If you were wondering the same thing I am - it's not about skills loss, quality, and less about money spent. It's more about frontier AI shops dogfooding their own models.
If it’s anything like AWS there’s hundreds of people making bespoke software factory setups, enhanced interfaces for ai tools, spinning up 10 parallel review agents with the best model available, etc because the budget is basically unlimited.
I find it a bit comical. AI tools are so easy to pick up, and it's unlikely you've pushed some highly critical features which couldn't have waited for a few months.
Talking to my friends working there, they said they were surprised when that happened and they didn't find Claude quality to be that much better than Gemini when using agy. So maybe after all the issue is not mainly the model, but harness and custom tooling.
I dunno, it could be about money, too. $100,000 budget per employee, per month. Why were they willing to spend that much on AI usage? I hope said employees also make that much in salary.
I think if you talk to LLMs and give feedback or openly say what works and what doesn't, you are essentially solving a captcha and produce accurate training data, while you pay for the token spend. I'd be a bit nervous with this lol.
Just one unsanitized input and you leak info. Or one hidden character and code may or may not belong to you anymore. Its very odd on many levels
These are companies famous for following the rules when it comes to handling other people’s data and IP after all, totally a non-issue and they would never violate contract law
Only if all you care about is developing models. I assume the rest of the business would rather just use whatever's best in class regardless of who made it, so I'm sure it's not that straightforward of a decision.
I can't be sure but I vaguely recall it being associated with the big Microsoft antitrust court case? Along with the delightful phrase "knifing the baby"
> Within Microsoft’s cloud and AI sector, monthly AI spending limits have reportedly been slashed from $100,000 per employee to approximately $10,000 in the majority of instances.
They were allowing $100,000 per month per employee?!
That's incredible. Just a few years ago that would have been zero.
My company took away my Claude because it’s too expensive. I feel like there is a reckoning coming. The accountants are finally realising the cost of token maxing.
That's pretty stupid. Most people who are incurring significant costs are just tokenmaxxing rather than being efficient with usage. You can get 99% of jobs and work done with Haiku/Luna in a collaberating working enviroment.
I feel like people who are later to the AI game just like to "oneshot" and sink a bunch of usage into generating garbage
There really is a skill to using it effectively. I've tried coaching some of the devs on my team. Some get it, some don't.
Our company has been tracking token usage and models used vs output (tickets, story points, PRs, deploys, etc...). A dev got chewed out, even after I warned him, because he spent over $2k in a single month almost exclusively on Opus while his actual productivity in terms of what he delivered was abysmal.
I get it but it goes against the grain for me. Isn't it ironic that we have to waste our precious and expensive human brain cycles to think about how to use AI cheaply so that it is not more expensive than us?
In other words I want to spend 100% of my mental capacity in the problem domain for the things AI cannot do for me, like steering, grounding, verification and not for things AI could do.
> In other words I want to spend 100% of my mental capacity in the problem domain for the things AI cannot do for me, like steering, grounding, verification and not for things AI could do.
It sounds like the parent is less talking about this, and more people burning tokens while not getting useful work done.
Claude, vibe code me an entire startup, the actual product doesn't matter, but it should all be based on the incredible pun "turn 'sorry' points into story points."
/goal get accepted into Y Combinator, you have an unlimited token budget, be bold.
EDIT: no, do not just make a product that gives away your unlimited token budget to users for free!
Attempting a serious but not-a-certified-whatever answer: "Points" do have meaning when properly used as a kind of moving-average tool for forecasting within a particular context.
Problems arise when people try to perma-peg them to particular tasks, or (worse) man-hours or (much worse) man-hours across teams. Even just encouraging the humans to answer in terms of hours/days taints the accuracy of the forecast by introducing a kind of bias.
So, what you do is you recognize every ticket has a somewhat variable “actual effort”; and, if you’ve been honest in approximate effort pointing, you’ll know your team (or your own) velocity.
From there you can run Monte Carlo simulations - say a few hundred thousand, and get a pretty good estimate of actual time spent.
Assume that for various practical reasons, we've decided it's one of the best functions out there. How do we use it effectively, especially when it has noise, and drifts over time with unseen changes to the human_estimator and the hideously complex world_state?
A popular option is to run it multiple times with different person/task combinations, putting a projected number on to each task. Afterwards, the tasks finished in sampling period ("sprint") become a quantifiable total for that period ("velocity").
Do the same process again with the next set of tasks, and you can figure out which ones are likely to fit if the velocity doesn't change much. If you know the velocity will change due to losing staff or vacation days... well, we apply a multiplier and hope for the best.
Trying to "fix" the meaning of points is maladaptive, because they reflect many changing things which are outside our control and can't be independently measured.
My wife, despite loving all things French, just doesn't "do metric". She wants my height in feet and inches, my weight in pounds, boom done. So while I know my mass in kilograms, she needs the conversion done before she can even begin to have a reference point.
