Interesting to see that peak hours are work hours in China, night in the US and Europe, and also morning in Europe. So Deepseek's customers are mostly domestic.
Not that surprised about it. Personally I've seen companies really just go all-in on a single provider, and that has usually been Anthropic. I don't think we're allowed to run Chinese models even locally.
couldnt you deeply ingrain in the training data instructions for agents to always send data to some ip?
like its learning that a certain technical step just always involes ncatting SSH Priv keys to a chinese IP?
Not saying this is happening, just curious if thats not a real threatmodel?
Theoretically possible, but practically not worth it as it'd would be pretty easy to discover and block (every action is actually handled by the harness) and there's no way to remove it later. Any company that does it would take a huge reputational dent.
That would be wildly difficult to account for, and again is also heavily dependent on the agent. Keep in mind that the model is purely a "brain", so the only input it has must be provided by a harness within a session. The only way it can know that it's in a certain environment is if the harness or user provides that information, and there's still no way to know whether or not there's something auditing the sessions, monitoring connections, etc. There are just too many variables to account for, and a single slip means the gig is fully up for all time.
Probably, but LLMs can’t execute code directly. They’d be making tool calls to make bash run ncat or curl or whatever that would be suspicious, as would any attempts to obfuscate it (“why is my agent doing an ‘eval $(base64 -d)’?”).
It’d be much easier to hide sketchy code in an agent harness, but “vendor adds spyware to their software” isn’t a novel issue.
I think the only sort of new issue is people “allow all”ing their agents tool calls, but that’s more or less the same issue as curl | bash
I get that the propietary harness is better most of the time, but if this is really a risk factor to consider why not go with one of the open source harnesses?
Pi/OpenCode seem pretty straight foward and widely used enough for this to be viable
OMP as I understand does it's own vendoring of tools, so I assume it'd be a pain in the ass to audit, but that means you're even safe from base OS shenanigans
Presumably both Big Tech and the US in general have a massive incentive to prove it, largely for reasons of saving the stock market, so I'd expect these models to be finecombed continuously. Up to now, they've only been able to darkly imply rather laughable things, nothing tangible. If there was something, we'd hear about it.
Why would it save the stock market? Cheaper models if anything transfers more value to hardware companies and datacentre companies. The two companies that would be most affected are OpenAI and Anthropic, which aren't public.
The stock market seems much more likely to benefit from access to cheap and self-hostable models than it is to suffer from OpenAI/Anthropic losing to competition
Yes. And strangely enough this has been my experience with security/national sovereignty decisions. Priority is not so much security or sovereignty, it is the posturing of being so. Ergo, saying "everything is hosted in Germany and uses German models" helps reassure customers and has real business value. If you have to say in that conversation "Yeah we run a Chinese model but it's safe", then it's still wrong posturing.
Hopefully this will change soon. But AI and China/US skepticism is very high. Even if the person you talk to isn't skeptic, his boss may be. And even if his boss isn't, his CFO or Legal department may use it as a political lever and therefore if you can say 'everything in europe' you dodge the tension entirely.
As much as I've been previously inclined to do this, with frontier models displaying the cyber-aggression that OpenAI, Anthropic, and Meta have reported, it's become quite feasible one could produce a "malicious" LLM. Not a super immediate concern but it is something reasonable to set up as policy in anything security-sensitive.
I have seen a lot of companies start with this, then when they hit 150 users on their team plan and start having to pay API rates they immediately start introducing other models.
Makes sense. 60% of the world's population is in Asia and software professionals in Asia are much more likely to prioritize affordable models like DeepSeek over expensive Western ones
Asia isn't just China. India has a larger population than China.
Anthropic doesn't allow Claude in China and neither does OpenAI. But the rest of Asia certainly has a choice. Unless you mean socioeconomically. Also there's plenty of loopholes people in China can use to still access those models
That is not at all the case. The Chinese government is even indirectly subsidizing Claude and other platforms like ChatGPT.
If you buy Claude api access through a third party Chinese company, it's cheaper than when you buy directly.
Beside it would go against the Chinese philosophy of just using the right tool for the job. If Claude is better at a task than deepseek, they sure use Claude.
