That is some insane value.
I've been using GLM Coding Plan Max with GLM 5.1 for a while and i've tested DeepSeek V4 Pro maybe for 3 weeks now and I found it to be better than GLM 5.1 for complex coding tasks. I've used 65m tokens and with that price it cost me $1.5, that's really cheap.
Anyone using deepseek through a gateway (not sure if right term) so there's no data retention? At work we're going through a few hundred million tokens a day in our app (using anthropic models), and we're looking for something significantly cheaper
I have been using deepseek via deepinfra, afaik they provide no data retention. Im probably going to deploy the full model on their infra instead of paying credits at some point, so far the experience has been pretty good
Using Cortecs.ai too in combination with DS4Pro and Mistral Viba as harness, but unfortunately DS4 on Cortecs is the opposite of cheap. So I just use it for privacy centric tasks.
If DS4flash works for your case, then https://tensorix.ai/pricing is offering at pretty much the same rates as deepseek themselves, with EU data residency and guarantees.
their MLA architecture cuts KV cache by ~5-13x vs standard attention. that's why inference is actually cheaper to run, not just a price war to gain market share.
That's also a game changer for local inference. It unlocks long contexts, batched inference and storing the KV cache to disk on ordinary consumer platforms.
I've "adjusted" my workflows now to use the cache. (Basically read all the files in your project very early on in your session, etc., simple stuff like that.)
Props to them. That makes DeepSeek v4 Pro extremely cheap compared to others, even in the same category. Look at these prices per million outputs tokens:
It's actually even cheaper when you look at the cache read costs. Those costs can dominate in agent workflows and DeepSeek's cost for cache reads is insanely low comparatively. At $.003626/M tokens, the cheapest other thing on your list is >$.2/M tokens. That's on the scale of 100x cheaper.
Also, deepseek cache hit rates are pretty good. I use deepseek v4 flash model regularly for agentic tasks (more than 20 tool calls on average per run), and 70%+ of input tokens get served from cache.
The speed is absolutely bonkers too. I once misconfigured a mcp I was developing locally, and told it to use the tools provided by this mcp to get certain task done. It figured out that the mcp is misconfigured, and then automatically went ahead and started to fix the mcp, fixed it, and then started using it by passing raw jsonrpc messages using stdin/out, bypassing the harness integration (since it would have needed a restart).
It did all of this in under 30 seconds and made over 15 tool calls in all of this (yes, I use yolo mode in a container, so my agents have full access to everything in the container).
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.
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.
> 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.
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.
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.
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
It went the other way, you can't use other harnesses to connect to the cheaper versions of Claude. So clearly they think their current moat is Claude Code use, not the LLM itself.
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.
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.
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
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.
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:
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.
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.
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/
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 $15-20 with OpenRouter or similar. But would be interested to hear real world usage.
But as usual, there are far cheaper subscriptions with higher limits than Anthropic and OpenAI, that also provide DeepSeek v4 Pro. So you should use those subscriptions first until you max them out, then look at a different subscription.
I don’t even use Claude that much and was hitting limits in the 20$ 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.
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.
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).
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.
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!
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...
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.
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.
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 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
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
Inference is cheap. I bet the financials of these Chinese companies are much saner looking than any of the big US AI companies which are bloated by investors.
DeepSeek is very likely selling tokens at a loss. There're many cloud providers that provide you with DeepSeek V4 Pro via API, and those services at least twice as expensive as DeepSeek itself.
I see no evidence anywhere that "inference is cheap". To my knowledge this is a myth being spread to pretend ChatGPT or Claude will one day make any economic sense.
DeepSeek likely operates at a loss. How big the loss is anyone's guess.
Meanwhile I am happy using their model. It is really good, to a point I forget I am not using Codex or Claude.
US suppliers are fine and won't go bankrupt, they can just focus on serving bigger "Pro" class models from their large datacenters. In fact cheap AI makes the bigger and smarter models more useful because it's smart enough to draft a clear question to the model, which helps minimize wasted tokens.
> US suppliers are fine and won't go bankrupt, they can just focus on serving...
For a while, US automakers thought the same of Japanese, then Korean car manufacturers, and Musk laughed at Chinese EV makers in an interview >12 years ago. People learn and get better at making things until they catch up with the frontier.
Chinese EV makers have a few interesting technologies especially wrt. batteries but they're still very far from catching up to the frontier in a general sense. From that narrow POV Musk was absolutely correct.