Upper management is the same way. They have forecasts that they need to make, deadlines and budget goals that they need to hit. They only deal in the units of hours and dollars (or local currency). Every software engineer is accountable for their work in those units only. The conversion needs to be done before the management chain has a reference point.
One easy way to do this is to have each engineer estimate the time it takes to fulfill a story after it's been pointed; then, upon completion, record their actual hours spent. Their estimated vs. actuals tend to stabilize over time, so even if they misestimate a task, you can arrive at a good guess at the time it will actually take.
It's funny, the thing that makes effective prompt also makes effective documentation/communication.
It's bizzare to see people that made clown issues (not enough detail etc.) suddenly start writing detailed prompts just because it is AI that will do the task and not the human on the other side.
You mean that now rather than taking 30 seconds to write a ticket, because we all know what we're talking about, we must take 10 minutes to give all the background information to the AI?
there's a manifold to what "effective" means. The problem is once you get into the vibe flow, it's really difficult to eject yourself into the other realms of vscode or IDE or whatever it is you normal do because the vibing provides no anchor to what you're doing.
Even if these models are smart enough to reorient themselves, they get entirely stuck in a desert and now you're asking someone to just pull up stakes and digg them out even thought they only watched them get there and the UI provides so much speed that no human can comprehend how they got there in the first place.
It's like asking a pilot to take over in an emergency situation when they're not tasked with any of the every day requirements of the job. The orgs are relying on borrowed time of experienced professionals, and that's going to erode away and what replaces it is mostly people who understand how to navigate context but not use any of the _classic_ tools.
It's a real conundrum and won't be easily surfaced but for a decade.
I’m trying really hard to keep my skills up but it doesn’t feel productive when I’m using it to write code. It feels like I’m slowing down the AI to the point that it’s not as effective as just letting it go. But I don’t get all the learning that comes from that time along the way.
Have you found ways to stay sharp while using it? Or are you relying on other projects outside of work to keep your skills fresh?
I feel like it's only within the past few months that opus got to the point where guiding the model is faster than doing things myself. I tried out sonnet recently and it was not a net positive to my work. I feel like anything that I'd trust haiku to handle isn't worth doing in the first place.
For context, I'm doing a range of tasks, everything from one-shotting adhoc scripts to having 4 hour 10M+ token conversations debugging things.
in my company there are a few who keep sharing screenshots of reaching limits on 3 separate subscriptions, 2 of them their personal on top of the company subscription
Wonder when subscription-hopping attacks will become more often (jumping from a personal model to injecting instructions into the business account and exfiling data)
Or does a "collaborating work environment" mean that everything is basically spoonfed to them? Or do you only ever use ghost suggestions?
I genuinely cannot even fathom. Just how do you even get into a state where tasks are so clear and cookie cutter? These things are abhorrent. Not only are they not useful, it's an outright form of psychological torture to try and use them. They almost fight you.
Luna doesn't even respond to steers properly! You try steering it and it immediately gets distracted and then just stops.
I can imagine coercing Sonnet into doing some of my tasks okay, but Haiku? Especially 4.5? Really?
I think you might be overestimating the sort of projects most of us have worked on throughout our careers -- we haven't been doing much groundbreaking work. LLMs can easily and successfully write most code.
It's possible it's my role distorting my perception, cause technically I don't write software, I work an SRE role. None of my items come pre-chewed or paced, it's all good luck and god bless.
I'm desperately trying to classify and standardize my work items and delegate them to less capable models, because my usage is clearly unsustainable and this same sentiment as above keeps being pushed on me too. But all my tasks are genuinely fairly arbitrary, so there's no real way around the agent actually being able to reason about business and technical context proper. It's not even that they're hard, it's just that they're dynamic.
I can get Luna to do things like walk our observability stack and perform a healthcheck, then defer to a stronger model if anything looks super off, but if I'm being entirely honest, this could basically be just a script. Which Opus 5.5 will immediately write for itself if it doesn't yet exist, run that, and then off it goes depending. But Luna will never actually do an investigation proper. Heck, it can't even read our dashboards most of the time, tripping up on Grafana minutia.
It feels like that surgeon vs surgeon comparison, where you're made to decide based on their surgery success rate, and the better succeeding surgeon simply reward hacks the number by only operating on less dicey cases. Except there's no objective way to make this classification here, so jackasses like the above get to play with my insecurities with full obnoxious confidence, while I'm left desperately trying to slim my usage and failing to do so between two moments of crippling self doubt and blockers.
For the last few weeks I've been using Luna with Medium Reasoning for routine debugging, installing/updating dependencies, test creation/fixes, CLI/config miscellaneous problem solving, etc and it's been solid. I can let it churn away for a half hour and it barely moves the needle on remaining usage on my Plus plan. Previously I had been using Sol Low/Medium and would frequently hit the 5 hour limit doing those kind of tedious automation and routine problem solving tasks.