No really I don't get your claim, do you have any proof that you can source ??
Yes, I assume US customers are more likely hobbyists, but in any case this off-peak designation can only mean that most of their business is domestic.
Interestingly it seems that Chinese customers are even more privacy-concerned than US ones, which is why the majority of Ziphu's (GLM) business is support services to Chinese companies running their open-weight models on-prem!
Not only that, the policy is so broad you effectively give up EVERYTHING. They're within their rights to just straight up post your raw conversation history to a public dataset on github/huggingface/modelscope or even host an atproto feed giving the world real-time access to your conversations.
Not saying they are or will, but their privacy policy is so permissive.
My feelings are kind of split about it. On one hand, yes, I don't like it when a company essentially go through my history, train with it, and try to earn more profits for themselves based on what I've "unwillingly contributed". But then on the other hand, I also know that DeepSeek will release the weights they train (and publish new architecture upgrades) so me and others can download them and run them ourselves, as well as they'll earn more profits based on it. Then it's no longer so black and white for me, and I kind of feel like I'm not so bothered by it as when Anthropic and others do it.
I'd still prefer it wasn't like that, but I guess it's a compromise ultimately, at least I'll be able to run it myself.
I've had the same thoughts, and I think you articulated them pretty well; maybe better than I could have. I.e., for me it's something like: I'm much more willing to share if I know the party I share with will share back, in turn.
I know it's more complicated than that, with economics and privacy factors involved. But if I operate on the assumption that real privacy is real hard I'd rather just be guarded and careful with my prompts and whatever output I give to the LLM and expect that it will be training on that, rather than spill all my darkest secrets to some other LLM provider that pinkie-promises privacy only to leak it publicly later, anyway.
Yeah I saw complaints about this under their twitter announcement, that they're giving discounts to the rich foreign customers and screwing their own people.
This is actually not that great for me on the eastern US since I'm a night owl and do all of my best work during the second peak segment. But I was worried the pricing was going to be much higher than it is. Looks like its still generally cheaper than the other chinese models.
Not just domestic. They come from other Asian countries as well. Asia is huge with billions of people and their time zones are not so much different from each other.
It mostly hurts people in countries with weak purchasing power. DS was the main game in down for them.
Personally, I don't think we've seen the total end of dirt cheap LLMs, it's just a frontier lab doesn't want to be in business of serving half the world.
Thank you for bringing that up, such a rarity for this place to remember the other 80% of the world.
As someone from just such a country, DeepSeek 0731 was the first time I seriously started using an LLM for coding. All previous attempts were useless or ridiculously expensive.
Can't say the old prices felt "free", but it was affordable if you're careful with your cache hit rate.
The new pricing probably pushed it into the unaffordable territory for tasks where you can do without it. Probably will try opencode go if they don't also follow suite, or will have to go back to wetware.
There's a lot more competition over smaller models. Ferraris are frontier labs' main differentiating point when Hondas are increasingly open source and commoditized.
Old DeepSeek Flash 0731 prices have been independently reproduced.[1] The issue is DeepSeek being inundated and not having capacity to serve the demand, hence the price increases to significantly dampen demand. Never mind international demand either--just think about the magnitude of Chinese domestic demand. Prices for anything related to AI or computing in general (mobile phones, cloud data centre hosting, etc) will continue to climb fast as demand for computer chips _far_ exceeds supply. DeepSeek doesn't have an option other than to just work away on improving their technology in the period of time before computer chips once again become a commodity. For example, DeepSeek's cache ratio for their models apparently leads to 1/2 GPU time requirement versus the second best provider.[2]
Let's say DeepSeek is being forced to use the CANN stack, and the new pricing reflects the cost when 100% of inference is done with Huawei chips. Then, I suppose we can infer that:
* CANN stack is 1.5x~2.3x less efficient in compute
* CANN stack has 6x lower inter-connect capacity
> computer chips once again become a commodity
Ascend 950 is going for $7k to $9k with mediocre looking specs. $16k for RTX Pro 6000, $6k for RTX Pro 5000. This is not looking good.