What the hell are you talking about? They have batteries that charge 0-80% in 5 minutes even at -30F. More full featured EVs at half the price with similar acceleration rates and higher top speeds. Total ranges are comparable or better. What is this frontier you speak of? I think the only thing US companies are far ahead on is self driving.
US providers are burning VC money because they have been selling the idea of total world domination. Even the government has bought into that. Now suddenly they are not longer dominating the field and even need uncle Sam to protect them from foreign competitors.
They can still dominate wrt. the biggest and smartest models. DeepSeek does effectively nothing to change that. Of course these big models will be served at a very steep price in order to fully and completely recoup the investment, but there's no reason why that couldn't work if they really are smart enough and if the market value of smarts follows any kind of scaling law.
If you think heavily subsidizing AI models isn’t financially viable, I have some bad news for you about US AI companies.
Deepseek has made some incredible advancements in model efficiency, and more importantly actually publishes those advancements so everyone can benefit from them.
DeepSeek hasn't raised enough money to be actively selling tokens at a loss. They have a small team, extremely low overhead relative to other labs, operate in a place with the essentially the cheapest commercial electricity rates in the world, and their architecture lends itself very well to cheap inference.
Maybe not. I don't see how US inference providers can compete anyway with commoditized models. Costs are out of control here and the infrastructure is way worse.
For sure. But also they’re building an electrostate with 100% electricity redundancy and dirt cheap electricity. They might actually be able to sustain this.
They might be thinking, we already have the servers and the GPUs sitting there anyway so why not make full use of it? They're not even close to being at a mature state where they start to monetize.
Even at these prices I find claude and codex subscriptions to be cheaper than per-token pricing when my usage is hovering around the session limits. I guess the subscriptions are heavily subsidized.
I guess I got downvoted because people don't believe me that it's cheaper? But I spent $5 a couple days ago in one hour with deepseek v4 in a coding agent. That's way more expensive than a $20/month claude subscription. Even if I hit claude's 5h limit in one hour I can do that many times in a month.
you doing probably something wrong, I used Deepseek v4 pro with opencode and in a day used 100M tokens for ~$2. Majority of tokens are cache tokens and those are extremely cheap in deepseek bordering free.
I am more worried about accidental data leak (agent reading env file for example) with the Chinese hosted models compared to the US hosted models. Am I wrong to suspect that the Chinese government might be more likely to scan all chats and save useful information compared to the US government or company?
I hesitated to even post this comment as it sounds biased and xenophobic. I would love for someone to convince me I am wrong. Does anyone have any insight into the company behind deepseek hosting, and what their history of respecting data privacy is?
User data integrity definitely should be a concern. It's also known that regulations is being outpaced, so the cost of being/using frontier products is a double-edged sword for sure.
I think there is a nonzero chance of that happening. Beijing could at any point decide that DeepSeek has become too powerful and/or is a major export and start to insert themselves (assuming they have not already).
There are widespread reports about how foreign actors (not limited to China) have infiltrated critical networks across many industries in the US en masse and are simply waiting for the right time to exploit them. Frontier models are simply another attack vector (and much more easily exploitable when you think about it).
The fact is that there is potential for this with any cloud-hosted model, whether it is intentional by the actual company building the models or a malicious actor is able to exploit a vulnerability.
It's not an unreasonable concern, which is why most US companies prefer to go with AWS bedrock, or even one of the AI labs, and typically request zero data retention agreements. But leaking is a concern no matter where it's hosted, it's just the incentives that change IMO. For example, the labs do scan every chat and train on data not covered under enterprise ZDR agreements. Law enforcement can request access to all user data with a valid warrant or in an emergency context [1]
If you're interested in trying DeepSeek V4 privately, you can try Tinfoil (tinfoil.sh) where all models are hosted in an attested secure hardware enclave, making the inference end-to-end private. Full disclosure: I'm one of the cofounders.
More likely? US tech leaders have been fully capitulating to the surveillance state for over a decade. Why do I care what China does with my data? I don’t live in China and never plan to.
The tech bro threat model has always been pure jingoism and xenophobia. Ironically, the worst thing a Chinese company has done with my data is sell Tiktok to an American technofascist.
This is a risk although then this is fortunately a model that isn't tied to Chinese hosting. But indeed something to consider if using straight DeepSeek.com.
I would not be shocked if they do that. I would not be terribly shocked that the US-headquartered models do that for another government either. As far as data confidentiality goes, I wouldn’t hold my breath. Microsoft checks all those enterprise boxes, right? Yet, Azure still gets breached once in a while.
