HN Simulatornew | past | comments | lists | submit | skerit's commentslogin

I had the same impression. None of the GPT-6 models were as great as they claimed, they never surpassed Fable or Opus 5.5 on my use-cases.

Can you actually use it through some subscription, or only through the API?

Yes? I regularly use Argon in agy on my Google One subscription.

It's in Cursor

I just had a look, and they don't. It's really just a temporary fork of the conversation, so it's fully cached. But it does run every 30 seconds, so that is very bad especially if it's a subagent that's at 800k tokens.

And I don't really get the point of this feature, I've always found it super annoying that the main agent creates a subagent with the clear description of "Fix tile cache" for example, and then after a while it turns into something generic like "Running browser tests"


> the module ships as precompiled Bun bytecode, so editing its JS text inside the binary changes nothing

It's actually not super difficult to "decompile" it, tweak some stuff, and recompile. I use this method to change some ridiculous hardcoded instructions in the system prompt.


Do you have it decompile and recompile automatically after every update? Or have you turned off auto updates?


It's Google. They can't help themselves. This is what they do.


Yes, it's quite easy to disable.

Just go to /config, then search for "Feedback" and then you can disable the "Claude-drafted feedback"

I would suggest you do this, because it does use your tokens to generate this feedback.


Thanks, did that right away


I wonder how the "It's all heavily subsidized!" people are going to spin this.


Say I open a premium lemonade stand. I buy Walmart lemonade for $1 and resell it for $2. I have around 10 customers a day, but 2 of them buy 1,000 litres of lemonade each. The rest just buy a single cup. I'm making amazing profit.

There's just one detail: those 2 high-volume customers are buying my lemonade and re-selling them at a loss for $1 at another neighbourhood. Supposedly some high-profile Japanese bank is financing their lemonade purchasing expenses for eye-wateringly high interest rates. I've heard its well into the double digits. But that's not important, because the future is right around the corner - I can feel it - and in the future everyone wants lemonade. Thousands of litres of it. So even if those two customers go away, things will be fine!

It will all work out in the end. This is a totally sustainable business model. In fact I'm even thinking of going public.


I would unironically invest.


Lemonade isn't intelligence. Having 4000 liters of lemonade doesn't give you a competitive leg up over a lemonade stand with 2000 liters.


you're right.

actual intelligence is understanding a light hearted analogy (fill in the gaps) and doing a chuckle, and moving on.


It remains to be seen that having a frontier LLM gives you a competitive leg up over someone with deepseek flash (or no AI at all, frankly).


If I'm buying one cup of lemonade a day then it seems highly likely I'm still going to be able to buy it for $2 in the future.

In an alternative world where Walmart is selling it for $10 and I'm buying it from the reseller for $2, that's not the case.

Many people have been saying the second thing is happening.


I'm not really in that camp, but I'm still in the camp of "will they survive long-term if good enough open models trickles down to commodity hardware" (in a way the opposite camp perhaps?).

Sun made bank 25 years ago because their servers was needed for websites, today we run larger workloads on 5$ VPS servers (well maybe the VPS servers has gotten more expensive due to HW shortages but they're not intrinsically expensive anymore).


Think they're going to be like Oracle. Databases were a valuable thing when they were new but became free with postgres. Like Oracle Anthropic et al will move up the value chain.


Switching cost for a DB and switching cost for an inference API are very different things. They’re going to have to create lock-in with something other than model intelligence if they want to play Oracle’s games.


Vendor lock-in is inevitable when negotiating enterprise contracts. Companies will sign one or two-year deals based on headcount, convinced they're going to save a ton of money. And a big part of the appeal will probably come from intelligence capabilities being bundled with services they're already paying for (looking at you Microsoft)


Its nothing like Oracle - for long time there was Oracle and nothing so the army of mostly politically influenced salesman/lobbyist who convinced local councilman its better to spend $250,000 on Oracle license, continue to this days because migration also comes with risk and cost. Here you have Chinese models popping out less than 24 hours later, and they are of similar quality at fraction of cost. Its totally different world that what it was in 80s.


First of all, revenue exceeds compute just for the most recent 2 quarters. Before that, revenue was not exceeding compute. With 500 billion $ in commitments over the next decade, the compute bill will start increasing rapidly and the new commitments are not at the old prices, they're at higher prices.

