China may be subsidizing this for now in a way that US companies can't or won't - but if they keep building power infrastructure and the US doesn't, then it will no longer require subsidy from them. It will simply be absolutely cheaper (including profit margin) to serve tokens in China.

China is building for the future, while Western Democracies are afraid of the future, and of their own shadow.

I'm not sure how much of it is subsidies. If the open weight models are anything to judge by, China is taking price performance seriously, and the US model vendors are looking for performance at any cost. Like any other Pareto optimization, we end up paying 10x more for the last few percent improvement on benchmark scores.

Of course, like literally every other time this has played out in computing history, the companies focused on price performance will end up with more economic resources, and get to turn the upgrade crank more often and for longer.

Also, of course, China's way ahead of the US on things like renewables, batteries, and electrification of their economy. All of that feeds into cheaper power to run the models, but I suspect it's a second order effect vs. "improve the software".

It seems to me China is chasing widespread adoption, while the US is chasing the AGI dream.

They also banned crypto mining which previously was using the free to cheap electricity, so if AI data centres are using those now under utilised supply, very possible subsidies are very low.

And yet despite the ban, China's contributions to Bitcoin mining remain very large.

https://cryptonews.com/news/china-doubles-down-on-crypto-ban...

not just renewables, also massive nuclear capacity and huge modern coal plants. They can really crank up capacity if they want to. How long will it take to get a new nuclear power plant operational in the US?

> Of course, like literally every other time this has played out in computing history, the companies focused on price performance will end up with more economic resources, and get to turn the upgrade crank more often and for longer.

The iphone is the best selling computing device in history and is among the most expensive in its category.

Most smartphones being sold are Android, though.

True, however apple makes the overwhelming majority of the profit in the smartphone market.

For most people Apple's main selling point is about showing off the cute devices and battery life, but that's not going to play a role when users are free to choose the tool that will call the models.

You might be vastly overestimating what a majority of phone owners use their phone for.

I’ve never seen anyone show off an iPhone. What a weird take.

I was talking more about laptops, but haven't you seen people sms bubble colour-shaming?

> China may be subsidizing this for now in a way that US companies can't or won't

They're subsidizing this in many ways - Huawei chips, new DDR5 memory fabs, etc.

Ultimately, DeepSeek's architecture is significantly more cost effective than anything from Google, OpenAI, or Anthropic.

Presumably, they'll incorporate DeepSeek's MLA* architecture to get all the benefits for next year's releases (if not this year's upcoming releases) which will bring down their costs...

They need to actually make money, though, so that might still not give them enough room to make enough money.

Ultimately, hardware depreciation is like 80% of total spending. So power is not as big of a deal in cost. The bigger problem is if you can get the power at all, not how expensive it is.

If you want to bring down inference costs, using less hardware is far more effective than getting cheaper electricity.

Google is in a sweet spot, because they aren't paying 80% margins to nVidia for hardware. So they're probably paying half as much deprecation as everyone else is (or maybe 1/4th for inference - which is now the biggest percentage overall).

> Huawei chips, new DDR5

The US is subsidizing in exactly the same way through the US Chip Act (as well as state level tax subsidies):

> The act includes $39 billion in subsidies for chip manufacturing on U.S. soil along with 25% investment tax credits for costs of manufacturing equipment, and $13 billion for semiconductor research and workforce training

https://en.wikipedia.org/wiki/CHIPS_and_Science_Act

> Presumably, they'll incorporate DeepSeek's MLA* architecture to get all the benefits for next year's releases (if not this year's upcoming releases) which will bring down their costs.

You can be sure the frontier labs all have similar approaches, but they just don't talk about them. That's why eg Google Flash (the old versions!) were do cheap.

I mean Google published MTP a month or so ago and it has sped up Qwen models by 1.7 times.

If that is what they still publish you get an idea of what they aren't.

What’s the TLA architecture? I haven’t read about that.

MLA, not TLA: https://medium.com/data-science/deepseek-v3-explained-1-mult...

Look up the US deficit- they have been subsidizing everything since the 1980s.

It feels like the US for years has operated under the assumption that homeostasis for the global economy would always be “designed in California, assembled in China.”

Like there was something in the American DNA that was lacking in China and innovation would always need to happen here.

But China it seems doesn’t need the US to produce great cars, devices, robotics, or AI. We absolutely need China to help us build all of the above.

Might be more far to say: they needed the US until they caught up. The massive straight up IP theft helps a lot here. Though theft might be too strong since a lot of companies knew what they were getting in to

> The massive straight up IP theft helps a lot here

I think this is vastly underestimating what "catching up" means. All my life, people have been saying "China copies". Now they are objectively better at many things (including robotics), and... well it seems that we cannot "just copy".

I saw western companies trying to "copy" superior Chinese technology, talking to brilliant engineers explaining how much they were learning by actually trying to copy.

The lesson I got from that is that China did not "copy"; they learned. And it took time, and now they are better. Now the western world has to learn from them, I guess.

Growing up moving around both conservative and liberal parts of the US, from middle school to college, I distinctly remember several US history classes where I was taught the exact same narrative about Samuel Slater. About how he was an American hero and the Father of the American Industrial Revolution because he memorized a bunch of industrial patent blueprints and brought them over to the US.

It got told as: the evil English made it illegal to even import blueprints for factory machinery, to keep the colonies in resource-extractive poverty, so they'd have to send raw materials overseas to get processed, then import the finished goods. (My other history teacher, the Anno / Dawn of Discovery video game series, also cemented this bit about resource extraction in my head at a young age.) But then thanks to heroic ingenuity and cunning, I was told, the US was able to outwit the colonizers and process its own raw materials, eventually gaining full economic, military, and political supremacy.

Sounds familiar.

It's ip theft when the Chinese do it but when it's the American copying on Chinese it's called learning.

Producing great products is a game at which every player wins, because sellers must find willing buyers. It only fails if one participant panics and jumps out of the window, or if a significant number of people are not participating (this is always the case when wealth inequality is involved).

China is out producing us at new scientists and engineers.