I'm not important enough for anyone in China to go out of their way to attack me. And DeepSeek has to maintain a sufficient level of trust so that users keep using their platform--they can't just act like a keylogger attacking everyone's crypto wallets or trust collapses.
If I was working on something that the Chinese government considered of strategic importance, then I would certainly be worried about it. But I don't do that.
I'm much more worried about techbros in this country using their LLMs to extensively profile me and produce something vastly more dystopian in this country than the real or imagined social credit scores in China. The people trying to convince you that the Chinese government are the people you should be worried about (as an individual in the United States) are probably the people you really need to be worried about.
There is nothing biased or xenophobic about the fact that the Chinese government and Deepseek are functionally the same organization. There is no private business in China with any form of legal protections from the government.
Waaay too many people think China is structurally identical to the US with the only difference being the language.
Deepseek servers are CCP servers, there is no functional difference or any form of friction to keep the government "in check". In fact there isn't even a concept of "keeping the government in check".
And for the apologists who love to flood comments like this with whataboutism...look at all the shit Trump has tried to do that has been shot down or derailed. That shit doesn't happen in China. Xi Jinping has never been over ruled because that isn't even a thing that can happen there.
If he wants a team to do a daily read of chosen Americans deepseek conversations, he will have it tomorrow, and all he needs to do is say it.
I've had a ton of success when pairing Opus 4.7 for planning w/ DeepSeek V4 Flash in opencode. Best part is DeepSeek V4 Flash is Free through opencode Zen.
I've been using DeepSeek Flash to replace Sonnet once the subscription stopped working. Haven't really noticed a difference, although I don't usually have it doing anything very complicated.
Claude literally refuses to finish tasks in auto mode and just keeps saying, now is a good stopping point, when it's 1% done (and doing the EXACT OPPOSITE of what I tell it).
Codex is barely better...
May as well pay 1/20th the price for DeepSeek.
Claude seems to have something that looks at how long you've been a customer and then just massively degrades quality.
When I started my subscription, Claude had none of these problems.
2 months into subscriptions Claude is completely unusable garbage, and Codex is not much better.
They’re playing games behind the scenes to massage and manage their earnings.
China is gonna win long term there’s no doubt. The fact that the American firms haven’t created immense escape velocity despite the disparity in spending is quite telling.
The nice thing about hosting inference locally is that you can be sure you're not being rug-pulled in any way. This doesn't really help China 'win' though, it's just freeloading on them making their weights openly available.
How expensive is ram and SSDs going to be in 2.5 years? A top of the line macbook is already $10k and thats when Apple was able to purchase ram and SSds for a fraction of what is being sold for now.
That was my experience with Claude code too. Someone will come and tell you you're doing it wrong. Hard to do it right when it'll just stop randomly, especially when it ends with something like 'let me know if you want me to continue!'.
I'll keep running Flash locally for the stuff I care about data privacy, but the value of Pro through their API is unreal for anything else (and I want to give them my training data as long as they keep putting out open models).
Even with the V4 Pro discount, the V4 Flash model gives you the best performance per unit dollar, and better performance overall for agentic, tool-heavy workloads. V4 Pro is smarter in one-shot reasoning, but at a significant speed difference. The performance, cost, and speed, makes V4 Flash our top flash model today by far.
> (2) For all models, the input cache hit price has been reduced to 1/10 of the launch price. This price adjustment takes effect from 2026/4/26 12:15 UTC.
There is no end date. Currently, it's 2% of the input price for DeepSeek V4 Flash and 0.8% with this new V4 Pro pricing, which is extremely low compared to competitors to the point that it affects the unit economics a bit and I thought it would be temporary.
In the case of V4 Pro, the effective cost is ~$0.04/M input tokens given the caching (based on OpenRouter's metrics: https://openrouter.ai/deepseek/deepseek-v4-pro), which is significantly cheaper than even small models from competitors.
And their disk-based caching is amazing. I got a long 700k context session spanning more than a week, with pauses in between that was longer than a day, and some rewinds mixed in as well.
Anthropic's caching requires you to pay a $0.75/Mtok for Sonnet and $1.25/MTok for Opus as a surcharge on top of the original input token cost. It's not even automatic.
If you are reading ~8 times (8 total back and forth tool calls) that means that cache reads in some sense cost ~$0.4 / M toks (Amortizing the write surcharge over all reads).
It's really quite ridiculously expensive considering what you are paying for is some residence on a VRAM that sometimes gets offloaded to NVMe.