I wish we had such a platform (and was properly adopted). Maybe then the necessary context would be properly organized, and these lesser models could be effective here as well, especially if combined with harnessing integrations too. I still have a hard time accepting that Haiku/Luna tier models can be effective there even then, but I'll just have to take your word for it I suppose.
I think they mean the new Haiku, which is mildly above Luna now . If you have a plan written by a smarter model (so the hard parts are solved) they can be great at implementation.
That's funny. In a meeting with my manager recently, they specifically called me out for not spending enough money on tokens. It's not like I didn't use it, just apparently not enough, and apparently on too low a setting.
Since then I've had fable cranked up to 11 for even the most trivial of tasks.
I know plenty of people working at very well known and very large companies who's CEOs were boasting about "not hiring anyone anymore", "all the code will be generated in 3 months", etc. they all went from "unlimited budget per dev" + public dashboard with ranking to flex how much credits everyone was burning to hard caps at $500-$1000/month/employee real fast. Some are even not allowing their devs to use the more expensive models
Do you know the details of the Claude Code plan you and your company are using (if not part of some enterprise deal)? Does your individual capacity out run something like Claude Max 20x ($200/mo)?
I use Opus 5.5 heavily but only spend around $800/week at API rates. I mean, I say "only"... That's a lot in absolute terms, but trivial compared to my salary and EASY worth it.
And someday I hope to understand why they do that. CEO: "Let's see, I can pay $200/month for Bob's tokens, or I can pay $2000/month or so, and then hope he doesn't screw up and rack up a seven-figure bill. The service is the same either way. Hmm."
Is there something in the ToS that says you can't give each of your employees a personal $200/mo subscription... and if they happen to use it at work, well, that's their business?
Even if so, it might be worth spinning up a separate LLC for each division, given what they are charging for the API.
The reckoning started years ago when we did the equivalent to token maxxing hiring coders for everything to crank LOC
Software is inherently a physics problem not all the job titles and specializations made up the last 20 years as dev job salaries kept attracting people
That was all illusory social construct to prop up jobs
Still a whole lot of that in tech but it's all at the top of the org now. Leadership sensory experience and thus innate habit to forecast future been programmed by years of yes men they refuse to accept the jig is up for them too
Sensory memory of being a useless figurehead fosters a lot of existential dread in priests, politicians, and the like. Completely aware their day to day effort is insufficient to sustain them they know how co-dependent they are. They'll dig in harder.
See Chris Matthews flame out shrieking about socialist execution squads. Dude seriously thought everyone wants to hang him from a lamp post. The reality is people just want a sense of control back and not have their perception dragged along by Chris Matthews.
Does all this velocity translate to increased income to the business though. At some point if we are releasing 20x more features, more games, more music, who is actually buying it all?
At this point we're solidifying a lot of stuff for the product: disaster recovery, getting workflows to scale so we can actually sell it to more people, paying off the Fort Knox of tech debt, so, at least in my case, yeah. YMMV.
Is that not a widely-known word in project management circles for scrum and other methodologies? I've heard it for years. Basically how much work you get done through the lens of how long it took to get done via a pointing system.
I didn't know people were using these non ironically.
Might as well go back to counting the number of lines or number of commits. It doesn't sound as good as "granularity" and "velocity" though. The good thing with velocity is that it doesn't care which way you're going as long as you're going there fast, so you can never be wrong
which is quite sad because opus 5.5 is really good. i say this as an anthropic hater. i wish I could move away to other models like 6.1 sol or deepseek or whatever, but they just all lack something. i _trust_ opus 5.5
i hope other labs catch up, especially chinese labs.
I mean, we’re not far from a situation where instead of how many story points you completed per sprint the metric to optimize is going to be what was your efficiency? How many story points did you complete while minimizing your token usage. In fact, that’s a pretty good idea. I’m going try and implement it at work with some sort of complexity normalization function
If true, it is a huge blow to Anthropic’s revenue stream. IIRC it was reported that the quarter of their revenue comes from just two clients and as the ex-Meta guy who left this July, I am convinced that Meta must be one of the two.
> The company's financial trajectory already shows how quickly that equation is changing. Anthropic's revenue run rate was about $9 billion at the end of 2025, according to the company, before rising to more than $47 billion by May. Anthropic has projected revenue of at least $10.9 billion for the second quarter of 2026, more than double the previous quarter, on track for its first quarterly operating profit of $559 million.
> The company has told a small group of shareholders that its adjusted operating income will be positive for the second consecutive quarter, according to multiple people with knowledge of the matter.
These are leaked and self-reported numbers, but no matter how much skepticism you pile on them it still looks likely that Anthropic in 2026 have had some of the fastest revenue growth of any company in history.
There is a conspiracy theory that the tokenmaxxing period was anthropic/openai ipo play. They do the tokenmaxxing to revenuemaxxing first, then time the ipo window to show they had the the huge growth to justify their price tag.
