atop the employee's NI contributions). Pension: 2-3k. Apprenticeship levy is £300. 3yr-amortised recruitment fee is £4000. Hardware costs: £1000. Office space £5000. Software/tools: £2500. Benefits: £1500. Training: £1000. Other admin overheads £500.
You pay that person for ~250 working-days, but they only attend for ~220, due to annual leave and sick pay, so you get around £62k worth of attendance out of that person in exchange for £70k, of which the employee sees £35k.
But that is 20% not 100%. And in most non retarded countries brutto is actually brutto, because there is no need to lie to people about how much the government takes away
Historically, this has nothing to do with lying, but is all about the founding idea of the social security system that all parties (workers, employers, state) carry part of the burden. Employers were supposed to pay their fair share because they also benefitted from the system (a sick or injured employee is not a productive one). Or saying it differently: the employer pays an insurance premium to reduce the effects of sickness. That premium is tied to the „value“ of the employee as measured by their salary.
There is plenty to improve with the system but to call it „retarded“ considering how much good it has brought to the world seems quite wrong to me. I don’t want to work in the pre-Bismarck era
In the UK, employers pay a stealth tax of 15% (recently increased from 13.8%) on top of the quoted salary minus the first £5k (recently decreased from £9,100.)
So your "£50k" salary actually costs your employer £56,750, and that's before all the other expenses mentioned elsewhere in this thread such as hardware, office rent etc.
A quick google tells me that software devs usually count for 20% to 40% of the total workforce in a software company. The rest is overhead that increases with every added dev.
And if one were to compare cost of a dev vs cost of an LLM, the dev comes with the cost of workspace, computers, sick pay, summer party, conferences and etc etc.
In the US, over and above salary, payroll taxes add 7.65%, pension contributions might be up to 5%, and employer healthcare and other insurance contributions can be in the thousands, plus other benefits, equity compensation, and per-employee software licensing, and lots of people just estimate 2x salary as the “total cost” of an employee, although that probably overstates it a bit.
I think you are broadly correct, but just to pushback on a few points:
(1) Ability to solve hard problems in days vs weeks as immense value
(2) Back-end improvements (if done right), should improve platform speed, stability, scalability etc. which should have revenue implication
(3) Ability to on-board a SWE equivalent entity in minutes, have them work on a specific hard problem and then off-board them in minutes can have value
All of the above, of course, depends upon Fable consistently being a 2x-3x SWE at minimum.
You're not really solving problems, you're retrieving the best match of solved problems from compressed corpus. And that corpus is available to many companies, meaning "hard" problems stop having "hard problem" value the moment they enter the weights of any model via the internet ... or distill from one model to another. Anthropics business model is commoditising knowledge, but as we see with the Fable model card, they only want it done to the knowledge of other businesses, in their own field, they totally hate it.
I don’t think that’s an accurate or useful characterization of modern AI like Claude at all. It is not simply regurgitating knowledge. It applies its knowledge to create bespoke solutions to the problem you pose to it, and is able to self evaluate its progress towards the completion criteria. If you don’t think that counts as “problem solving”, your definition would exclude nearly all knowledge work and engineering.
People underestimate the vastness of training data (internet) and overestimate their ability to recognize if something is really bespoke. Not to say the no problem solving is happening, because there are many problems that we inefficiently solve again and again and the LLMs are making the solutions more accessible to everyone with a subscription.
> It applies its knowledge to create bespoke solutions to the problem you pose to it, and is able to self evaluate its progress towards the completion criteria.
It imitates applying knowledge. The imitation may be uncanny, but assigning LLMs intentionality and ToM is a category error.
Does "applying knowledge" necessitate human-like intentionality and theory of mind? If you insist it does, and this is a category error, then we need a new category.
By analogy, consider that many have referred to classical, deterministic computing as some kind of "thinking" for the last half century+. Does this stop being kosher when the computer has an uncanny propensity for human language? Perhaps, but the computer is still clearly chewing through problems that would have required a lot of human thinking (e.g., arithmetic) in ages past.
I haven't seen any genuine proposals for words to replace the human mind analogues, let alone proposals that the anglosphere would plausibly adopt en masse.
> You're not really solving problems, you're retrieving the best match of solved problems from compressed corpus.
This is not correct. LLMs interpolate in a high dimensional space, so you're actually composing the best matches in a compressed corpus to find novel points/paths in that space. That is problem solving.