Ok, not my favorite narrative, but assume asymmetric application of intellectual property rights was a big factor. Wouldn't the US exploiting asymmetric labor wages, rights, and conditions be the even bigger story? It still feels like a short-sighted own goal. The US abandoned its ability to manufacture. Maybe dark factories and robotics can bring it back, but manufacturing supply chains are just so much more advanced in Asia than in the US.

> Wouldn't the US exploiting asymmetric labor wages, rights, and conditions be the even bigger story?

Yes, but "the US" is reductive. The exploitation wasn't done by the towns having their tentpole industries shipped overseas, it was done by the people shipping them overseas and pocketing the profit. US capital owners made a deal with the Chinese Communist Party that was good for both of them and bad for the US.

That's really well said.

The promise was always to get cheaper goods and services in the US, so long as the Chinese firms never competed. Guess what, they compete now.

IP theft may only be part of the story though. it’s a question of priorities. US optimizes for profit which can place limits reinvestment. China seems to optimize for ubiquity and dominance, and has the capital to throw at those goals. when you’re beholden to the shareholder/ceo/investor, you make concessions to stay within their will. when you’re beholden to the state, you do the same.

Talking about IP theft with a straight face in context of AI. lol. Not that kind of IP theft, that doesn’t count.

Lol it was not ip theft it was American and European companies building factories in China themselves teaching them how to manufacture use their cheap labour. Well they learned and as they were the dong the manufacturing got better at it. I believe the current aerospace industry which the US leads in is also result of IP theft from the British then out innovating them.

Wait until you hear about the history of US industrialization. This trope of 'they stole our ideas' needs to fade away, it's a coping mechanism based on the assumption of inherent superiority of American society rather than the natural wax and wane of civilizations due to varying structural factors.

This so much. You can also read up about when Germany sent industrial spies to Great Britain. And the first documented case of industrial spionage was against... China.

It plays this way: you're behind, you ignore IP rules. You're ahead: you create them to defend your newly-gained status.

Also please no moralizing here on IP when the entire OpenAI/Anthropic playbook has been "massive straight up IP theft". The irony.

At some point we can’t keep blaming IP theft for obvious innovation and investments being made by China.

We also can’t blame subsidy. All countries subsidize their industries.

This video on the auto industry covers a different industry but has a lot of the same rhymes as far as China’s strategy:

https://youtube.com/watch?v=UhhZu0ZHdw4

The gist of it is that China does the following:

1. Treats low margin industries like mining and utilities as areas to focus investment and come up with incremental improvements, making those available to all companies. The West, by contrast, allows private companies to handle those industries, who logically don’t bother investing in them since their investors consider those basic industries to be low-value segments of the production chain. But now we see those advantages in China where investments have been made (e.g., the best battery chemistries and mining/refining, the cheapest power (when was the last time your local utility company focused on reducing pricing?)).

2. Because all companies in China have access to the same excellent infrastructure, they must compete furiously on quality/features/price of their products.

3. China allows foreign competition so long as they operate in China (see: Tesla) further insisting that their domestic products be globally competitive and that foreign products sold in their country benefit their local ecosystem.

>"IP theft"

Can we stop this crying baby already. Every country has stolen from the other. Did you really expect countries to settle on sewing closes and ship all profits to foreign companies for eternity? The IP is just an artificial concept that participants follow for so long as it benefits all parties.

As did the big AI providers.

I would appreciate some reading pointers about this.

> Samuel Slater ... known as the "Father of the American Industrial Revolution", a phrase coined by Andrew Jackson, and the "Father of the American Factory System". In the United Kingdom, he was called "Slater the Traitor" and "Sam the Slate" because he brought British textile technology to the United States, modifying it for American use.

> He learned of the American interest in developing similar machines, and he was also aware of British law against exporting the designs. He memorized as much as he could, and departed for New York City in 1789. Some people of Belper called him "Slater the Traitor", as they considered his move a betrayal of the town where many earned their living at Strutt's mills

https://en.wikipedia.org/wiki/Samuel_Slater#Early_life_and_e...

dig into https://en.wikipedia.org/wiki/Håkan_Lans for example

What? How is someone born in 1947 relevant to ip theft in the 19th century?

'usa ip theft 19th century' in your fav search engine

Well, I did, and to save others the time, the most relevant resource I found appears to be the book "Smuggler Nation: How Illicit Trade Made America” (2013) by Peter Andreas

Sure but I think what people are actually concerned with today is China copying a product and dumping cheaply back in the country it was taken from. That scale and speed is not what was happening in the 19th century.

I personally have little issue with countries doing that for domestic use (I hate using term "IP theft"), but to re-export so quickly you can't run a viable business in your own country is not fine.

The one major area they are still behind in is CPU tech, but they are hungry and thus moving quick.

Looking at Loongsons processors for instance. About 15 years ago they coudl barely compete with a Pentium 2. Now they are about 4-5 years behind Intel/AMD. Further behind on some more specific work loads (SSL decoding for example) Not great but that is a decent jump. The jumps between generations are pretty decent.

LA446 was a decent enough processor core but had an awful memory controller that held it back as soon as it needed to reach outside of cache. As such it was SLOW.

But they learned the lesson and now the LA664 almost entirely fixed that issue. I think a big part of performance issues is that they are working domestic 5 to 7nm processes, so a good 5-7 years behind.

They are launching the LA864 later this year and are touting some decent performance gains. That is just marketing so far but something to keep an eye on.

Considering that these chips are using their own ISA, own designs, domestic manufacturing and they aren't terrible is a big thing.

I suspect in the next 5 years they have the chance of completely closing the gap. But it can also go the other way that they end up stalling as smaller nodes get much more difficult to attain.

How much does corporate espionage help them?

> Like there was something in the American DNA that was lacking in China

In most Americans' eyes, unfortunately, there was. It was just known by the name "American Exceptionalism". Yes, it's nonsense, but unfortunately it is nonsense that has historically been used by most empires throughout history, and believed just as fervently by said empires' populi since it's one of the central elements of imperialism as a whole.

The US models are still better though, let's not get carried away. Ours are better, theirs are cheaper. That's how it's always been.

people might not wanna admit it because it feels politically incorrect - but that belief is massively due the idea of "western (white) supremacy".

cz if you're smart & pragmatic - then you will know innovation can come from anywhere - but western elites choose to continually bury their heads in the sand.

> attracting the best minds around the world to a free and stable society.

China is comprised of ~91.5% ethnically Chinese citizens. [0]

> Tump voters threw it all away because they couldn't stand non-whites coming to America and doing better than old stock Americans.

The U.S. is more diverse than it's ever been [1], and under Trump we're still below the deportations of Obama's terms.

Sounds like open-borders immigration was never necessary in the first place, given that we're being beat by a country with a similar demographic skew that we had like 80 years ago. Coincidentally, when we arguably had our best economic opportunities for citizens. Who'da thunk.

Clearly, the only solution to our fading relevance is opening the border again and importing 500 million more ""doctors and engineers"" all the while China is investing in their *actual* doctors and engineers, and has extremely strict immigration policies [2].

[0] https://en.wikipedia.org/wiki/List_of_ethnic_groups_in_China

[1] https://en.wikipedia.org/wiki/Historical_racial_and_ethnic_d...

[2] https://en.wikipedia.org/wiki/China#Population_policies

You're conflating Mexican border hoppers with skilled immigrants.

I'm absolutely opposed to illegal immigration and have a more extreme position on how to deal with it than most Americans.

What I'm irked by are Trump's attacks on legal immigration and the general worsening of the environment. ICE's kidnappings, the 100k H-1B fee, and the recent Green Card thing have deeply eroded America's attractiveness to legal immigrants.

I think when MAGA came after H-1Bs, it became pretty clear that it's not about law and order, it's just a race thing.

And if you want to go gloves off, I'll just say it: the main problem in America is that its 3 major ethnic groups are infected by anti-intellectualism and slothfulness, whereas the Chinese and various other cultures are not. The direct benefit from skilled immigration is so that we can increase the ratio of people who actually value education and hard work vs the failing old stock Americans whose broccoli-headed kids dream of becoming YouTube influencers instead of astronauts.

That's such a gross misrepresentation of reality.

First of all, the only group of immigrants targeted by the admin are those critical of certain middle eastern regime.

Republican racists mainly care about the immigrants that do not take their middle-class jobs anyways.

Anti-Indian hate is restricted to a minority of software engineers and anti-Chinese hate is virtually non-existent.

I do believe it is idiotic to have your universities full of Chinese, your manufacturing in China and, at the same time, treat China as a geopolitical enemy.

Propaganda. We americans ate that shit up.

There's nothing special about anything we design in the US other than time and money commitment to create it. China did have some espionage of course going on, but the vast majority of shit isn't some secret. And with the US shitting on China with restrictions, we increasingly caused them to invest time and money into things they otherwise would have passively accepted as coming from the west. ASML sees the writing on the wall for themselves in particular.

I like this take a lot and agree with it. The US for too long has been asleep at the wheel on many areas, power generation one of them. China with no doubt has conducted very deep and sustained espionage campaigns and even with LLMs there is enough evidence that most of the initial gains was training off of western models. Again no complaints here but I think it’s important to acknowledge both which can be true at the same time.

As john oliver said on conan many years ago: "an inflatable barbecue!".

China can certainly design an inflatable barbecue. China can certainly biuld an inlfatable barbecue. But will the chinese people ever want and buy an inflatable barbecue? ... never. That is why the US will remain the premier consumer economy.

The US is the richest consumer market in the world.

And yet BYD is likely to outsell Ford worldwide this year (despite being banned in the US)

https://en.wikipedia.org/wiki/List_of_automotive_manufacture...

Remember kids, in the west it’s “investment”, in China “subsidy”

Put another way: if the average US citizen doesn't subsidize the costs of these trillion dollar companies, China is gonna come get you. Funny that you talk about being afraid of your own shadow.

I have some exposure to utility regulation and from what I can tell some of the AI companies are "good actors" and willing to shoulder some of the burden. But others are pretty adversarial and want a free lunch.

Power is foundational to pretty much everything. Cheap power is going to give China a massive advantage in everything; AI is just incidental.

Cheap power at what cost for our planet?

Not long ago we were crying death to bitcoin, it’s going to destroy the planet.

Come AI, with unlimited power demand. Everybody screaming we need more power.

We need infrastructure, clean energy, even nuclear. We are doing all in the wrong order.

China added 315GW of solar in 2025.

For context, EU added 65 and US 43.

In one year, China _added_ almost the total capacity EU has.

China is the one place where AI actually can use clean energy…

Possibly France, too.

- 70% nuclear

- 26% renewables

- 4% gas/coal

France cannot really add more of it. Not fast and cheap enough, anyway

China also has 1,271 GW of coal capacity, and is planning 500GW more.

And their coal capacity factor (ie the percentage of time they use their coal) is dropping at about the same rate.

...and China manufactured almost the totality of the EU and US solar capacity.

China is the leader in solar?

For a while already

> Put another way: if the average US citizen doesn't subsidize the costs of these trillion dollar companies, China is gonna come get you.

The future is blatantly going to be electric. Between cars, heat pumps, ranges, etc, the quantity of kilowatt hours consumed will rise dramatically per capita because they are replacing burned fossil fuels.

We don't need to subsidize the trillion dollar companies, we can settle for just not cancelling wind and solar projects, and generally updating the grid infrastructure.

A rising tide lifts all boats. If the subsidies go to common infrastructure, that's good for everyone. There's no need to complain about a road being paved because it will benefit FedEx in addition to everyone else.

All public infrastructure benefits the public but the role of our governance is to correctly prioritize. $100 billion spent on nuclear power plants is $100 billion being withheld from other critical social services.

The US could very causally spend a couple $100B less on their military and not have a real reduction in capability.

> $100 billion spent on nuclear power plants is $100 billion being withheld from other critical social services.

What? No it isn't.

There are many places the government could use to appropriate funds, not just social services. The military, for example. Other subsidies. Tax credits. Simply increasing the debt.

No, the money is not coming from a fixed box. When the US wants to do something (typically starting a new war), they never ask where the money is coming from. This tells you everything about how the decisions are made, if it is a priority for them, they will spend the money first and ask questions later. If green infrastructure was a real priority they would invest the money and later find ways to pay for it.

But those wars are typically fought to maintain the US status, to preserve its ability to debase the national currency effectively siphoning wealth from the world economy. Self-preservations comes first. I'm just describing the system.

Which of the wars in Iraq, Afghanistan, and now Iran were fought to maintain US status?

> not cancelling wind and solar projects

Tell it to the guy doing just that, as much as possible.

windmills cause cancer and kill bald eagles so we can’t do wind. /s

I believe you are right. These models are at worst a 6 month lag to the costly frontier models, but the ability to scale energy production is years ahead of where the US is. That advantage is often under appreciated

Their cost of energy is what matters vs the US as much as speed buildout.

I'm still not entirely clear on the problem <-> capability matching. E.g. it seems like Kimi K2.6 with good context would already be able to solve a huge chunk of problems. What share of prompts require frontier models?

It's not really a bottleneck. US capital is building data centers in South Asia, MENA and SEA. Many of these countries offer tax breaks because they want US data centers, and they have abundant equatorial land for solar.

You might say that US would prefer sovereignty but that's a separate argument vis-a-vis strategic competition with China in particular.

Wonder if they are finally exploring installing anti air defenses on these datacenters given they are massively expensive and devastating targets of extreme opportunities.

> China is building for the future, while Western Democracies are afraid of the future, and of their own shadow.

Yes, countries where compromise is not required, where social, capital and human costs are non-factors and where regulations are bendable at will by who's in power can be more effective at achieving some goals.

> China is building for the future, while Western Democracies are afraid of the future

who are the decision makers in china?

> while Western Democracies are afraid of the future, and of their own shadow.

Trillions of Dollars being invested against AI infra would indicate otherwise. US is in fact betting a lot of its economic future on AI.

yup - good read: https://www.thebignewsletter.com/p/the-efficiency-moat-why-c...

> while Western Democracies are afraid of the future, and of their own shadow.

Well, yeah. This is a technology that has the potential to make large chunks of the population unemployed.

Chunks of the population that took on debts prior to late 2022 with the understanding that there would be a way to pay those debts back with their labor.

> Chunks of the population that took on debts prior to late 2022 with the understanding that there would be a way to pay those debts back with their labor.

I’m calling it now, the future is indentured servitude.

> then it will no longer require subsidy from them

Is there actually a huge Chinese consumer market for these products? If not then I'm not sure how you ever actually achieve this endpoint. Chinese wages and American wages are not nearly the same thing yet.

> It will simply be absolutely cheaper (including profit margin) to serve tokens in China.

It will simply create more pollution and environmental destruction too.

> China is building for the future

That's the plan. Whether that's true requires an honest analysis.

> while Western Democracies are afraid of the future

Developed nations take fewer risks than undeveloped ones. Do you assume this pitched dichotomy will naturally sustain itself?

> and of their own shadow.

Yea, it's funny what having open and fair elections can do for a country.

You got me with fair. Gerrymandering, PACs, two-party system, electoral college.

Where do we start...

We start logically. Do you presume your handful of cases exemplify the entire Democratic system? Do you assume that "China" is best understood as a single centralized entity?

You completely walked past the argument to pick at a meaningless nit.

Handing out lessons in democracy from the record-holder country in foreign intervention (https://en.wikipedia.org/wiki/United_States_involvement_in_r...) had equal civil rights only in the 1960s, pardoned the perpetrators of Jan 6, has its supreme court in entirely political hands, and has the awesomest repressive force in the world, together with the incarcerated population to go with.

Maybe I picked like 4 meaningless nits as in: US politicians respect so much democracy that they constantly reweight "one person, one vote" to suit the interest of the incumbent, they do not have their outrageously expensive campaigns financed (legally) by private interest groups, the popular vote is represented, and elections are uncontested (unless the wrong candidate wins, where the Supreme Court promptly fixes the issue), and it has room for more than two (quite similar I may say) viewpoints in representation.

Maybe.

But please don't call “Yea, it's funny what having open and fair elections can do for a country.” an argument.

Please don't take one sentence out of a larger context and pretend it represents the argument.

Which, again, you've managed to completely ignore.

The argument, ironically in black and white, so you can sense it, "this isn't a black and white scenario and seeing it as China vs USA blinds you to the complex differences and global geopolitical forces involved."

I get that you don't personally like America, for whatever reason, but you've blinded yourself to sense in your rush to convey your rather negative and absolutely common sensibilities.

"China is building for the future, "

Meanwhile, the USA is paying for its past excesses, with interest on its debt being the number two most expensive line item in the budget.

https://fiscaldata.treasury.gov/americas-finance-guide/feder...

In the last 40 years, China has been building while the US has been wasting money and lives fighting wars. Can we learn to really put America first for once?

If you look at total debt instead of just national government debt, then China is even worse off than the US.

Article in Fortune: https://archive.is/53Vu0

> The formerly "fiscal conservatives" that I know are working overtime explaining how the debt isn't a bad thing and we can just move numbers.

Sounds like they're just catching up to what Democrats always used to say whenever a Democrat was in the White House and some Republican would complain about the national debt. "A government isn't a household, debt doesn't work the same way, you don't get it."

That's interesting, because I thought it was common knowledge that Republican presidents actually add more on average to national debt...?

They wouldn’t be ahead because they can’t buy Nvidia compute racks anymore and they don’t have EUV machines.

Blackwell is 10-20x more efficient than H200. Vera Rubin is expected to be several times more efficient than Blackwell.

The US has way more compute installed in Gigawatts because China can’t get enough chips. https://epoch.ai/blog/trends-in-ai-supercomputers

I do wonder how most Chinese employees at OpenAI and Anthropic feel about their employer constantly spreading anti China propaganda to decrease competition. Perhaps money solves almost all things so they go along with it.

This is the next phase of the OpenAI deception: give us as much money as we want or you'll be labeled anti-US and pro-China (guaranteed by the propaganda arm of openAI).

I have not fully seen or appreciated most of the negativity. Obviously there are exceptions to that but in my eyes it has largely exposed how vulnerable the west is due to poor infrastructure constructs and a lack of building out generation and transmission.

To be honest, I’m sort of annoyed that the datacenter around the corner from my home closed. It was a five minute walk on 3rd street and I know of it because we used to have so many cages there 15 years ago. Now I have to drive to Fremont.

Yes, and yes!

Yes, I want cheap clean power.

What are data centers doing? I'd never heard of anybody having had a problem with them until about two months ago.

Amusing that just when the big three AI providers from US raise prices significantly, even for the mini models, you’ve got a Chinese model slashing their already-cheap offer by 75%. Not to mention you can run this model on your own hardware, although admittedly even the flash stretches the meaning of local for individual people.

My guess is that the popular US providers get a lot more traffic and are supply-limited. No point in lowering prices unless you can serve the traffic that will result.

Yesterday I did some testing on the cost to solve the same simple problem on openrouter with different models using cline. Simple problem but it had a few nuances to solve it properly and so required reasoning.

After reading comments like this I was expecting (hoping?) that DeepSeek or similar would be cheaper.

However I was surprised that DeepSeek v4 cost about 5.5x GPT-5.4 to solve the problem.

- Deepseek-v4-pro-medium cost $2.47 - GPT-5.4-medium cost $0.45 - GPT-5.5-low was $0.86

That doesn't sound right. Were you using the actual Deepseek provider? The one time I spent 3 dollars on Deepseek in a day, I had 615k output tokens, 96M cache hit input tokens, and 5M cache miss output tokens.

Yeah, I struggle to use more than a few dollars a day using Deepseek V4 Pro (max reasoning).

* Some people suggest not using max reasoning due to overthinking and looping issues, this may consume more tokens than needed.

Nothing weird about it. It’s all supply and demand.

The US providers are at capacity limits and are increasing pricing as demand increases.

The Chinese providers are relatively unknown and not even allowed for a lot of applications. They have to cut the price just to be attractive.

Can they actually make money at these prices?

IPO metrics juicing is a bitch

Capitalist competition at its finest

Kudos to the DeepSeek folks for making tokens not only affordable but also open source. This is a race to the bottom for token costs in a good way.

Open weights aren’t open source. Source is the learning data and algorithms, and that is closed.

And this is purely a way to undercut American models. If/once they’re ahead, it’ll stop being the case. Already qwen is doing that.

I'm not entirely sold on this idea, open source models aren't really hurting Deepseek or Qwens bottom line.

99.99% of people cannot run these models on their own hardware, they are forced to rent it from someone. That someone is almost always the big China players themselves anyways.

Once they have their own coding agent which they seem to be working towards, I may start predominantly using their models. They seem to be doing all the "right" things, open sourcing models, publishing research, and keeping prices low for everyone.

You can use V4 Pro with Claude Code [1].

I tried it and it's impressive.

[1]: https://api-docs.deepseek.com/quick_start/agent_integrations...

I'm working on a custom launcher for hooking up Claude Code with various providers (groups env variables in profiles) cause DeepSeek doesn't have vision and sometimes I need browser use with screenshots or Opus reasoning, for other tasks it's fine: https://ccode.kronis.dev/

  # After installed (or when run portably with ./ccode)
  ccode init-config
  ccode edit-config
  
  # Run with default profile
  ccode
  # Run with named profile
  ccode --deepseek
  
  # Set default profile
  ccode set-default-profile deepseek

Also turns out that with a local proxy you can get Remote Control working and see the DeepSeek sessions in the desktop app, screenshots on the page. Other than that, I'm happy that it works pretty well and the discount is enough to make me consider going from Anthropic's Max subscription to Pro and using it only where DeepSeek is insufficient. With that proxy I eventually hope to be able to transparently switch models mid-task, if I need Opus for like 5 turns or something.

Overall though I'm not sure exactly how well Claude Code would stack up against OpenCode, since the latter overall feels a bit less hacky with 3rd party models and is even getting niche but nice features like a locally runnable web version: https://opencode.ai/docs/web/

I've been using V4 flash consistently with Claude. Pretty great fast and darn cheap. I use it about 3h/day and so far haven't crossed $1 USD/week.

FWIW, I this is what I have in my settings.json

  "env": {
    "ANTHROPIC_AUTH_TOKEN":"sk-nope_not_real",   
    "ANTHROPIC_BASE_URL": "https://api.deepseek.com/anthropic",
    "ANTHROPIC_MODEL": "deepseek-v4-flash",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "deepseek-v4-flash",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "deepseek-v4-flash",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "deepseek-v4-flash",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
    "CLAUDE_CODE_EFFORT_LEVEL": "low",
    "CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING": "1",
    "CLAUDE_CODE_DISABLE_THINKING": "0",
    "CLAUDE_CODE_ENABLE_AWAY_SUMMARY": "0",
    "CLAUDE_CODE_SUBAGENT_MODEL": "deepseek-v4-flash",
    "CLAUDE_CODE_MAX_OUTPUT_TOKENS": "8000",
    "CLAUDE_CODE_FILE_READ_MAX_OUTPUT_TOKENS": "4000",
    "BASH_MAX_OUTPUT_LENGTH": "20000",
    "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "60",
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "200000",
    "CLAUDE_CODE_DISABLE_GIT_INSTRUCTIONS": "1"
  }

3h/day and how many parallel agents? 1/3/10?

I think out tokens would be a better metric.

Why not use higher thinking effort?

Hi, is it comparable to Opus?

V4 Pro is between Sonnet and Opus. But it is cheap. Slow but very cheap. Very diligent.

I run a proxy that allows me switching back to Opus when necessary.

Deepseek isn't like Z.ai which is bit cheaper only on the surface. Or like Qwen 3.7 Max which is Opus-level but very expensive.

Deepseek is my favorite since V3 but V4 is definitely catch-up to newer Anthropic models

thank you so much for sharing ir

How does the cost compare using the API vs the $20/month plans with other providers?

I did some back of the envelope calculations and it seems like you would pay $5/month using DeepSeek directly or $15-20 with OpenRouter or similar. But would be interested to hear real world usage.

It is still more expensive per-request than the common Anthropic and OpenAI subscriptions, but the math changes a lot based on your specific use case. https://codeberg.org/mutablecc/calculate-ai-cost/src/branch/...

But as usual, there are far cheaper subscriptions with higher limits than Anthropic and OpenAI, that also provide DeepSeek v4 Pro. So you should use those subscriptions first until you max them out, then look at a different subscription.

I don’t even use Claude that much and was hitting limits in the 20$ using sonnet, I’ve deposited 5$ with deepseek and haven’t hit the limit after spending 60million+ tokens. So no way it’s more expensive.

The link you shared is just a large table of data, which is hard to browse on a phone.

Could you please elaborate on the far cheaper subscriptions that we should be using?

I've been using it pretty extensively over a month and I'm at maybe $7. It thinks for quite a while, but the results have been better than Sonnet for me.

I'm not curious what tasks you tested it for. Im working on coding agent writing code dynamically on request for customers. i'd say code itself very simple and aggressively cached, and patternalized, e.g. we adding lots of hints to the system.

the only real family models that work were claude and openai, surprisingly, for tasks that needs faster speed, gpt 5.4 is very impressive. Deep seek was very average , doing things somewhere in gemini flash 3.0 domain.

I am curious - Is there a way to switch between models depending on the task? Because I believe Deepseek V4 is not multimodal and it will be good to switch back to Claude if vision or other capabilities are required.

I was looking into something similar because I wanted to test a local model for doing basic coding and smart model (deepseek) for planning.

It's basically not possible with claude code, the api endpoint is a single environment variable and whatever models are on that endpoint are what's available.

HOWEVER, if you run a proxy like LiteLLM, you can configure it to send requests to different api endpoints on the back end and expose them as different "models" on the front end, then configure claude code to switch between those virtual models.

Found this: https://github.com/farion1231/cc-switch

It allows for switching models in Claude Code.

Right that says it has a proxy feature so it can probably do what I was describing with LiteLLM

Check out the project called superpowers. It can use different models for different agents. I use it witb opencode to have different models for reaearch, planning, execution, testing etc

There is a tool called deepclaude, which runs a proxy in the background capable of doing this, by simply doing /model in Claude.

i've been trying that, in reality every time you try to save it, it's not worth it, the cost of mistake is so high , you can spent 2-3h on just wrong assumption, you lost your time and all the burned tokens.

It seems you can use the Claude Code CLI harness without a Claude Pro subscription now, which I don't think you could a before?

I've been using Deepseek v4 with Cline in VS Code as a replacement for Github Copilot, and it's not been too bad.

I'm my experience claude code is kind of shit.

Pi works very well with deepseek though

The npm install of Claude Code deprecated, since Feb 2026.

Surprised Anthropic hasn't done anything to restrict Claude Code from using other providers.

At this point in the AI wars, it is probably better to have more users of Claude code rather than restrict which LLMs it can connect to. Claude code is probably (currently at least) stickier than the LLM model itself. Getting people into the Claude code ecosystem is worth it.

Later, they can always lock it down more or add Claude LLM only features to it.

The value of Claude Code the harness isn't that great. There's a lot of other good harnesses out there.

I thought so, and then I tried Opencode and Codex and started to appreciate Claude Code a lot more. They've actually done great work with the small details.

I actually have't looked back since trying opencode The ability to properly see what the agent is doing in tool calls and subagents is really unmatched, CC strips all reasoning and return values, only displaying tool calls, and you're unable to expand a single subagent, it's expand everything and scroll endlessly or show everything collapsed with basically no info at all (read x files, ran x commands) Just seems like extremely basic features are missing

And it gets dragged down by Anthropic actively injecting unhelpful things into prompts without telling users about them (https://github.com/anthropics/claude-code/issues/58262).

What’s your favourite harness? Is there any benchmarks for harness like LLMs have for swe verified?

There Seen to be more and more harness benchmarks out there, pretty interesting read:

https://neuralnoise.com/2026/harness-bench-wip/

You can check my profile for which one I like most :) I do think there have been efforts to benchmark different harnesses.

Personally I'm not going to choose one harness or another based on +/- a few percentage points in a benchmark. I'm going to use one the one that I find the most ergonomic, that isn't too bloated, etc. The models are the primary lever, not the harness.

Good or better? Curious which would be in either bucket.

Probably a matter of taste. I prefer the harness I wrote, I don't want to go near Anthropic's bloated mess of a harness with a 10-meter pole.

IMHO the ergonomics of their tooling are not great. I'd rather use Codex or even OpenCode. Configuration alone is very arcane with lacking documentation. Sandboxing/permission system is quite confusing too.

That's interesting. I thought Claude Code is not as good, therefore people want to use Claude model with other alternatives. This is the other way around.

Which begs the question, regardless of the model, which Claude Code alternative is better? (I keep saying "Claude Code alternative" because I don't know the term... LLM CLI?)

AFAIK the two most popular open source harnesses right now are OpenCode and Pi. They take a pretty different approach, OpenCode includes a lot of features while Pi is very minimal by design and focused on extensibility, to the point where many people are just asking Pi to write a plugin for itself whenever they want it to have a new feature. I personally like Pi's philosophy more and I think its developer justified the choices really well in his blog post:

https://mariozechner.at/posts/2025-11-30-pi-coding-agent/#to... (the pi-coding-agent section)

Author blocks referrals from HN, weirdly dramatic, especially considering they have 1086 karma here. I wonder what we did to them.

Oh damn, I haven't noticed because my browser removes the referer header. But I think the image on the block page is a pretty good answer to why he did that.

What's the image trying to convey? Genuine question, I just come here to read nerd stuff and I'm not aware of any controversy

The image shows Garry Tan, the CEO of Y Combinator. He has lately been on a huge AI psychosis streak, bragging about things like "shipping 37000 lines of code every day" and "using Claude Code so much it burned out his USB-C power connectors". He's in a lobster suit because he's talking about OpenClaw, an AI agent assistant which those same AI psychosis types lean into too much by giving it full read-write access to all their life and then getting surprised when it accidentally deletes all of their emails.

Pi's developer is obviously not anti-AI, and he definitely doesn't hate OpenClaw, since it's based on Pi. But there's a growing number of people who take those things too far, and a lot of them are on HN. You can easily find them in the comments of any AI-related post here. I assume that's the type of people the image is portraying.

The common term for a tool that wraps an LLM with a workflow is “harness”.

I've seen good results with opencode connected to glm 5.1 on ollama cloud... for $20 a month you get similar performance that you get with opus 4.7

I love oh-my-pi, but I'm not sure if it's "better". Maybe just as good.

I use DeepSeek v4 flash with CoPilot and it works pretty good.

It works very well with OpenCode. My team keeps hitting the 5h limits on other subscriptions and it's pretty good to have Deepseek as a backup. I just put 50 bucks on there and it feels like it'll never run out.

It's not good enough to fully replace any of the frontier models yet but it's definitely great to have as a backup!

Why do you need them to provide a coding agent? Just use their model with any off the shelf coding agent. I happen to prefer Pi, but use whatever works for you.

I probably have an unfounded assumption that whatever coding agent they make will work really well with their models, better than external harnesses. I don't have a good sense for how all the model + harness combinations compare, nor any good way to compare them myself, but generally believe model companies train their models to work best with their own harness.

I've noticed that models have gotten less finicky with this over time. Harnesses don't need to be complex to get good coding performance from models, they just need to implement some sane primitives for code exploration and editing.

It is in the model's provider's interest for you to believe this because they get to lock you into their harness and inference. As models get better they will get better at using any harness, it comes down to how well the harness is actually engineered. I highly recommend you take an hour or two and check out Pi to either solidify or change your assumption. The harness is essentially just another developer tool and can be as opinionated, overly-engineered, minimal as anything else. I would think for DeepSeek, especially, they're efforts are much better spent researching how to make their LLM's better instead of working on engineering a harness that might get some marginal gain building it for their models.

Edit: here is a really good twitter thread about this exact topic: https://xcancel.com/kunchenguid/status/2057700714626105412

Yeah, I'm using Pi with their models through an OpenCode Go subscription and it works pretty well. 10 bucks and V4-Flash is virtually infinite.

What's the best way to use it with Pi, OpenRouter?

> What's the best way to use it with Pi, OpenRouter?

I can't claim it's "the best"...

But the Pi.dev and OpenRouter combo is what I'm doing at home, and I love it. Setup was easy, I can use /model to switch between any of the openrouter models and whatever I'm hosting locally via VLLM.

Open router is a 5% tax? If you use it seriously may as well skip it

I don't have an LLM-positive culture at work. I'm on a bit of an island. Or under a rock.

Anyhow, I'm pulling myself up by my own bootstraps.

For me a 5% overhead is fine... if it gives me better visibility of this rapidly moving field.

I only use local models myself personally. But yeah, OpenRouter would probably be a good option.

Qwen cli

RL with the harness inputs and outputs of users is one of the primary improvers of model performance, a self perpetuating flywheel.

Earlier this week I started testing Chinese models on my codebase. I haven’t really looked at interactive coding yet, but more at issue triage, bug auto-fixing, log analytics, etc.

I used DeepSeek, Kimi, GLM, Qwen, and MiMO against GPT-5.5 high as reference, all running in Pi harness without anything installed.

So far, Kimi and MiMO look the most promising to me. I haven’t tested them rigorously enough to make a strong statement, but my first impression is that, in practice, all those models may be less behind on typical daily tasks than people think.

They are a bit “work hard, not smart". Getting to same-ish results more slowly and using more tokens, but at a fraction of the price

I just did a little comparison using benchmarks for GPT 5.1 through 5.4 to map out the equivalent capability-level of some of the Chinese models.

Based on these benchmarks, here's a rough mapping:

- Qwen 3.7 ~= GPT 5.3

- Kimi K2.6 ~= GPT 5.15

- DS V4 ~= GPT 5.1

So yes, we have GPT 5 at home now. No need to pay the Legacy Labs anymore.

Here's the benchmark I used since I can't post images here: https://x.com/trydotworks/status/2058004995195490706?s=20

I switched to predomentantly using mimo this week, mostly out of curiosity to see how dependant I was on frontier models. Honestly I cant really tell the difference. I would say I work on pretty average codebases with well know frameworks doing pretty typical things and initial impressions is that mimo, kimi and deepseek can probably handle what I need more or less the same as gpt5.5 or claude.

I personally really like DS4 Flash - it's the largest I can run locally with decent speeds and I feel like it's good enough to maintain a codebase with less effort

What hardware and quant do you run it with?

maybe i need to give it second chance, surprisingly Kimi 2.6 consistently fail even to generate valid json plan, where gemma 4 was doing really good, but slow.

Are you going through OpenRouter or direct? I’ve had nothing short of excellent results from Kimi.

And not letting you opt out of being their training data.

I would prefer a coding agent to be somewhat independent of the model provider. Providers are trading off on quality, features, and price so frequently, and I don't want to keep changing my agent every time.

I am looking forward to things slowing down and stabilizing. I'm not saying that should happen today, just I am looking forward to it.

I think this will happen much sooner than we thought. Maybe it will happen in next 6 months

You can take Codex today and ask it to rewrite itself to work with any API

There is OpenCode and Pi, they both work pretty well

You no longer need "their coding agent". You can hook up claude code to use Deepseek. Works perfectly.

Zed's Agent natively supports a DeepSeek API key now. (do not use it through OpenRouter if you want to save the most cost)

Give pi a try if you haven't already. Avoid vendor harness lock-in.

You can use DeepSeek with my coding agent VT Code. Recently I've added DeepSeek V4 Pro and DeepSeek V4 Flash support with all providers, via: Official DeepSeek API, HuggingFace, Ollama Cloud, OpenRouter providers.

> https://github.com/vinhnx/vtcode

antirez's ds4-agent works quite fine. It runs on any Apple Silicon device with 96GB RAM or more.

I wonder how many years it'll take for the API token cost to exceed the money spent on ram.

The DS4 folks are unofficially testing ways to run the model with lower performance on lower-RAM machines. Similar efforts are going on with llama.cpp. The results are a bit of a challenge, prefill time tends to explode which is a limitation if you care about agentic workflows.

Anything that runs with 64?

You can just try it yourself, it will probably run with a heavy slowdown using SSD offload.

All the major coding agents already support DeepSeek.

open code works with them today. I've been using it fulltime for 2 weeks so far.

Using it with Pi and can only report good thing so far. I'm very impressed by how good it is (also it's way slower than Claude Sonnet and GPT-5.5 and often thinks "too much" before starting).

Why not OpenCode? Genuine question, not an expert..

Both pi, opencode and zed work amazing with deepseek.

You seem to have tried a few things, if you don't mind I have a few questions as someone currently on Claude Code but would prefer to not lock myself in a commercial ecosystem (and their pricing change regarding headless usage is annoying me):

- how do/would you add the WebSearch tool to your harness? pay for a separate service or does deepseek offer something with their subscriptions?

- do pi/opencode support pasting images in prompts?

- how do you handle reading images? deepseek is not multi modal IIRC? do you pay for another model and route to it?

Any of these missing would really annoy me in day to day use...

Brave, Exa, and Tavily all offer a free tier for websearch, after that it comes out to like 1¢/search, very easy to ask pi to build a web search tool using any of these providers.

They support image locations like a file or url, but not regular images (opencode desktop might though?)

Both pi and opencode make it very easy to change models so you can easily call to 5.4-mini or whichever multi-modal LLM for reading images. I'm sure you could even create a skill to automate the process too, having the model use the cli to send the photo to the multi-modal and give it back a description.

I use them for pure coding, but I think they do curls when needing something from the host machine.

Yes I'm also using it for coding: I often make the agent use WebSearch in the research phase when deciding on a stack or a library or research best/modern practices to do achieve something. As for images I find it super useful to be able to paste snipped screenshots to show the agent when something is wrong in a UI/frontend or just something I can't copy paste easily.

there already is a open-sourced deepseek-tui coding agent. besides, you can always connect to opencode.

i have done some amazing things for 5 dollars, using opencode. give it a shot, it is incredibly cheap

If you have not tried DeepdeekV4 you're missing out. The pricing makes it unbelievably good.

The chains of thought for Deepseek are very very interesting reads. Open code won't show them but do read them and you'll be surprised at how underrated the model is.

My model usage is very low but I still do pay directly to Deepseek regularly as my tribute and contribution to them open sourcing their models as my gratitude and showing support for what I deem positive for overall social good.

It’s good and cheap, but don’t talk about politics to it or it might trigger some sort of censorship rule. You can see it think, then suddenly erase everything and suggest to switch to another subject, without explaining anything. I also had it output some sort of generic message about how the news outlets are in the service of the people. Both times I was surprised because I didn’t make any sensitive requests, neither illegal nor subversive. But it was a remotely political topic and it was enough. There was something both chilling and refreshing about it, since censorship in the west is usually more subtle.

How is the censorship when asking for "dangerous" actions like writing a port scanner?

The base model doesn't have these problems FWIW

How are you running the base model?

vLLM in a docker container, FP16 quantized on an 8x MI300X cluster. Very lazy hackjob, I didn't even set up an interface. Was constructing curl commands from string templates. I worked out if I paid that compute cost over a whole month, it was twice as expensive as the monthlies you'd pay for owning a very nice 2000sqft non-coop apartment in Midtown Manhattan. I was paying rock bottom prices, too.

do most people use llms to chat about politics?

I once was doing unrelated work to politics, but Qwen3 said that my input had homonym characters to those in the name of Xi Jimping. It gave the strangest response twice before finally admitting the source of the error and allowing me to convince it that no, I was not trying to intentionally misspell the man’s name.

"do most people use llms to chat about [specific thing]?"

No, of course not, why do you ask?

Yes - the model is REALLY good. I try Claude at work and Deepseek personally and this is the only model that works without trying to actively bankcrypt me.

Perhaps unintentional, but I find 'bankrypt' to be a thoroughly interesting portmonteau.

I'm not sure if it's when you run out of crypto, or when your bank gets hit by ransomeware.

Ironically you spelled portmanteau incorrectly. OP very well could have made a similar error for bankrupt. Maybe not, interesting to think about.

I'll be honest, I carefully scanned your comment for a similar mispelling.

its misspelling all the way dwon

I thought of it as crypt in the sense of "underground vault that acts a a burial place". So, not just ensuring you're bankrupt but with maybe a chance to start over, but "bankrypt", so bankrupt that they make sure you're buried.

Either way, something interesting about that accidental misspelling. It will probably become someone's band name one day.

Opencode absolutely will show you. You just have to toggle “Expand Reasoning”

I live V4 Pro for certain things but I've been quite impressed with V4 Flash for coding. It's terse, to the point, tends to make few mistakes and is pretty fast.

same here. gpt-5.x medium was my default for coding and v4 flash (max) has completely replaced it. it's the first open source model that made me feel like I could just let rip and not worry any more than Claude or GPT.

When planning small-to-medium sized changes, I found that it was a little bit faster than GPT-5.5 (high) and produced equivalent results. on large changes its results were fine but GPT's were more thoroughly thought through. DS v4 beats the absolute pants off GPT when it comes tone and style though.

Reasoning display can be toggled in opencode