However Elon was able to ipo before them which took a lot liquidity of the market. Their financials are exposed. The market condition and sentiment now is in the gutter. It would be very interesting to see how these would pan out
Considering that’s a healthy portion of a salary for an additional employee per person, the fact they’re slashing spending sure makes it look like AI wasn’t even a 2x multiplier at minimum.
You could only conclude that if they would have dropped LLMs completely. It just means they see the benefit/cost optimum at a lower point than 100k/programmer.
I also doubt a longterm 2x multiplier for most developers.
I don't think you need the LLMs to be completely dropped to indicate this. Even entry-level faang programmer salaries are in the 300k range, so the fact that LLMs aren't worth 100k/programmer does tell us something about the marginal usefulness. If the drop was motivated was motivated purely by price-per-output, it would indicate the LLMs are less useful than adding another 1/3 engineer. However, as discussed in other comments, there are other motivations as well, notably dogfooding homegrown tooling.
We're sitting on a year of more or less capable coding models and I have not exactly seen a revolution in new software being released. AI is likely a force multiplier for specific subsets of individuals and workflows. Coding is not a bottleneck for a lot of systems.
I suspect that SaaS is silently being eaten alive as an industry.
I think SaaS will be fine. How the field works might change, but in the end marginal cost of software is still zero and when you've got a quality, well-tested and well-designed product it's going to be obviously better than things vibed together.
You often pay for regulatory compliance and liability with SaaS. Would you risk running your own vibe-coded payroll or accounting system for your company?
It’s mostly from people using their personal accounts to run LLM services that serve a larger team or organization. At least at Microsoft, it’s still impressively hard to get access to an LLM for service usage with high enough rate limits to be useful, making running services on dev boxes much more appealing (despite the countless drawbacks that few people seem to care about around security, compliance, reliability, etc).
Just to sprinkle more details here about MSFT. The usage was trimmed down even more for some orgs and some specific accounts (I think those were vendors and the like). But if you are an FTE and somewhere in CoreAndAI then your limit is like 1M GitHub credits ($10k as article mentions) that could be spent on any model you have access to, e.g. Opus 5x or GPT-x. It is hard to pin down how many people use Claude vs how many just use internal harness over the Opus though. Importantly, Anthropic models are more expensive (not even talking about Fable here as it is a special case) as seen in usage metrics. Peeps can check their usage which has a good breakdown that shows how many tokens were burned on which model and how much that "costs" (the prices as you can imagine are not consumer ones), so IMO this change will drive more people to optimize based on their usage patterns if they are heavy hitters. The easiest path is to just switch to something like Sol/Astra and have similar perf for lower price.
> Peeps can check their usage which has a good breakdown
Must be nice. At Sony we use Copilot (as an MS enterprise customer) and restrictions to token usage circa June have resulted in a hunger games of who-can-use-them-all-first, with the pool usually drying up by the 10th day of each month.
As for usage breakdown: we got a manually-published table most days just showing per-user token usage, until eventually we were told even that was somehow too difficult to provide.
Between that and learning how much the whole of Sony Corporate Group spends on Copilot (spoiler: it's hilariously little), I've sort of run out of disbelief.
> Importantly, Anthropic models are more expensive (not even talking about Fable here as it is a special case) as seen in usage metrics.
Not sure where you see that. The API prices of Opus are cheaper than Astra and there was a study suggesting that the Anthropic abos are offering more bang for the buck:
Right! it seems obvious why: Both these companies want to dogfood their own coding models and stop paying competition.
You can also read this as diminishing returns / AI isn't good enough, etc, but the simplest explanation is that they don't want to send money to Anthropic.
This, and in addition to dogfooding, incentivizing employees to be more effective with the cheaper models. A lot of problems don't need anything fancy, but it takes more brain power and engineering effort to make that work. By default humans will take the path of least resistance if it's available.
But Microsoft and Meta are not blocking competitor tools for internal use, they're merely trying to reduce costs and divert a fraction of use to their own technologies. Microsoft and Meta are both still spending nine figures a year on Claude, and the article does not state or imply they're even considering a complete halt.
I feel like "good enough" was reached around Opus 4.6 - 4.8. All I wanted after that is improved speed, continued tweaks to the tooling to get the most out of it and quality of life features added.
The first things I install on Windows 11 are OpenShell and ExplorerPatcher to get rid of their crap taskbar and start menu and bring back the Windows 10 taskbar and start menu. What they did with Windows 11 is an abomination.
What does it even mean to use "Copilot" these days?
I use the Github Copilot app at work but it's really just a tool that gives models access to our Github repos and my local clones of them, and it's been super helpful to just tell it to reference other repos for patterns or integrations or whatever, get agents and an orchestrator to look at multiple repos to make a change across several projects and then implement them, etc.
And yeah you can use a selection of whatever model, I was using Opus 4.6 for a while, now I'm on GPT-6 Sol. It's been a very good experience.