20x the cost means you need to have fable to be 20x better than the alternative, which is a tall order. And there's more options out there too, perhaps the 4x cost is enough.
This means if the deepseek / under 1k alternative is at least x1.2 improvement, fable needs to be x24, which I think is very2 unreasonable. It is possible for it to worth if it can x2 a $20k SWE, though I doubt it can do that.
“Ability to solve hard problems in days vs weeks as immense value”. Citation needed.
LlMs are incredible don’t get me wrong, but they are good on tiny contexts (writing a script). Not on software engineering (adding features to Chrome).
Honestly, LLMs been OK at adding features to software since around Opus 4.5. From what I've tried of Fable, it's a decent step up from the Opus models and I can only see things getting better.
The thing about AI-generated “solutions” is that they often go down bad rabbit holes and need to be re-run, or since they are so “cheap” to create they are often just thrown away and rebuilt when requirements evolve. Plus, just more stuff is created and needs to be maintained. So in the end, your efficiency gains go out the window.
I work at a smaller tech company (<300 people), and my friend showed me everyone's spending.
Our top user is at 10k a month, but the next highest is $2,000.
I would say the average is around $1,000-$1,500 for a developer.
We have completely unrestricted access to Claude, Codex, and Cursor.
Funny enough, the guy spending 10k is not even a dev by trade but an SME in what we work on that just vibe codes apps and somehow has not been cut off yet lol.
I have a single thread of GPT 5.5 medium running basically all work hours and I am around $1,500 a month in spend on Enterprise pricing.
At my company, most devs are under $1500 a month as well.
I’ve heard of a few cases of devs racking up bills fast, but it has typically been due to inefficient context usage. Like they just have one super long session with Opus 1M and are getting killed with input token costs and cache misses.
With careful context management and some thought into good approaches to problems, I have personally only rarely even hit $1k in regular use.
> Funny enough, the guy spending 10k is not even a dev by trade but an SME in what we work on that just vibe codes apps and somehow has not been cut off yet lol.
I'm guessing he's producing pretty valuable work. We have a few SMEs that vibe code tons of stuff with Claude. The only thing they really need tech for anymore is deployment and helping get their wheels unstuck on occasion.
Interesting! Would it be fair to say your company spend $100k to $150k per month on this?
Multiply this times many, many companies, and you can see how providing AI could theoretically be a good business to be in. Margins may be tight, though.
Also -- I'm convinced someone will figure out more use cases beyond software programming, which will result in many more companies spending $1k+ per employee per month.
It remains to be seen how much of this is a bubble.
No it doesn’t and will not be. Companies have not realised the cost yet, wait till the end of the financial year and you’ll see a different direction.
DeepSeek v4 is pretty decent, and probably on par with sonnet. I see a future of hybrid models where opus or fable might be used only for complicated features or bugs, but general day to day would be DeepSeek or whatever good models that will be released later.
>I switched to DeepSeek v4 Pro with OpenCode and am on track for a few hundred dollars of spend this month.
I was about to say that. Deepseek is just magnitudes cheaper and absolutely good enough for most things. Anthropic and co just try to milk the cow while its possible. If they cant compete with Deepseek pricing I do not see a bright future for them.
Its too bad my boss views China as the big evil country so he wont ever make the switch to Deepseek but then proceeds to throw all our data to US companies like OpenAI or Anthropic...
I recently switched off Max flat rate to Enterprise API pricing and I went from 200/mo to 10k/mo with the same usage pattern on Opus. They don’t offer flat rate to enterprises.
So what keeps your management from just buying everyone individual flat-rate Max subscriptions, or at least buying them for the users responsible for the sky-high token invoices?
I see a lot of comments like this but I don't understand why some people willingly pay so much more than others for the exact same service. What are you getting that I don't get as a $100/mo Max subscriber?
With GPT 5.5 on the $100 plan, it's hard to hit any 5h/7d limits - while allegedly being better than DeepSeek 4 pro. Not sure why, or how you spend "a few hundred dollars of spend".
With that said, I still had the Pro plan on Claude, I didn't expect much, but it blew up my 5h allowance on Fable with one simple single prompt, and it didn't even complete lmao
It's not just Pro! I have Max 5x and Fable absolutely blew up my 5h window. Didn't complete the code review either, and got downgraded back to Opus 4.8 on the really important memory safety parts I actually needed it for. It's an excellent model but Anthropic's not providing a good experience.