DeepSeek V4's KV cache is very efficient due to its heavily compressed and sparse attention architecture.
DeepSeek V3.2 which uses DSA only (sparse attention, but without compression from HCA and CSA) is a smaller model but uses 10x more memory at 1M context window compared to DS V4 Pro.
Also, I have to say, DeepSeek's API has a very good cache hit rate. With the same workload, I see ~80% KV cache hit rate with the DS API vs ~50% with the major western inference providers for open weight models.
I will testify I have used V4 Pro as a coding agent and it did a great job solving a complex problem. It worked with Pi over something like an hour, iterating and running tests. I paid API rates via OpenRouter and it cost me less than $1 I think. I've had single prompts cost that much with Anthropic. I was very impressed.
I've been extremely impressed with DeepSeek V4 flash.
We've been working on a project which can be thought of as an agent, just not for coding. So we've been building everything: agents, sub-agents, RAG, dynamic intent detection, changing models based on what's being done, etc. In our tests, DeepSeek V4-flash is the cheapest model with acceptable replies (few hallucinations, while finding the right information). It's not the cheapest one we run overall (we're actually surviving with 3B models for some tasks), but it's definitely the one powering the system and driving the main "agent".
I use it with Pi and with Gptel and I'm extremely happy about the price. The speed of deepseek-v4-pro though leaves something to be desired. I do love how detailed its chain of thought reasoning is, and it's pretty wild watching it think at ~2400 baud. It much more transparent than Gemini 3.5 flash in that regard, but maybe 4-5x slower? For my Latin language morphology and linguistic tasks it seems to be up to the job, and on the plus side I can analyze a handful of sentences parallel without worrying about breaking the bank.
I just can't get past the deepseek-CCP connection... as good as it might be I'd wonder when your machine gets backdoored by the CCP or at least your data gets stolen
The same model hosted by other providers is much more expensive [0]. So either DeepSeek can host it much cheaper than anyone else, or their business model is different. I suspect the latter, especially since their privacy policy [1] says personal data, including “User Input,” can be used "To improve and develop the Services and to train and improve our technology".
They haven't raised enough money to be selling at a loss. And selling at a loss to gain market share in an industry with zero switching friction between sellers is not a strategy. That doesn't make sense.
Loss leading only works when
- it leads to a situation that allows you to prevent competitors from selling to your customers (gilded age railroad and pipeline industries are great examples). Then you can eventually raise prices and not lose back any market share.
- or when it allows you to remarket to customers and make back the difference (selling a single console at a loss to sell a whole library of high margin videos games, or selling jet engines at a loss to lock in 30-year maintenance contracts).
Yeah, cool theory, but they are selling at a loss. We know that because their model is open and available on other providers too. No other provider even sells a quantitized version of DeepSeek V4 Pro at that price.
Also, in case of LLM, market share = more people uploading their whole codebase/legal documents/unfinished books/literally everything to your servers for you to use in future training. So the incentive to sell at a loss is much stronger than other kinds of service.
We are missing the fact that they have created their GPU's that are now just 4-5 years behind. And considering it's China, which does everything-hardware at insane scale, and efficiency, my guess is that they are at step-1 now... gain market share at loss, and at the same time, gradually, start plugging their in-house cards to power these models to gauge their performance on real workloads.
Once they cross a certain threshold, nVidia can say goodbye to it's monopolisitic profit margins of over 70%.
GPU infra capex is the biggest spend for the inference providers as of now, power, second biggest.
China has already cracked the power part, they are now close to cracking the GPU part.
Didn’t the DeepSeek team release a paper documenting inference improvements that showed they were still making a profit even under heavy discount?
Why would it be impossible for them to make a profit now, with a new model and more research?
Before DeepSeek, no one sold cheap tokens anyways and then DS showed the profit margins.
Inference stack efficiency: Many of these providers take off the shelf sglang / vllm / trtllm and hope for the best. Meanwhile DeepSeek team is known for pushing the boundary of optimizations.
Now, sglang and vllm are great pieces of software, but take DeepSeek's Sparse Attention (DSA). Introduced 1.5 years ago (https://arxiv.org/abs/2512.02556), used by DeepSeek 3.2, GLM 5, DeepSeek V4. Only now is it slowly strating to get optimized in the major inference engines: (https://github.com/sgl-project/sglang/issues/19380https://github.com/sgl-project/sglang/pull/22851 etc.). Of course, DS V4 adds extra optimizations into the model architecture on top of DSA, and those will take more time to be taken full advantage of by the open source inference engines.