AFAIK there is no such explicit, company-wide effort at meta. The article seems to try to slip Meta in there with whatever reporting they are doing on Microsoft, despite there being no such evidence for meta.
To also add an important context to the drop in Claude code users reported for meta - this elides that we are absolutely still using the Anthropic models full steam ahead, but are moving towards internal interfaces and harnesses.
So I would question if that 50% drop in CC users is more of an interface change than anything else.
I for one have stopped explicitly using codex or CC entirely. But the interfaces I am using still use those harnesses under the hood. I wonder how that is counted.
From what I could see these numbers would usually be hit by people doing vast amounts of analysis over files, they would usually rock up tokens very fast. Then you combine this with the selection of Opus models which are much more expensive internally than GPT ones and the limit gets reached quite quickly.
I assume that at shops that both employ engineers and are developing an AI product, internal usage is not about improving productivity, it is about improving the offering. Of course they want employees to use internal tools.
Maybe I am slow here and everyone is using Claude with credits at max use. But isn't Claude Teams like $25/month per developer for ordinary use? What the heck of these guys doing that makes it get that phenomenally expensive for their use cases? These are presumably well capable engineers who started to use this as an aid right not just throw Fable at everything and loop to the max?
There are people out there building AI building orchestrators for orchestrators for orchestrators for agents. The author of that blog post later claimed to be spending the equivalent of $122k/month on tokens (by rotating their usage between 21 accounts).
As far as I can tell, the only thing that this level of spend has produced so far is an indie 2D RPG video game.
> Gas Town was intended to be reusable, but I only ever wound up using it to build itself. Gas Town fell apart at the seams with Opus 4.7. Up through 4.6 it was working brilliantly. With 4.7 we saw the introduction of the "just two more things" tic, which prevented Opus from ever converging on being ready to do real work—it always wanted to fiddle with Gas Town itself. The Opus tic never went away, so Gas Town effectively burned down. It had other problems, too, but 4.7 was the final straw
I hadn't seen the $122K figure mentioned previously. $87K for API-style pricing was mentioned in the above post, and ~$2,800/mo for multiple Claude Max accounts:
> My solution has been to create a token tap on $200 Max accounts, which for me work out to ~30x the list-price equivalent. So in reality I'm only spending about $2800/month out of pocket for my $87k "worth" of tokens. Though that number keeps growing alarmingly.
(He never seemed to provide numbers for Gas Town initially, whether what he paid, or what the API-style pricing would have charged, so it was interesting to get some actual figures. Sounds like a lot to me, but if he's genuinely getting multiple people's-worth of work out of it, then...)
(Also in the article: a little morsel of Emacs content, which was nice to see.)
He's always come across to me as at least a little bit like this! His AI stuff doesn't feel enormously out of character to me. (If you hadn't encountered him before Gas Town - he's been posting online for like 20+ years, got a Wikipedia page and everything, so you can find his back catalogue and judge for yourself.)
I should probably know more about Claude's TOS, but it is probably a mistake for these companies not to leverage the plausible deniability of their usage and turn it into a massive distillation resource for their own models.
At Microsoft, the recommendation is to use cheaper models for tasks that don't need frontier models. I personally use Lua and it's more than capable for complex problems.
It doesn't say they're cutting back on OpenAI usage.
To me this seems mostly related to the way Anthropic showed the level at which they monitor sessions, plus wanting to limit how much training data they're directly feeding into a company that competes with their own products/investments.
The cybersecuritynews.com news one simply republishes details of a story published by The Information. At least they have the decency to LINK to that Information story:
CEO's nephew showed him how good the Chinese models are?
I am only half joking, I heard something like "my son or nephew did this cool thing with $X so we'll take $this_radical_step because of it" enough times over my career.
There is a perceived opportunity cost from someone using a lower-tier model on their task. What if the better model did a "better" job? what if my trials and tribulations are due to model quality?
If you are used to talking to opus5.5 medium, going to GPT6.1 luna low will feel like a step down. Why would any employee take the (personal) risk?
It may bei cost efficient, but is it wise? We use not only the big US models, but also Chinese ones. This way we can compare who makes the difference. Simplified: Knowledge comes before economic aspects.
You're opening yourself up to data right and privacy risks with that. My company demands that I use their enterprise account because they can claim full ownership of all produced output and have full logs of every interaction.
I think that gets legally murky, if the employee is the one who pays for the tool.
Wait, okay maybe I'm in a bubble, but every org I've worked at understood that giving us whatever hardware we want more than pays for itself, and very quickly.
I can't fathom the logic of paying 10k a month for a developer and giving them an 800 device.
Insane. Your leadership is severely broken. If you can, work somewhere that values you.
Just one unsanitized input and you leak info. Or one hidden character and code may or may not belong to you anymore. Its very odd on many levels
At best you produce some noisy signals that are going to have a tiny impact if even that.
And that's on a personal plan where you didn't opt out of sharing usage data.
Business plans offer zero data retention. This is a non-issue.