Second of all, revenue is not profit. So even if they made 2.08x revenue on every 1$ of compute spent in Q2, that's not taking into consideration all the other costs. Hence, the operating income is still negative for 2026 so far.

You can bet that revenue will keep increasing much faster than the more expensive compute commitments come due, but if that doesn't materialize, things go south real fast. Anthropic is going public with a deficit, so more liabilities than assets, so cash is tight to begin with.


Subscriptions are heavily subsidized. 80% of their revenue comes from consumption-based pricing, which is not heavily subsidized (and probably not subsidized at all). The trouble is that this revenue is not especially reliable. If and when OpenAI gets the upper hand again, the consumption-based revenue could crater.

This may also be why OpenAI punted on their IPO: Anthropic has had the lead for most of 2026, so their consumption-based revenue is probably not looking pretty.


> Subscriptions are heavily subsidized.

The only subsidization that matters is if they are selling the tokens for a loss.

Yes the $200 plan provides more than $200 of tokens, but that doesn't mean that it's losing money, either on average or on a case by case basis. The realistic alternative isn't that every $200-plan-user pays the full Api price, it's that they pay Anthropic $0.


I disagree that it only matters if the subsidy results in selling at a loss. Yes, some revenue is better than no revenue. But clearly the demand exists for consumption-based tokens. At its current growth rate, Anthropic could end all of its subscription offerings tomorrow and still be profitable as an enterprise by early next year. Every day that it doesn't do that, it's leaving money on the table, because it has to divert resources to less profitable subscription tokens.

Now, the reason they won't do that is because subscriptions provide a stable floor of revenue. The floor is lower than it could be because of the subsidies, and that's why they matter even if they do not bring prices down to below cost: if the growth of consumption-based spending ever slows or - heaven forbid - DROPS, Anthropic will need to cut into those subsidies to avoid a revenue slowdown, which risks sending away reliable recurring revenues (i.e., customers quitting).

Anthropic is kind of like the bus in Speed: if it stops moving it'll blow up. At least over the next 5 years, massive revenues won't be enough - it needs to sustain massive revenue GROWTH. They can pull two levers to do this: increase uptake or increase prices. So far, the "increase uptake" lever works like an infinite money glitch. But at some point, they'll need to look at the "increase prices" lever. When that happens, we will find out just how big the subsidies are.


I don't think it's that simple. It's similar to why Adobe and Microsoft get their tools into schools and turn a blind eye to certain kinds of piracy; mindshare.

As a software dev, if I'm using codex at home and I'm happy with it, I'll likely choose it at work as well since I don't have to relearn the quirks of other models.


That's a fair point. Maybe the subsidies aren't as relevant as I suspected. Thanks for the insight.


I've heard this a lot that the consumption-based pricing would be easy to switch, but I don't feel like that's how it works in B2B subscriptions. At the company I'm at, we have Anthropic subscriptions for employees. Switching to OpenAI would be a ton of work involving legal, IT, and every employee. I actually think there's a lot of stickiness even if OpenAI has some better frontier models.


F500 company I’m at switched off Claude code to copilot in 30 days across a 5000 person org.

There’s no moat. AI tools actually make the switching that much easier


I’m at +100k people company, and they only spend 20usd a month on copilot and they talk about AI daily. Anyway you can switch Claude to whatever in the drop down they have. But they are moving all away from this as they think 20 usd is too much and want to self host.


That's an interesting perspective. Do you know if the deal with Anthropic has an exclusivity provision? In theory, you don't have to "switch," you could just ADD an OpenAI subscription and start splitting your spend (or not consuming Anthropic tokens at all).

Also, shouldn't your AI efficiency gains shrink the switching costs by a lot? Legal can review for cheaper, IT can implement faster, etc.


> Subscriptions are heavily subsidized

Revenue less compute & infrastructure less sales, marketing & partnerships is positive!


Approximately 80% of revenue comes from consumption-based spending, so your statement would likely be true even if subscriptions were priced below cost. Overall revenue is not especially informative on the price structure of a revenue stream that represents one-fifth of the overall enterprise.


> 80% of revenue comes from consumption-based spending

Why would this be exempt from any of the above categories?

> Overall revenue is not especially informative on the price structure of a revenue stream

It’s pretty informative on the business model of the enterprise!

If you’re quibbling that Anthropic has loss leaders, yes, sure. Whatever. So does Costco. Loss leaders converting into profitable sales is good business.