Privacy: Betting that people will pay extra for inference hosted outside China. This is especially true with DeepSeek, because DeepSeek is transparent about using API data for model improvements.
And few other things (scale (matters a lot for MoEs), reliability, soft enterprise lock in, etc.)
---
There is also, likely, tacit collusion at play here. Look at GLM 5 and GLM 5.1 prices. GLM 5 and 5.1 cost the same to run, but providers decided to charge much more for 5.1 because it is much better model, and because Z.AI raised their price as well.
Another factor is that DeepSeek is not just doing inference, but also training models, so they can use underutilized compute nodes for training during off-peak hours, as described in their DeepSeek v3 article: https://github.com/deepseek-ai/open-infra-index/blob/main/20...
But I agree that the main driver is that they are really good at optimizing. They will have chosen their architecture in such a way that it will be as efficient as possible on their own infrastructure, so they have a massive head start. Inference framework developers still have to catch up.
Probably a dumb question, but looking at OpenRouter, are there really no providers outside of the US, Singapore and China offering DeepSeek? It seems like such an obvious thing for a European or other Western provider to offer. I'm sure it's a quantum leap ahead of Mistral.
I'd love to give these models a try, but I'd rather not use a provider that trains on or stores my data (beyond standard legal requirements of course).
Just checked, Crof.ai links to "Nahcrof LLC", and the terms and conditions say "These Terms are governed by the laws of the United States."
Though to be honest, I'm not sure I want to trust business workflows to a website where the only contact is a Gmail address and no physical contact address. That site looks incredibly dodgy.
I really hope Huawei ramps up Ascend production and DeepSeek open sources their optimized inference engine (they already open source a lot of their kernels -- kudos to them). This could shake things up.
DeepSeek's official privacy policy explicitly states: “To provide you with our services, we directly collect, process and store your Personal Data in the People's Republic of China.”
US companies dont sell AI services in China (as far as I know) but deepseek markets to US companies and customers.
Havoc | a day ago
Sphax | a day ago
DeathArrow | a day ago
ReptileMan | 22 hours ago
bel8 | a day ago
First accessible model with useable 1 million context window for me.
kingjimmy | a day ago
chvid | a day ago
belinder | a day ago
bel8 | a day ago
I recall reading about that in an issue or in their Discord server.
But I would contact them formally to verify that.
BeetleB | 23 hours ago
What's frustrating is that they give no information on who the provider(s) are!
mlcruz | a day ago
goobatrooba | a day ago
MaKey | 7 hours ago
wkcheng | a day ago
You don't get the discount that Deepseek is providing, but it's still a cheap model (v4-pro is cheaper than sonnet)
Phelinofist | 21 hours ago
Aldipower | 20 hours ago
freakynit | 11 hours ago
Aldipower | 9 hours ago
DS$ Pro on Tensorix. That is not exactly cheap. Input:$1.75 / 1M tokens Output:$3.50 / 1M tokens
freakynit | 7 hours ago
From what I've read online, people have reported that DS4Flash-xHigh works even better than DS4Pro-xHigh .. so, you can try. No harm in trying :)
cold_harbor | a day ago
zozbot234 | a day ago
vitorsr | 23 hours ago
trollbridge | 22 hours ago
Nearly all requests are cached now. It's amazing.
Reubend | a day ago
DeepSeek V4 Pro: $0.87
Qwen 3.7 Max: $7.50
Grok 4.3: $2.50
GLM 1.5: $3.08
Opus 4.7: $25.00
GPT-5.5: $30.00
Arcuru | a day ago
freakynit | 12 hours ago
The speed is absolutely bonkers too. I once misconfigured a mcp I was developing locally, and told it to use the tools provided by this mcp to get certain task done. It figured out that the mcp is misconfigured, and then automatically went ahead and started to fix the mcp, fixed it, and then started using it by passing raw jsonrpc messages using stdin/out, bypassing the harness integration (since it would have needed a restart).
It did all of this in under 30 seconds and made over 15 tool calls in all of this (yes, I use yolo mode in a container, so my agents have full access to everything in the container).
marksully | a day ago
onlyrealcuzzo | 23 hours ago
It doesn't matter how good Opus is if 2 months into your subscription they make it worse than GPT 3 to save money.
cassianoleal | 21 hours ago
niwinz | 24 minutes ago
gck1 | 8 hours ago
Turns out, it's possible to do the inference efficiently if you're not given permission to just burn money without constraints.
alyxya | a day ago
lambda | a day ago
hootz | a day ago
alyxya | a day ago
wolttam | a day ago
wyre | 21 hours ago
Edit: here is a really good twitter thread about this exact topic: https://xcancel.com/kunchenguid/status/2057700714626105412
satvikpendem | a day ago
apitman | a day ago
lambda | 22 hours ago
schaefer | 22 hours ago
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.
brianwawok | 18 hours ago
lofaszvanitt | 20 hours ago
tequila_shot | a day ago
ammar_x | a day ago
I tried it and it's impressive.