I think this data is likely worthless compared to curated RL tasks.
Like its a no brainer to force your employees to use your own models, then RL train them to be better.
I am not convinced that's the case.
... in 1988.
https://en.wikipedia.org/wiki/Eating_your_own_dog_food
They were allowing $100,000 per month per employee?!
That's incredible. Just a few years ago that would have been zero.
I feel like people who are later to the AI game just like to "oneshot" and sink a bunch of usage into generating garbage
Our company has been tracking token usage and models used vs output (tickets, story points, PRs, deploys, etc...). A dev got chewed out, even after I warned him, because he spent over $2k in a single month almost exclusively on Opus while his actual productivity in terms of what he delivered was abysmal.
In other words I want to spend 100% of my mental capacity in the problem domain for the things AI cannot do for me, like steering, grounding, verification and not for things AI could do.
It sounds like the parent is less talking about this, and more people burning tokens while not getting useful work done.
Define productivity, and while at it, quality, maintainability , modularity and so forth.
/goal get accepted into Y Combinator, you have an unlimited token budget, be bold.
EDIT: no, do not just make a product that gives away your unlimited token budget to users for free!
Problems arise when people try to perma-peg them to particular tasks, or (worse) man-hours or (much worse) man-hours across teams. Even just encouraging the humans to answer in terms of hours/days taints the accuracy of the forecast by introducing a kind of bias.
So, what you do is you recognize every ticket has a somewhat variable “actual effort”; and, if you’ve been honest in approximate effort pointing, you’ll know your team (or your own) velocity.
From there you can run Monte Carlo simulations - say a few hundred thousand, and get a pretty good estimate of actual time spent.
I’ve seen it work before with shocking accuracy.
estimate(human_estimator, task_description, world_state) -> numeric_effort
Assume that for various practical reasons, we've decided it's one of the best functions out there. How do we use it effectively, especially when it has noise, and drifts over time with unseen changes to the human_estimator and the hideously complex world_state?
A popular option is to run it multiple times with different person/task combinations, putting a projected number on to each task. Afterwards, the tasks finished in sampling period ("sprint") become a quantifiable total for that period ("velocity").
Do the same process again with the next set of tasks, and you can figure out which ones are likely to fit if the velocity doesn't change much. If you know the velocity will change due to losing staff or vacation days... well, we apply a multiplier and hope for the best.
Trying to "fix" the meaning of points is maladaptive, because they reflect many changing things which are outside our control and can't be independently measured.
My wife, despite loving all things French, just doesn't "do metric". She wants my height in feet and inches, my weight in pounds, boom done. So while I know my mass in kilograms, she needs the conversion done before she can even begin to have a reference point.
Upper management is the same way. They have forecasts that they need to make, deadlines and budget goals that they need to hit. They only deal in the units of hours and dollars (or local currency). Every software engineer is accountable for their work in those units only. The conversion needs to be done before the management chain has a reference point.
One easy way to do this is to have each engineer estimate the time it takes to fulfill a story after it's been pointed; then, upon completion, record their actual hours spent. Their estimated vs. actuals tend to stabilize over time, so even if they misestimate a task, you can arrive at a good guess at the time it will actually take.
It's bizzare to see people that made clown issues (not enough detail etc.) suddenly start writing detailed prompts just because it is AI that will do the task and not the human on the other side.
Even if these models are smart enough to reorient themselves, they get entirely stuck in a desert and now you're asking someone to just pull up stakes and digg them out even thought they only watched them get there and the UI provides so much speed that no human can comprehend how they got there in the first place.
It's like asking a pilot to take over in an emergency situation when they're not tasked with any of the every day requirements of the job. The orgs are relying on borrowed time of experienced professionals, and that's going to erode away and what replaces it is mostly people who understand how to navigate context but not use any of the _classic_ tools.
It's a real conundrum and won't be easily surfaced but for a decade.
Have you found ways to stay sharp while using it? Or are you relying on other projects outside of work to keep your skills fresh?
For context, I'm doing a range of tasks, everything from one-shotting adhoc scripts to having 4 hour 10M+ token conversations debugging things.
There is no way I can beat even local models at generating complex Python scripts fast.
hn is filled with uber geniuses.
Or does a "collaborating work environment" mean that everything is basically spoonfed to them? Or do you only ever use ghost suggestions?
I genuinely cannot even fathom. Just how do you even get into a state where tasks are so clear and cookie cutter? These things are abhorrent. Not only are they not useful, it's an outright form of psychological torture to try and use them. They almost fight you.
Luna doesn't even respond to steers properly! You try steering it and it immediately gets distracted and then just stops.
I can imagine coercing Sonnet into doing some of my tasks okay, but Haiku? Especially 4.5? Really?