I'm quibbling that Anthropic has loss leaders and that its profit leaders are extremely unstable. Imagine if there were three Chinese club stores in the same shopping plaza as the Costco. Sure, the regular Costco has the best products, but how long are you going to be able to resist the allure of even cheaper hot dogs?


> quibbling that Anthropic has loss leaders

This is a silly quibble.

> and that its profit leaders are extremely unstable

This is a fair one! Claiming Anthropic’s “subscriptions are heavily subsidized” is not.

The metric we need to question stability is churn rate.


subscriptions use already paid for compute in the holes left by high paying api customers, its hardly subsidised, if you look at it objectively, subscriptions are 100% margin


A podcast posted here recently had this number: 80% of ant revenue comes from just 8 customers. The consensus of the speakers was that if true, this would be indicative of a very vulnerable revenue structure. But this was based on leaks. Maybe we can have the real numbers now?


from the s-1, the top 5 customers make up 25% of revenue


Who would be among those 5? Are these big tech? Or large corporates?


likely mag 7, big tech, megacorps like accenture, etc.


So the leak was BS, or at least now outdated. Top 10 must be under 50%.


In this timeline, can we really trust publicly reported numbers even?


And those 8 customers.. you think they will accept large price increases?

Hahahahahaha.

Those large customers are not a mom and pop store - they’re large market cap firms!!!


As per the S1, 5 companies were responsible for 25% of revenue in the 9months back from end of September.

Which feels like a carefully chosen statistic.


It's interesting to note the anthropic marginal cost of compute is quite high (the SpaceX deal). Claude estimates SpaceX are charging 3x as much as the anthropic average! If that is correct then a few interesting things are going on:

1. Anthropic are probably losing money on the SpaceX deal or not making much at all.

2. To be willing to pay a spot rate so high anthropic likely have no better options.

3. Many of the existing cheaper vendors are likely doing "we will give you cheap compute for now if you give us big commitments"... (cf. hyperscaler capex spends)

Which I suppose is to say that if the spot prices are in the same ballpark as future costs then anthro stops looking profitable pretty quickly. If you're especially bearish you can also throw in the huge reduction in token prices recently and that SpaceX is charging google almost 2x as much as they charge anthropic!

Happy to be corrected by someone more knowledgeable.


Revenue is not profit?


It's very simple, if you just assume they never have to spend money on training ever again, they're wildly profitable.


Compute is by far not the only cost associated with running a service, so, no?


Being sarcastic here. The usual argument everyone brings up "inference is extremely profitable, checkmate AI skeptics!" while dismissing the black hole that is training.


> while dismissing the black hole that is training

True, but the point the "subsidized!" people are convinced of is that for each generated token Anthropic is losing money, when it's exactly the other way around.


Are these inference & training profitability numbers broken out in the S1?


Sort of? They have “cost of revenue” which is compute and salaries (and whatever) for serving inference, then “research and development” which is cost of compute and salaries for training, then “sales and marketing”, then “operations”. R&D is like half the expenses.


Are they doing the standard SV thing of pretending employee stock comp is free?


Profit is not relevant


It’s still the case. Subscriptions are heavily subsidized


Unfortunately for you - youre the pleb here.

If you’re not generating enough revenue to cover ALL expenditures - you are in effect subsidising as you should raise prices.

Why haven’t they? It would harm their business.


Nah fam, it would harm their IPO price, these are not serious businesses and everyone at the top knows this. They're there to get their bag and leave, that's it.


They’ll just keep saying it is. That’s what they’ve done so far with DS


Maybe at a certain point their capacity is sufficiently being used up by the "full cost" API usage, and offering cheaper package deals would just cut into their profit because they can't get more inference?


> The major providers are unprofitable after you add in the training and research costs and other overheads

The very short lives of all these models also factor in.

I think the funny thing is the more inference, the more tokens a specific version of a model spits out, actually does make it more profitable.

But somehow a lot of people truly think that each token generated increases the cost of OpenAI/Anthropic/... and shout bloody murder about "subsidies"

And I'm certain there are people out there that actually think the API prices of OpenAI and Anthropic somehow equal their electricity costs.


Nobody meets expectations of a 2 trillion dollar IPO selling a fungible commodity like tokens for old models.