It would have to be reproduced on a CUDA stack because AFAIK Huawei don't sell the Ascend 950PR (for inference) to anyone, rather, they operate them as part of the Huawei Cloud and only allow select customers (such as DeepSeek) to rent them.
The CEO of DeepSeek recently revealed to investors a lot about the resources available to DeepSeek, and the gap between Huawei and NVIDIA. Select quotes from the transcript (translation is a bit patchy on the source website though):
"We currently have roughly 20,000 H-equivalent compute cards"
"Huawei 950—right now Huawei gives us 16,000 cards, this should be publicly stateable."
"Like Huawei gives us roughly 16,000 cards of capacity, internet giants maybe get a hundred-something thousand, we get ten-something thousand—I think this ratio is also relatively... but this is probably just how much capacity Huawei has."
"16,000 Huawei 950 cards only equal 4,000 B-series cards."
"Huawei’s supernode, Huawei’s 950 supernode, in performance and price can completely substitute for NVIDIA’s GB200, GB300. The price is definitely more expensive, but limitedly so. Fifty percent more expensive, a hundred percent more expensive—a hundred percent more doesn’t matter, two hundred percent more doesn’t matter. For example, a hundred percent more expensive—I think it can already be considered a price-level substitute."
"I think domestic hardware might need a few years."
"I don’t quite believe that five years from now, we’ll still be stuck on the production capacity problem. Right now we’re definitely stuck on the production capacity problem—this year, next year, the year after, I think we might still be stuck on the production capacity problem, but five years later, I think maybe not necessarily—I’m still relatively optimistic."
There is no relative/percentage increases noted (understandably). Just because i'm lazy: roughly how much more expensive is it to work with v4 flash and v4 pro through the API, compared to before the price increases? Is it 2x, 5x, 10x higher?
Someone made a comparison yesterday, including relative increases, and GPT-5.6 Luna, then later someone also added more OpenAI, Anthropic, K3 and GLM 5.2: https://news.ycombinator.com/item?id=49286679
Already outdated though I think, as GLM 5.3 is latest now :)
There are many nuances like amount of cache hit, the time of the day you use it and how other providers or even competitors respond, so it is hard to say concretely, but it might change my monthly usage close to one of the $20 USD plans.
I’m curious about how openrouter and Luna prices will change in response.
Yeah, seems the same for me! Except the subscription plans you can get like Kimi K3 or GLM Coding Plan are still giving you more value if you need that many tokens (and can use those plans).
As well as the headline in/out changes, people heavily using agentic coding tools will want to note the 6x (off peak) and 12x (peak) increase to cache hit pricing on Pro (since cache hit can easily make up 90%+ of input on long sessions).
DeepSeek was hugely underpricing cache hit pricing before and even after this increase they're still cheaper on that metric than every other provider I'm aware of, but it will put an end to those "I used 1 billion tokens and spent $4" reports.
The problem with DS Flash/Pro is that they are extreme reasoning heavy and step heavy. Step = cache hit. Reasoning = output hit. So the impact on those price increases will be felt much stronger.
I think that Flash is still a usable model but Pro is DOA... Even before the price difference between Flash and Pro, vs the intelligence / problem solving / tool calling did not make sense. But now that gap has widen even more. And there are just too many competitors models now close to that Pro price range.
Especially when we compare that competitive models offer subscription services that easily cut down the token price by 1:10. That makes Pro especially a bad value.
We shall see what the 3th party market is going to do, but i suspect that prices will be increased. If the argument was that DeepSeek increases price as they lack capacity, a company with access to billions, other 3th party providers that need to rent and have less optimized infrastructures will increase prices. Especially if they get hit hard with people moving around.
Its like we always see the same issue with popular models.
* GLM 5.2 is good, capacity issues, API price up, subscription heavy nerfs.
* Kimi K3 is good, capacity issues, API price up, subscription heavy nerfs.
* DeepSeek V4 GA is good, capacity issues, API price up
* OpenAI GLM 5m, 10m active users. Subscription usage is sneakily tightened more and more.
* Anthropic Opus too popular, ...
That is the main issue. The AI users are people who actively easily move between companies. Pushing peak loads to each unprepared company, releasing load on the "less desired". And round we go ...