It's just glue code
It's not complicated. Someone just has to be there to squeeze the bottle
I'm desperately trying to classify and standardize my work items and delegate them to less capable models, because my usage is clearly unsustainable and this same sentiment as above keeps being pushed on me too. But all my tasks are genuinely fairly arbitrary, so there's no real way around the agent actually being able to reason about business and technical context proper. It's not even that they're hard, it's just that they're dynamic.
I can get Luna to do things like walk our observability stack and perform a healthcheck, then defer to a stronger model if anything looks super off, but if I'm being entirely honest, this could basically be just a script. Which Opus 5.5 will immediately write for itself if it doesn't yet exist, run that, and then off it goes depending. But Luna will never actually do an investigation proper. Heck, it can't even read our dashboards most of the time, tripping up on Grafana minutia.
It feels like that surgeon vs surgeon comparison, where you're made to decide based on their surgery success rate, and the better succeeding surgeon simply reward hacks the number by only operating on less dicey cases. Except there's no objective way to make this classification here, so jackasses like the above get to play with my insecurities with full obnoxious confidence, while I'm left desperately trying to slim my usage and failing to do so between two moments of crippling self doubt and blockers.
I wouldn't even tried it, i would still just go with even Opus (we don't have that many alerts) but it really surpsied me.
When i ran into usage limits a few days ago i switched most to Sonnet and again was surprised how good it is now.
Optimally? Opus will pay for itself if you save just 10% of your time
So be less snarky?
Since then I've had fable cranked up to 11 for even the most trivial of tasks.
>Great Depression style collapse and all the current AI companies go bankrupt.
Oh this is just a 33 day old doomer account.
But even $200/month is worth shaving if it doesn't generate value.
um.
This takes some doing and now is the time where it's dawning on the finance departments.
So it's not $200 a month but it can easily reach $200 a day, and unless you're a startup playing with monopoly money the maths don't work
Even if so, it might be worth spinning up a separate LLC for each division, given what they are charging for the API.
You had budget for your normal salaries, for externals and now suddenly you have a few millions additional.
What do you do? You compensate.
Business people doing business things.
Since this summer coding on Opencode Go + Codex for a total 28$/month gives me more intelligence and token than 400$ did in may.
Also, SOTA models are increasingly useless for anything even barely tangential to security work.
Software is inherently a physics problem not all the job titles and specializations made up the last 20 years as dev job salaries kept attracting people
That was all illusory social construct to prop up jobs
Still a whole lot of that in tech but it's all at the top of the org now. Leadership sensory experience and thus innate habit to forecast future been programmed by years of yes men they refuse to accept the jig is up for them too
Sensory memory of being a useless figurehead fosters a lot of existential dread in priests, politicians, and the like. Completely aware their day to day effort is insufficient to sustain them they know how co-dependent they are. They'll dig in harder.
See Chris Matthews flame out shrieking about socialist execution squads. Dude seriously thought everyone wants to hang him from a lamp post. The reality is people just want a sense of control back and not have their perception dragged along by Chris Matthews.
Our velocity is twice as high as it was before Claude, so I doubt that we'll ever go back, but I could see efficiency being a priority.
Is this the new buzzword for the quarter? Last quarter was "granularity", I didn't get the memo yet
Might as well go back to counting the number of lines or number of commits. It doesn't sound as good as "granularity" and "velocity" though. The good thing with velocity is that it doesn't care which way you're going as long as you're going there fast, so you can never be wrong
I'm sure investors will love it.
Now we're starting to see real impact from AI, people are learning how to use it, and OpenAI cut prices by no less than 60% like a week ago.
You think now is the time they're going to cut the spend?
i hope other labs catch up, especially chinese labs.
In 2026 their revenue has gone up by a factor of more than 10x, and they no longer have just two whale customers.
I heard a rumor recently that customers spending less than $100m/year aren't even considered their "top tier" now.
Where are you getting the 2026 figures from? The Bloomberg article that was "on track to generate" and just prediction?
The most recent reporting from Reuters themselves (somehow not included in their more recent article about the IPO stuff): https://www.reuters.com/business/anthropic-ipo-valuation-hin...
> The company's financial trajectory already shows how quickly that equation is changing. Anthropic's revenue run rate was about $9 billion at the end of 2025, according to the company, before rising to more than $47 billion by May. Anthropic has projected revenue of at least $10.9 billion for the second quarter of 2026, more than double the previous quarter, on track for its first quarterly operating profit of $559 million.
Here's the FT: https://www.ft.com/content/4564e6a5-69e9-40a6-bf0f-a888f2f4f... - "Anthropic tells investors it will be profitable for second straight quarter"
> The company has told a small group of shareholders that its adjusted operating income will be positive for the second consecutive quarter, according to multiple people with knowledge of the matter.
These are leaked and self-reported numbers, but no matter how much skepticism you pile on them it still looks likely that Anthropic in 2026 have had some of the fastest revenue growth of any company in history.
However Elon was able to ipo before them which took a lot liquidity of the market. Their financials are exposed. The market condition and sentiment now is in the gutter. It would be very interesting to see how these would pan out
what. I can see a team of 20 costing 100k per month (but rare), but per person?