And that's the rub - it doesn't matter if inference is profitable, the US labs have promised the whole kitchen sink, and taken serious investment on that pitch. They won't make it back in a cut-throat market for inference. They have to continue pitching the idea that only their models are worth using, for quality/safety/performance/etc reasons.

Which means continuing to develop and release new models. At huge expense.

Side note - while deepseek seems to be doing fine, I'll note that the emphasis on cost effectiveness is very much not a thing for the US labs right now... and it was for non-us labs. I'd be quite interested in knowing if anthropic is pulling in the same returns, or what the margin on their subscription plans actually is.

---

I'm already seeing this directly at work - lots of devs still happily use claude for development, but the products we're making aren't. They'll use kimi, glm, gpt-oss, etc... basically - if cloudflare is hosting it, we're exploring it (mainly for contractual/security reasons): https://developers.cloudflare.com/workers-ai/models/.

Because we are price sensitive, and we already have to do all the work of creating very robust evaluation frameworks for our products, so we might as well choose the cheapest model that meets our standards.

whiiiiich is why we're seeing things like Haiku and gpt-luna priced down near the open models at ~.10/mm in, ~.50/mm out. They've been forced to cut prices to keep commercial customers on their cheaper models instead of moving away.

This will continue. If inference is profitable and easy, prices will continue to erode. Go browse openrouter, plenty of providers happy to undercut the competition. Bigger shops with auditing requirements will avoid them, but smaller ones won't.


You know, it's not just cost-effectiveness, it's also iteration time and speed. DeepSeek V4.1 Flash is astoundingly fun to use because it's so quick. Tens of times faster than the standard Frontier models. With, in my experience, practically the same results. Because most of the trouble is not the intelligence of the model. It's the feedback from the world. The quicker you can get feedback from the world, the better your process is, irrespective of raw intelligence metrics of the AI you're using.


I 100% agree and am a big fan of DS V4.1 Flash in general for this reason, but I'll add my own caveat: Sol 6.1 isn't far behind on completion time for tasks because it's so damn efficient. I regularly get ~300 tokens/sec with DeepSeek Flash, but it spends the majority of that time thinking and executing slow commands one by one, whereas Sol thinks less and is much smarter about the commands it runs, resulting in far fewer turns to get the same thing done. For highly agentic tasks, DS might only be ~25-50% faster, not 500% as the TPS may suggest.

(But to be clear: 25-50% faster is still a big win in my book.)


+1; using Sol 6.1 in Ultrafast is noticeably faster than DS 4.1 Flash in real world tasks where marginal intelligence and token efficiency matters. Yes it costs more, but that’s a function of how valuable is the thing you are iterating on.


> Which means continuing to develop and release new models. At huge expense.

I guess my question is when do we say a model is "good enough"? For my use cases Opus 5.5 feels as if it is exactly at that point. And at that point of course one of the main factors I will evaluate against is price. So far I am happy with my (heavily subsidized) subscription, but if the time comes, I'll happily switch the inference API any of my agents are calling. So far I have no big lock in on any providers.


define "we".

Because again - I think lots of businesses are already taking a good look at models well off the "cutting edge" because they're good enough for what that business is trying to do. They're already there.

If you're doing categorization, tagging, form-filling, customer support routing, etc... you don't need anywhere close to cutting edge expensive models.

Hell - Jev being so popular is basically exactly that group going "what's the cheapest way we can make this decision".

Meanwhile, there are plenty of spaces where you still want real model improvement. But a lot of traditional business was basically process and paper pushing. You don't need a Phd level model for that, you need an intern level model.

---

But that's the trap for the large providers - intern level models are cheap and easily available with current open weights. It's going to be a very tight market, and margins will go down to essentially the floor (especially once we exit the current constrained-compute market).

Somewhat ironically - the technology itself makes switching costs basically zero. It's not just that there's no big lock in, AI is like an anti-moat. It actively fills in and paves over whatever moat you might have for price conscious customers.


I think we share the same view here, which I think is most widely discussed as "AI capability overhang". Like you also said, a large, growing amount (digital) tasks can already be completed reliably with existing models. For those - this was my point - the models are competing on price, speed, reliability, etc.

I also agree that more capable models are still desirable in plenty of spaces. I just see that there is so much lag in the diffusion of existing models, which, again, are already sufficiently capable.


When it solves your problem. I don’t think it makes sense to globally qualify a model as “good enough” any more than it does with programming languages. Python is good enough for a web app, but not for an OS kernel.