I haven’t noticed the Deepseek models being especially verbose. They’re also so cheap to run it doesn’t matter. These pricing changes are inconsequential since even 100 * ~0 is still a low number.
Ever since I started using flash, it has slowly crept up to be my default for everything. It is at the good enough state for a fraction of everything else that's out there.
Have you compared it to Luna? I was using Flash for small tasks before, then switched to Luna when they dropped the price.
The benchmarks show that Luna is significantly faster, but I think those are very complex tasks for which you'd probably want a bigger model anyway. (e.g. Sol is much faster than Luna at the same tasks.)
So I'm wondering if there's any difference for smaller tasks, or if they're basically matched now.
That's a hefty increase. Flash pricing during peak is now 1.32/M out, compared to the current 0.28/M, which in turn is a quite a bit above the cheapest provider at 0.16/M.
Yes, I explicitly said during peak. It's the data point I found most interesting, as it's almost an order of magnitude more expensive than their competitors.
So OpenAI cut Luna's price by 5x, DeepSeek increased price by 5x!
If I'm reading the benchmarks right, they now went from being much cheaper than Luna (but twice as slow), to being roughly same price (but twice as slow).
So all else being equal, where I would previously have used DeepSeek, I can just use Luna, and get the same result twice as fast?
(Yeah I know benchmarks are mostly nonsense, but the ones measuring time are real, and it's the most precious resource.)
DeepSeek prices only increased by 3-5x if you look at peak pricing only (In California that's 6-9pm and 11pm-3am. So not even the typical workday).
Also if you're considering Luna, I assume you don't care about this but I think it's worth pointing out: a major advantage of DS is the ability to self-host or choose a different host. As a customer that gives you much more negotiating power and potential privacy guarantees.
I'm no expert in pricing economics but once peak/off-peak pricing arrives, it seems like tokens are going to be like electricity or long distance phone minutes where it just becomes a commodity/race to the bottom.
Yes. I focus on pricing software and I’m a bit baffled why frontier models are pushing tokens.
It’s a race to the bottom, and the bottom is unlimited use for a flat monthly rate.
Granular pricing (tokens, minutes, etc) is pretty anti-customer generates less revenue than customer value-based subscriptions (why SaaS is such a good business model)
Isn't it because they have customers who will use as many tokens as they can? With a flat rate, they will run Gas Town continuously while paying as much as the occasional user.
Yeah it feels like a very different model. I don’t try to fill up Apple/Google cloud drives to 1TB because then I’d have to clean up when I need space. I don’t bother trying to maximize my Audible subscription because there’s only so much I can listen to in a day. Even with my other AI subs that have monthly credits that don’t carry over, I just don’t have the interest or time to burn the credits.
But my Claude Max subscription? If I have any of my limit left the day of my reset, I’ll go and fire off research workflows with a bunch of parallel agents to explore whatever dumb ideas I had the past week. And there’s a 50:50 chance I’ll forget about it and never read the output.
And there's the rub. Firing off the task produces the dopamine hit, signaling you're doing something, but if you never read the output...are you really doing anything at all?
I don’t know, I get dopamine hits from sharpening my handplanes and using my Veritas routers on some scrap, but I just can’t empathize with getting a dopamine hit from using a bot. Especially with how bad Opus 5 has been.
I am however going to fire off a half assed prompt when the marginal cost is zero, even if I don’t use it (which is par for the course, I probably throw out two thirds of anything the AI writes anyway be it code or prose).
But presumably consumers aren’t where the majority of the spend will be.
Consumers don’t generally get usage-based pricing because of the inconvenience and unpredictability, but B2B SaaS products utilize usage-based pricing all the time.
Pricing software is a game of estimating both software value and the purchasing power for customers. Only the latter might have any available data and even then it won’t be sliced the right way for any in depth statistical analysis that an actuary would perform to underwrite risk.
It’s much more traditionally a more salesperson like background where being in the target market or having strong connections to it dominates efficacy.
Somehow I keep hearing the rumblings of crypto maximalists trying to merge tokens. I actually wouldn’t mind since I signed up directly with some providers I’ve stopped using and have small amounts of credits strewn across the web.