I'm on $200 plan which is supposedly 20x usage of $20 plan. With few Fable prompts (I'm working on u-boot port) I got 10% of my 5h usage, so that's already 2x of $20 plan usage and that would be 40% of $100 plan.
So Fable is just not usable for $20 plan and barely usable for $100 plan.
I am because CEOs are. Look where the puck is going. Sorry to update your p(doom) priors in this way, it was obvious to anyone paying attention years ago conditioned on uplift trend persisting. Trend persisted and here we are.
Welcome to the new world. People start to repeat what tech founders preach. They do not require humans in the mix. Peter Thiel gave a good example of that mindset in a (mostly) recent interview where he didn't have an answer on "Should humanity survive?"
Yes. Hiring people has various benefits, I will lay them out for you:
- They learn the domain of your product, which means long term ownership and knowledge establishes itself. If you've only ever shipped SaaS slop, you might not know, but lots of companies are solving real world problems that have no better solution. Owning and understanding the code and the domain is key.
- They will learn from their mistakes (no LLM does this).
- Human skill is a REAL moat. Once you build a team that fully understands and is skilled in the domain you work in, these people are going to be the thing that sets you apart. If some of them are particularly social or charming, let them sit in with you for meetings and watch them provide loads of value, for no added cost.
- If Claude or OpenAI is down, they will continue thinking. In fact, they will continue thinking even when off the clock! This is a neat little hack called "consciousness" where you get a lot of work for free!
- You can hire people who punch above their weight; not everyone you hire needs to be a 500k/year staff software prime engineer of doom, you can just spend some time and effort to hire good juniors/competent mediors who will think for themselves (gasp!) and get work done.
- You still get ALL THE BENEFITS OF AI!!!! They can use AI just like you can, or better!
- You get people who you can brainstorm with, which is distinctly different from LLMs because your employees are less likely to want to suck you dry in every sentence just to make sure you spend more tokens. Employees don't care if you love them, they care about the quality of their work if you manage them correctly and reward that.
- They are quite loyal if you treat them right; spend a little more on their well-being, and they will stick around, come in to work every day and deliver cool things with you.
- Humans can only manage, review and give tasks to so many agents. If you add more humans, you can handle more agents.
An expensive LLM and a lot of extra tooling gets you some of this, yes, but not all of it. With humans you can still do the expensive LLM and extra tooling if you end up making enough money anyway.
- AI isn't bound by need for rest, vacations, sick days, or labor laws
- AI doesn't bounce from company to company, taking your business knowledge with it (actually this isn't technically true based on the practices of AI companies, but that's not a technical requirement)
- AI doesn't join a union and stop work in demand for higher pay or workers rights
This is what CEOS and capitalists are thinking. For capital, the best outcome is to not have any labor at all. And if you can do that when your competitors can't, then you have a huge market advantage. (Slop notwithstanding)
I'm not saying this is a "good thing" but this is what drives the market. Less labor revenue in the long term and money printing machines.
The issue is, of course, that the quality of work is not good, and this will eventually show itself, likely in the total collapse of the US economy, but until then I wish them good luck with this.
I used Ory Kratos in a Go application a couple years ago by installing it as a dependency. It worked pretty well but in hindsight I would have hosted it as a separate application because it was a pain to bring along all of its dependencies.
One of my biggest complaints was that one of the Account Recovery flows was just an emailed 6-digit code. So a 1 in 1 million chance that somebody without access to any of your stuff could hack you by just hitting reset and guessing "123456". It's actually surprising how many other Account Recovery flows across the web I have noticed recently that do the same thing. Not sure if Ory has added the option for more entropy in this code as of today's release though it's been a while since I've used it.
Otherwise it was a great project to work with that has tons of knobs to customize. I commend the authors, aeneasr especially. It must be a ton of work to keep up with all of the auth standards and offer this in an Apache2 licensed package all while building a business around it as well!
Sure, but say the implementation lets you try 5 codes in that 10 minutes with a 30 minute lockout. An attacker could trigger Account Recovery, blindly try 5 six-digit codes immediately, and have a 0.0005% chance getting into your account.
They could script this to run over a long period of time targeting 1 account, or they could target many accounts at once, and would probably have success.