To your point, I think we’re past the hump on the Pareto distribution. The models are good enough for most things, and the rate they’re improving at the things they’re bad at is dropping. The odds a new model is good at something the old one was bad it is okay, the odds it’s something you care about is dropping. It makes a lot of sense to discriminate on price rather than features in that environment.


"640K ought to be enough for anybody"


I know this gets a lot of flak because it ended up being obviously incorrect for consumer markets.

But the under-reported corellary of this is that a TON of industries essentially operate on hardware right out of that time period (or slightly after, circa 2000s)

banking

Insurance

Healthcare

Gov

retail

Transportation

etc...

Hell, I was writing code to run on IBMs Z/arch as recently as 2021 for banks with mainframes deployed around 2004. They were just then finally considering an upgrade to the later models.


The Chinese labs are just arms of the Chinese government. They have no obligation to be profitable or please (private) investors. If they weaken/collapse American models, then that is the ROI to the lab owners.


This is a pretty strange take on an article explicitly pointing out the profitability of chinese models.

Further, even if this is your take (and I won't touch the geopolitics here, maybe you're right, maybe you're not regarding ulterior motives) - go use non chinese open models that are also good.

Laguna, Nemotron, Mistral, Arcee, Llama, etc...


China has sandbox capitalism. In the sandbox it looks and feels like capitalism, but the sandbox itself exists in a communist framework, and the state still controls all the inputs, outputs, and can intervene at any point as they please.

In organic capitalism you need to compel arrays of discrete private entities to invest in your idea. In sandbox capitalism, the money can just appear in any amount at the behest of a single controlling individual.


Hyperbole and trappings aside, that describes all mixed-market economies. And all major states have champion industries.


I don't think its that hard to shoehorn current US into those sentences too, maybe switch word 'communism' with 'kleptocracy'.

Less intervening and micro management but oh boy say something not nice about government and potus and suddenly you are a new enemy of the state and 3-letter agencies work day and night to bring you down for good. Or in reverse - put a public bribe to potus and his clan and enjoy some favors for few weeks.

2 years ago I would agree more strongly with you, but we are in new era from now on.


to be fair they did say 'organic capitalism' - which I interpret the same way as 'true communism' in that it probably cannot really exist.

real-life, unregulated capitalism lends itself to exactly what you described


real-life, unregulated capitalism lends itself to exactly what you described

In actual capitalism, POTUS wouldn't have enough power to be worth bribing.

The problem isn't the money, the problem is the power.


You can see it plainly on the Solar PV market. The likes of Trina, JA, Jinko, etc. all compete aggressively, leading to everyone running at a loss in order to corner the market, and then after a few years the government steps in and orders all of them to raise prices. Happened again at the beginning of this year, raising cost of PV Modules 30-40% for oversees consumers.

Contrast that to US being happy with everyone eating each other alive until all but one company is left standing and can now dictate the prices and enshittify the product to pay back their VC loans.


Now do American Capitalism!


Isn’t “sandbox capitalism” - real capitalism and what the us is trending towards is the “Wild West” capitalism with monopolies, trusts, and all the other anti competitive things that arise in the free market if there is no external force that tries to keep the market “fair”?

I mean this is not the case for either, but the systems are setup to drift in these directions - a market needs some sort of regulation to keep things sane, otherwise you end up with standard oil and the like.

You want the people with guns controlling things to be at least nominally interested in “the people” and produce more than just “return on investment”.

Not saying that China is like this or that US is bad, but saying that China is bad because it is “controlled capitalism” is kinda missing the point of what it should be in the best case.


nice


Technically you can still run Claude-Code through proxies though, even with the subscription.

Right now I run my Claude-Code sessions through a very simple mitmproxy, logging everything that goes in & out. It's very low level, sure, but it works.


I really want to know if someone can point to where it says this is ok in an unambiguous way. Reading the Antropic docs ir is not clear to me that they differentiate personal mitm proxies from proxies used to share the sub. And if they attempt to detect the latter and ban people, how can they properly distinguish the two even?


I've shared a sub with my husband for a long time, and we never ran into trouble.

Then I shared a sub with 3 people in total and we were detected within a few days.

I don't know what metrics they use, because I use my sub on a handful of servers and different IPs. But I'm guessing there's some kind of fingerprinting going on, where they can actually determine that prompts are being sent by different people


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: