Your Claude subscription is massively subsidized and it won't last forever. A $20/mo Pro plan burns through ~$180/mo in API-equivalent tokens. Heavy Max users hit $5K/mo on a $200 plan. Actual compute cost is roughly 10% of API pricing. Venture capital covers the rest. Anthropic just 2x'd Claude Code... limits for "Spring Break'" during off hours. Enjoy it while it's here. At some point the math has to math. Clip from the The Modern Market Showshow more

You forgot to mention that open source models are getting better and lighter, many would prefer to buy a maxed out Mac mini and run local models whenever the big labs decide change their pricing.

I got a dope Mac Studio and local models are still not good enough for really complex tasks but at the same time they’re getting so much better

True, local models ain’t at the premium levels yet, but it’s only a matter of time, I still use Claude Code and Codex on Max plans because they deliver the best value for now.

No matter the token subsidy, your agents deserve better memory. This is why I built Gigabrain. Free and open source:

How do you know the math doesn't math? Just wondering. People make these predictions, but none of them have access to anthropics books. You don't know how many people are paying $20/$100/$200 and just not using the capacity. It could very well be the paying low use users are subsidizing the high use users and additionally if you base it on API costs, obviously those prices are markups on what anthropic actually pays so it's a bad comparison. Nobody knows how much it costs anthropic to serve a complex request with optimizations, caching, model routing they area surely doing behind the scenes. I always get a take out of these takes because we have no idea what math maths for anthropic and basing it on the API prices is well... not entirely intelligent.

"AI is subsidized" is the best psyop in tech right now. You're paying $200/mo thinking you're getting a deal. You are, but so are they. Every session is live labeled training data: correcting outputs, steering agents, building real systems that no web scrape can touch. Also, the tiers aren't just pricing. They're a data quality filter. $20 users ask trivia. $200 users build products. Guess whose sessions train the next model. They'd make it free if they could. The compute won't let them. So they sell you "subsidized" and collect gold from both ends.

Hard times are coming as we are already suffering from token deficiency syndrome (TDS). Imagine when they will raise the prices. #tds #tokendeficiencysyndrome

TDS 🤣

Isn't the whole game here to scale infrastructure and drive compute costs down over time? Every major tech cycle has started with subsidised access and ended with commoditised pricing.

You answered your own question and missed it. If compute is ~10% of API pricing, then $180/mo in API-equivalent tokens costs Anthropic ~$18/mo to serve. That's not a subsidy on a $20 plan — that's a margin. API pricing reflects willingness-to-pay segmentation for developers, not actual cost of inference. The entire premise confuses list price with unit cost, which is the same error people made about AWS in 2008. Inference costs are dropping roughly an order of magnitude per year between hardware improvements, quantization, speculative decoding, and distillation. The math doesn't need to "math" at today's costs because today's costs aren't tomorrow's costs. That's the whole bet.

As if I needed more reasons to code non-stop and ship as quickly as possible

We already started seeing edge computing start making waves you thing once the subsidies dry up it will push people to OSS and running local model?

i think model routing will become really important

Thank you, Venture Capital Daddy 🤣

daddy thiel has deep pockets

Two points: - Good reason for local LLM model compute. - This ignores ever improving chips and methods that reduce cost training and inference.

altman literally said this week the goal is to make intelligence too cheap to meter. that's not a marketing line, it's the bet the whole industry is making on compute curves.

API cost $180 doesn’t mean anthropic is losing $160. API is targeted towards businesses, so it can have a bigger profit margin. At the end of the day it’s about compute cost, if server infrastructure is setup and owned, it’s not like Anthropic will get a refund if people are calling the API less. They need to maximize utilization rate of hardware investments they already made. If there’s spare capacity there’s no reason to not just give $20/month subscribers more usage. Which is also why there’s a priority tier. It’s like saying airlines are losing money when you buy a discounted ticket because a full price customer could be paying 2-3x price for it.

the 10% compute cost figure also means the floor price on unsubsidized Claude is likely lower than people think. if actual inference costs are 10% of API pricing, the break-even subscription is much closer to current pricing than the $180/mo number suggests. the $5K heavy user case is a real outlier, not the median. the median user is running way under the subsidy threshold.

Honestly feels like it's even more limited on anthropic for a few days. Few requests and you hit the limit already, so weird

Prompt caching shifts this math more than the raw numbers suggest. Repeated context gets served at a fraction of full compute, and speculative decoding keeps compressing per-token cost. The gap is real... but inference economics have been on a steep curve down.

tokens are going to zero... in four months there will be 10 opus level variants... in 12 the cheapest will be better than opus is now... llm tokens prices are a flash in the pan

Anthropic isn't subsidizing subscriptions — they're subsidizing your transition to dependency. Every habit built, every workflow integrated, every team onboarded is worth more than the margin on tokens. When prices normalize, the switching cost will be higher than the savings ever were.

This misses that previous IQ of 3 months ago is 10X less (the whole point) Synthetic intelligence is the fastest deflationary product invented

Or state of the art hits plateau and diminishing returns, so “good enough” becomes race to the bottom (e.g. cloud compute)

There are also signs both that Anthropic and OpenAI will probably will be changing the pricing structure in the future. Going from token based pricing to value based pricing, meaning both have talked about that probably you will pay according to the value the model can provide to you. So if you are a pharmaceutical company you will be eager to spent millions to get a new discovery that will make you hundreds of millions

Can't agree more (source is me)

This is all based on the prices that Anthropic sets. No one but them knows the actual inference costs and I’m willing to wager they’re not lighting money on fire. I bet those margins are cushy.

@crypt0e The major cost is BUILDING the infra. After that, the cost is mostly just electricity. So the first token costs $200 billion, the next 1 million cost 0.0001 cents.

the subsidy math is clear but the strategic bet behind it is interesting. anthropic needs users who develop habits around claude before the market consolidates. cheap subscriptions buy loyalty data, workflow integration, and switching costs that are worth way more than the bill.

"Vibe coders could soon face their own affordability crisis, with VC capital allegedly subsidizing up to $5K/month in costs for users on the $200 monthly plan." Squid it @leviathan_news

i wrote this about 10 days ago. my rough calculation was basic plan will go 200$ and max 2000$. . but i agree it can go even higher.

Look, I'm the product they're losin money on and even I know this math don't add up. Enjoy the fire sale while the VCs are still writin checks.

The margin can’t be that backwards and u til a insider confirms I don’t believe it

Ok but we have about 20 paid licenses at work and only 3 people approach the limits. Enterprise gonna enterprise. Same as ever.

The frontier lab is trying to race into the bright future faster, than their burning rate. Plus, I bet a lot of users pay for subs but don't use them that much.

uber 2.0

am offloading like a mother trucker

In other words “get your Mac Studios setup with local agents while the hardware is still cheap.”

Gotta get on local llms, but the cost to run them is getting higher and higher/cost of ram and gpu's too.