This is my biggest gripe with email auth or any kind of security code via sms/mms. I pray for the day I can fully move to a passwordless setup and break free the mess of email addresses spaghetti and phone numbers.
I sometimes dream of what it would have looked like to become a doctor (or PA or similar) instead of choosing Software. Mainly the allure of interacting with and helping more people.
This young person sounds like they are motivated enough to succeed at any study they put their mind to. Of course many companies will deny a young person employment based on age, just like they would deny them employment based on a lack of a formal degree.
But one day you turn 25, you are the right age, and you have the right degree. Then the praises for saving the company 70% on their cloud computing costs stop, and the same managers start asking you to work the weekend to fix other people’s code. And if you oblige, the burnout will become as real as a Doctor’s burnout, I imagine.
I still like AWS all these years later. It’s trusted in the enterprise and you can empower people to do what they need to themselves with IAM. And it’s pretty reliable.
Flash removal broke multiple government sites. I couldn't take a required training course for a few months after flash support was removed and the site was taken offline for an upgrade.
I’m sure ActiveX and Silverlight removal did too. And iframes not sharing cross domain cookies. And HTTP mixed content warnings. I get it, some of these are not web specs, but some were much more popular than XSLT is now.
The government will do what they do best, hire a contractor to update the site to something more modern. Where it will sit unchanged until that spec too is removed, some years from now.
Flash was dependent on a proprietary plugin from a single vendor. XSLT styled documents are compatible out of the box in any web browser from multiple competing vendors, even old Internet Explorer.
The iPhone never supported Flash. But thanks to web standards it supports viewing RSS feeds and other weird XML/XSLT artifacts from the past to this day.
What are those low-power devices (can you identify any?) doing with XSLT, then? If they don't have the power to do the transformation, it seems pointless for them to possess the template needed to perform the process.
XSLT is by and large single-threaded, and most jobs in the print domain get horrifyingly ginormous due to basic conceptual flaws of XML/XSL. Your Operations guys might have a panic attack when they see how that impacts on the server side. But then, to mitigate that, someone's going to need to cook a queueing system with some kind of notification/email doohickey, and now the InfoSec guys are also having a panic attack.
You're probably going to save money, end of the day, just homebrewing some XSL drop ins with a real programming language.
I gotta say, as a mostly-defense XSL guy[1] who also knows his way around TS and Py, this is probably going to be a real boom time to kick this dumb XSLT work to the curb and do some high dollar contracting making JS/TS drop-ins for govcons and defense.
[1] Who also thinks XSLT is a joke told by an idiot. My morale? Oh, it's great.
There would be no reason to fix this if the chrome people had kept up their end of the bargain by supporting the standard. We can quibble as to whether or not XSLT should have been part of the standard to begin with but it IS part of the standard.
Google says it's "too difficult" and "resource intensive" to maintain...but they've deliberately left that part of the browser to rot instead of incrementally upgrading it to a modern XSLT standard as new revisions were released so it seems like a problem of their own making.
Given their penchant for user-hostile decisions it's hard to give the chrome team the benefit of the doubt here that this is being done purely for maintainability and for better security (especially given their proposal of just offloading it to a js polyfill).
Commercial enterprises can only support standards if it's commercially viable.
It's commercially beneficial to make the web standard so complex that it's more or less impossible to implement, since it lets you monopolise the browser market. However complexity only protects incumbents if you can persuade enough people to use the overcomplicated bits. If hardly anyone uses it, like xslt, then it's a cost for the incumbent which new entrants might get away without paying. So there's no real upside for Google in supporting it. And you can't expect commercial enterprises to do something without any upside.
I expect commercial enterprises not to be allowed to engage in anti-competitive and consumer-hostile behavior. Like it or not and regardless of their contributions to tech/the web Google is notorious for pulling the rug out from under open industry standards only to replace them with their own proprietary or, as you described, "standards" that are so complex it's more or less impossible to implement so you're "forced" to use/buy their product.
They will be as anti-competitive and as consumer hostile as they can get away with. Adding and removing features from the standard is so ambiguously motivated that I almost can't imagine them being successfully prosecuted for it. In a way it's pretty clever.
Nobody is going to do things you agree with all the time. That doesn't mean everything they do should be condemned by default, without thorough investigation into their motives.
I don’t quite understand the part of the article that deems that you can skip all the checks under the assumption that this is an older browser, and that there is no CSRF vulnerability.
The algorithm seems sane for modern browsers. But you could probably find an outdated browser - older Android device WebView would be common -where the whole thing breaks down.
So I think tokens can be a thing of the past for modern browsers. I like the middleware, I hope it does show up in ASP.NET proper soon. My guess is they’ll keep tokens middleware around alongside it for some time once it does though, and the decision on which to use will come down to whether or not you want to make sure older browsers are secure.
I am the Product/Eng Lead and a Co-founder of a company formed ~1 year ago building AI-native developer tooling for Platform Engineers. Have been able to iterate very quickly through PoC phases and get initial feedback on ideas quicker. For features that make it into production code, we do have to spend some time re-working them with more formal architectures to remove "AI slop" but we are also able to try more things out to figure out what to move forward, so I feel like it is a net gain.
Part of "AI-native" means being able to really focus on how we can improve our Product to lessen upfront burden on users and increase time-to-value. For the first time in a while, I feel like there is more skill needed in building an app than just doing MVC + REST + Validation + Form Building. We focus on the minimum data needed for each form upfront from our users, then stream things like Titles, Icons, Descriptions, etc in a progressive manner to reduce form filling burden on our users.
I've been able to hire and mentor Engineers at a quicker pace than in the past. We have a mix of newer and seasoned Engineers. The newer Engineers seem to be learning far quicker with focused mentoring on how to effectively prompt AI for code discovery, scaffolding, and writing tests. Seasoned Engineers are able to work across the stack to understand and contribute to dependencies outside of their main focus because it's easier to understand the codebase and work across languages/frameworks.
AI in development has proven useful for some things, but thoughtful architecture with skilled personnel driving always seems to get the best results. Our vision from our product is the same, we want it to be a force multiplier for skilled Platform Engineers.
Not in all states. California does not have a lower tipped minimum wage. It's at least $16 here last I checked (except $20 for fast food because "reasons")
Ah yes; for a bit my wife was making less as a preschool teacher than the minimum wage at McDonalds. I understand it caused a bit of turnover at the local public schools, since cafeteria workers and aides were making less than $20/hr in 2024 as well (I don't know if they still are).
Reminds me of when I was visiting family a few years ago in Kentucky. I kept seeing tons of ads everywhere about hiring for plumbers, and some warehouse roles.
The listed salaries were not that far off from what even the local McDonald’s was paying
There was a (at least local) McDonald's inversion just after Covid; they were paying $20/hr which was competitive or better than the local factories.
It's one of those "reset" things you need to do now and then, because it's really easy for an industry like CNA or similar to end up paying less than the gas station for more annoying work.
So it does make great sense for a CNA or a preschool teacher to be quite a bit more highly paid than a gas station attendant or fry cook, due to the much higher responsibility level and like you said the annoyance.
However, I don't think anything that's happened in the last 5 years has helped that. If anything, the inflation has cost everyone dearly, but if I put 20% of my income into stocks I am less impacted than poor people who put 100% of their income into goods and services whose prices have gone up as a result of everyone's wages.
The reason we are not seeing this in mainstream software may also be due to cost. Paying for tokens on every interaction means paying to use the app. Upfront development may actually be cheaper, but the incremental cost per interaction could cost much more in the long term, especially if the software is used frequently and has a long lifetime.
As the cost of tokens goes down, or commodity hardware can handle running models capable of driving these interactions, we may start to see these UIs emerge.
Been using gRPC with json transcoding to REST on a greenfield project. All auto generated clients across 3 languages. Added frontend wrapper to pre-flight auth requests so it can dynamically display what users are allowed to do.
Claude Code has been an absolute beast when I tell it to study examples of existing APIs and create new ones, ignoring bringing any generated code into context.
So Fable would cost me 20k/mo at Enterprise rates. That’s around the average cost of a loaded SWE in the USA. “But I’m >2x more productive” doesn’t justify doubling the opex of the Software/IT department for most companies when revenue isn’t even up 10%.
I switched to DeepSeek v4 Pro with OpenCode and am on track for a few hundred dollars of spend this month.
Rewriting your stack from Ruby to Go in 2 days where it would’ve taken 6 months is impressive and fun. But that isn’t upping revenue.
Iterating on net new business features and ideas that are niche that the LLM isn’t trained for are much harder. Is 20x the token cost worth it there?
reply