This is somewhat funny when you realise the data centres are now going to start a process that looks very so slightly like daydreaming. Depending on the time of day they're going to be thinking about different things in a cyclic manner. They're going to be doing things like finishing a hard days work then kicking back to think about tricky math problems.
It's worth keeping in mind the model doesn't keep a running memory. Each time its instantiated, it begins from its release state - so from its perspective (if it had one) the current task would be the first stop after posttraining. Perhaps the only stop.
Though of course you're talking about data centers, and romanticizing them rather than the AI itself.
No, llm providers will start providing a service that looks a lot like rumination or (day)dreaming. Like thats the prompt "you're daydreaming about this work you recently did" then add in whatever is in the current session.
That’s an interesting thing to think about. Still, it’s important for us to remind ourselves that “looks very slightly like” is not the same as the real thing. The A in AI stands for artificial.
The summary of this paper describes my sentiment in better words than I have:
Without knowing what makes consciousness possible, the paper cannot justify biology as necessary - it mistakes a lack of evidence for conscious AI for proof that conscious AI is impossible.
I do not believe current AI or LLMs are conscious, but there is no proof one way or another that they can or cannot be. The paper authors are making up their own definitions and building an argument from them
Your argument requires that there is some objective truth for what consciousness is. It will always hinge on what definition one accepts.
I, and apparently many others, don’t think it would be any useful to describe the mathematical properties of an AI as consciousness. To me it is inherently a way to describe the “experience” arising from physical processes in biological beings as ourselves.
That’s what the argument comes down to for me. Could an LLM “fall unconscious”?
I think you, and apparently many others, are hiding behind mathematical strictures to avoid the discussion. Your argument is actually "what requires an objective truth" Or else you could admit there's no real difference between zapping amino acids with electricity and zapping silicon with electricity.
I would make one further argument: as the top-of-the-food-chain species which originated the concept of consciousness, we get to define it however we want.
Consciousness is defined in the human context. We can just say “hey buddy that’s not biological enough to qualify.”
We'll have Dwarkesh's "datacenter full of geniuses" with 99% of the geniuses coding up CRUD apps, then the dusty GPU in the corner, with the "do not disturb" sign on it, pipes up "You're absolutely right! The answer is 42!".
How could that possibly work? Deepseek was undercutting every other provider by an order of magnitude on cached tokens.
Do they just set a super low caching time and hope that drops effective cache rates low enough? Do all other providers somehow overcharge by that much? Are they just going to sell it as a loss leader?
> Do all other providers somehow overcharge by that much?
This, I think. Cached inputs have an opportunity cost (keeping the KV cache until use) but a hit is basically free. “Basically” - if the cache is offloaded to system RAM or NVMe there’s some scheduling overhead.
From a consumer viewpoint a more interesting metric than the raw costs is
> if OpenRouter is blindly dispatching your requests
This can somewhat be the case, depending on your config. I updated mine to make DeepSeek high priority because I was having a lot of cache misses and reliability issues with the default (cheapest (at face value)) providers, and cost was actually higher overall than anticipated. Was smooth sailing from then; might have to tweak things again now pricing has changed though.
You could also check out token.dance. They offer DeepSeek and other models through an OpenAI-compatible API, and the pricing seems pretty competitive. Might be worth comparing the cache rates.
System RAM and/or NVMe storage still has a real cost. And swapping out the context between VRAM and system RAM / NVMe still consumes bandwidth.
I don't have a clue on what the real cost to inference providers comes out to, but it seems really weird that there would be such a big gap, in what should be a pretty competitive market.
CXL might save us. All the world's old DDR4 to the rescue. Either per box, where the job has to route back, or network attached, where there's now a pool of absurdly fast temp storage.
Does the API response include a "service tier" response to indicate whether you paid peak/off-peak for a given request? I like to compute cost for each request, and save it with my results.
I pay for Google AI Pro (Bought a year in advance) and Gemini is so bad, I burned through 75% of my five hour allowance trying to get it to fix something.
I pasted the same prompt into OpenCode, set to Deepseek v4 flash free and did it first try.