I also doubt a longterm 2x multiplier for most developers.
I suspect that SaaS is silently being eaten alive as an industry.
Must be nice. At Sony we use Copilot (as an MS enterprise customer) and restrictions to token usage circa June have resulted in a hunger games of who-can-use-them-all-first, with the pool usually drying up by the 10th day of each month.
As for usage breakdown: we got a manually-published table most days just showing per-user token usage, until eventually we were told even that was somehow too difficult to provide.
Between that and learning how much the whole of Sony Corporate Group spends on Copilot (spoiler: it's hilariously little), I've sort of run out of disbelief.
Not sure where you see that. The API prices of Opus are cheaper than Astra and there was a study suggesting that the Anthropic abos are offering more bang for the buck:
https://news.ycombinator.com/item?id=49975345
You can also read this as diminishing returns / AI isn't good enough, etc, but the simplest explanation is that they don't want to send money to Anthropic.
Windows 11 shipped with a broken task bar, couldn't even ungroup items. No power users was ever involved in this.
I use the Github Copilot app at work but it's really just a tool that gives models access to our Github repos and my local clones of them, and it's been super helpful to just tell it to reference other repos for patterns or integrations or whatever, get agents and an orchestrator to look at multiple repos to make a change across several projects and then implement them, etc.
And yeah you can use a selection of whatever model, I was using Opus 4.6 for a while, now I'm on GPT-6 Sol. It's been a very good experience.
"Fine. Please no. I'm out..." in that order
To also add an important context to the drop in Claude code users reported for meta - this elides that we are absolutely still using the Anthropic models full steam ahead, but are moving towards internal interfaces and harnesses.
So I would question if that 50% drop in CC users is more of an interface change than anything else.
I for one have stopped explicitly using codex or CC entirely. But the interfaces I am using still use those harnesses under the hood. I wonder how that is counted.
$1.4B/year is not a small number, even at Meta's scale, when it's money going to a competitor
There are people out there building AI building orchestrators for orchestrators for orchestrators for agents. The author of that blog post later claimed to be spending the equivalent of $122k/month on tokens (by rotating their usage between 21 accounts).
As far as I can tell, the only thing that this level of spend has produced so far is an indie 2D RPG video game.
Regarding Gas Town, see also https://yegge.ai/essays/the-shape-of-things-to-come/ :
> Gas Town was intended to be reusable, but I only ever wound up using it to build itself. Gas Town fell apart at the seams with Opus 4.7. Up through 4.6 it was working brilliantly. With 4.7 we saw the introduction of the "just two more things" tic, which prevented Opus from ever converging on being ready to do real work—it always wanted to fiddle with Gas Town itself. The Opus tic never went away, so Gas Town effectively burned down. It had other problems, too, but 4.7 was the final straw
I hadn't seen the $122K figure mentioned previously. $87K for API-style pricing was mentioned in the above post, and ~$2,800/mo for multiple Claude Max accounts:
> My solution has been to create a token tap on $200 Max accounts, which for me work out to ~30x the list-price equivalent. So in reality I'm only spending about $2800/month out of pocket for my $87k "worth" of tokens. Though that number keeps growing alarmingly.
(He never seemed to provide numbers for Gas Town initially, whether what he paid, or what the API-style pricing would have charged, so it was interesting to get some actual figures. Sounds like a lot to me, but if he's genuinely getting multiple people's-worth of work out of it, then...)
(Also in the article: a little morsel of Emacs content, which was nice to see.)
Or maybe not, but the bar was set pretty low that going all-in on AI might have been worth it.
To me this seems mostly related to the way Anthropic showed the level at which they monitor sessions, plus wanting to limit how much training data they're directly feeding into a company that competes with their own products/investments.
The cybersecuritynews.com news one simply republishes details of a story published by The Information. At least they have the decency to LINK to that Information story:
https://www.theinformation.com/articles/meta-microsoft-work-...
... and of course the Information story is behind a paywall.
I assume this is being pounced on by "I told you so" AI skeptics. Sorry but it's not what you were looking for.
I am only half joking, I heard something like "my son or nephew did this cool thing with $X so we'll take $this_radical_step because of it" enough times over my career.
If you are used to talking to opus5.5 medium, going to GPT6.1 luna low will feel like a step down. Why would any employee take the (personal) risk?
After all, they should know how to compile their software. Any automation of that process is cheating their employer.
I think that gets legally murky, if the employee is the one who pays for the tool.
If my employer is putting scoreboards to see and champion who uses a tool which costs money to use, they shall pay for the tool.
Sorry, I'm not a ladder climber, yet I'm not mindless enough to bankrupt myself.
An LLM is not too dissimilar to a Work Laptop or an IDE license.
I can't fathom the logic of paying 10k a month for a developer and giving them an 800 device.
Insane. Your leadership is severely broken. If you can, work somewhere that values you.