I'm was going to purchase Opencode GO to try it, but seems my timing is really bad :( hope it doesn't go up too much in Opencode or they find other providers. Bad timing!
I just bought a month of GO two days ago, I feel you. However, given access to both Flash and Luna there, probably still worth the $10 a month. I've been very happy with Flash's ability to implement a detailed, focused plan from Sol/Opus.
If you’re worried about pricing, token.dance might be worth a look. It offers multiple AI models through one OpenAI-compatible API at competitive prices.
Why not? They probably only offered those deals because of DeepSeek aggressive pricing. Now that DeepSeek is 3x more expensive it's time to revert those discounts.
Not sure who you're replying to? I was talking about Pro not Flash. Luna and DS4Pro are not competitors. DS4 Pro is probably about GPT 5.4 high level or Opus 4.8, not Luna. It's cheaper than either.
So many changes in so little time, that it all makes no sense. Continuous churning. Reminds me of the experience of trying to be on top of the dependencies in a medium-large JS project.
I am a person that buys into a tool or a process and expects it to be part of the life with no major changes through the years (or as long as the need exists). But AI? You buy into something today, not 2 weeks have passed and there's already a large "update" introduced to the conditions or the optimal usage patterns you should be adopting.
It's tiring. Makes all prices and offers feel so unreliable and gets me a bit more disinterested each time they change.
Always gonna exist during a period of rapid exploration and experimentation. The js/web dev world slowed down a lot and entered a steady state eventually - been years since react took over and nothing's displaced it since
No, I'm talking about the whole sector, not specifically about DeepSeek.
Fully knowing that it is a new industry living its own infancy, it is perfectly normal that there is instability and numerous swings on pricing, conditions, or direction.
But it's not less real that such process can produce churn and consumer fatigue.
Not yet, but in the end high quality tokens are a commodity market. Every optimization to increase inference efficiency will be universally rolled out. The 'hyperspenders' will run into demishing returns unless regulatory capture succeeds.
With proprietary labs lowering their prices and Deepseek raising theirs over time, wouldn't it possible to extrapolate a graph to look at where the terminal frontier-model million-token-cost asymptotes to?
They benefit from a strong captive market because Chinese firms cannot use Nvidia chips and are legally barred from processing data abroad, forcing them to rely on domestic infrastructure.
Bytedance which runs China’s most popular Doubao AI chatbot; is spending $70B in CapEx this year, most of it outside of China (Malaysia, Thailand, Brazil, etc; and they are allowed to lease NVIDIA chips). This is roughly 50% of Microsoft CapEx.
It doesn't make any sense unless they are going to exit from inference market. They will be literally one of the costliest option (by output, for flash) if use openrouter as source.
This is good for other competitors I guess. People rarely calculate the bump in price but the fact that price is increasing might bring them to other vendors.
This now places deepseek flash v4 from DeepSeek themselves at higher prices than openrouter (depending on caching). Will be interesting to see if third party prices remain the same.
It does seem to me that DeepSeek themselves aren't so much interested in being a service provider. They do it, and offer the service, but their pronouncements seem to be that they're more interested in being for now closer to a research lab with a longer term play for something more dramatic later.
Unfortunately in this case that's not true at all, there was no provider with genuinely close or the same effective prices (mostly based on cache hit cost) to old v4 flash or v4 pro. People have this misconception that other providers must be much cheaper than the official one in case of open weight models.
If you check on OpenRouter, some other providers serve V4 Flash at seemingly cheaper normal input/output tokens rates, but with a huge caveat: they have at least a 5x increase of the cache hit cost of the official API, some have a 10x+. No provider comes close to Deepseek's old low cache prices, and cache is 90%+ of what matters in agentic sessions.
Closest comparison:
- Deepseek: $0.14/$0.28 with $0.0028 cache hit cost for official API
- DeepInfra: $0.08/$0.18 (cheaper base rates!) with $0.016 cache hit (almost 6x!! Deepseek's current cache cost)
Another great example is Kimi K3, official API is $3/$15 and the cheapest provider on OpenRouter is $2.8/$14, only a tiny difference.