[1]: https://api-docs.deepseek.com/quick_start/agent_integrations...
Scarbutt | a day ago
cortesoft | a day ago
Later, they can always lock it down more or add Claude LLM only features to it.
wolttam | a day ago
koolba | 23 hours ago
wolttam | 23 hours ago
chandureddyvari | 23 hours ago
wolttam | 22 hours ago
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.
Mkengin | an hour ago
https://neuralnoise.com/2026/harness-bench-wip/
crooked-v | 23 hours ago
rane | 23 hours ago
intuxikated | 19 hours ago
HWR_14 | 21 hours ago
odiroot | 3 hours ago
thisisit | a day ago
mewse-hn | 22 hours ago
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.
thisisit | 21 hours ago
It allows for switching models in Claude Code.
mewse-hn | 21 hours ago
mvanbaak | 18 hours ago
maxdo | 21 hours ago
longsword | 20 hours ago
wiradikusuma | a day ago
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?)
wrs | 23 hours ago
flexagoon | 23 hours ago
https://mariozechner.at/posts/2025-11-30-pi-coding-agent/#to... (the pi-coding-agent section)
rjh29 | 22 hours ago
flexagoon | 20 hours ago
SturgeonsLaw | 17 hours ago
flexagoon | 16 hours ago
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.
copperx | 21 hours ago
g023 | 21 hours ago
jijji | 19 hours ago
KronisLV | 23 hours ago
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/
rjh29 | 22 hours ago
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.
0xbadcafebee | 20 hours ago
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.
iammrpayments | 12 hours ago
stavros | 20 hours ago
maxdo | 21 hours ago
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.
hbarka | 21 hours ago
BiraIgnacio | 17 hours ago
FWIW, I this is what I have in my settings.json
hawtads | 16 hours ago
ed_mercer | 15 hours ago
chewz | 9 hours ago
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
oezi | 11 hours ago
I think out tokens would be a better metric.
itsthecourier | 9 hours ago
firecall | 17 hours ago
I've been using Deepseek v4 with Cline in VS Code as a replacement for Github Copilot, and it's not been too bad.
cultofmetatron | a day ago
sunaookami | a day ago
zozbot234 | a day ago
rjh29 | 22 hours ago
zozbot234 | 21 hours ago
vrganj | 12 hours ago
zozbot234 | 11 hours ago
LaurensBER | 23 hours ago
It's not good enough to fully replace any of the frontier models yet but it's definitely great to have as a backup!
raincole | 23 hours ago
ReptileMan | 22 hours ago
Guillaume86 | 21 hours ago
- 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...
ReptileMan | 21 hours ago
Guillaume86 | 21 hours ago
wyre | 20 hours ago
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.
smoe | 22 hours ago
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
c0rruptbytes | 22 hours ago
r0b05 | 14 hours ago
maxdo | 21 hours ago
JSR_FDED | 6 hours ago
_under_scores_ | 20 hours ago
try-working | 17 hours ago
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
potsandpans | 22 hours ago
minimaxir | 22 hours ago
jack_pp | 21 hours ago
vinhnx | 18 hours ago
> https://github.com/vinhnx/vtcode
jdboyd | 17 hours ago
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.
gaolei8888 | 16 hours ago
hawtads | 16 hours ago
akritid | 3 hours ago
linzhangrun | 15 hours ago
teekert | 11 hours ago
wg0 | a day ago
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.
tequila_shot | a day ago
seemaze | a day ago
I'm not sure if it's when you run out of crypto, or when your bank gets hit by ransomeware.
jeffadelic | 14 hours ago
aqfamnzc | 14 hours ago
seemaze | 2 hours ago
SyneRyder | 7 hours ago
Either way, something interesting about that accidental misspelling. It will probably become someone's band name one day.
abyssin | a day ago
ux266478 | 21 hours ago
cosmojg | 19 hours ago
ux266478 | 18 hours ago
joewhale | 5 hours ago
customguy | 3 hours ago
No, of course not, why do you ask?
cassianoleal | 21 hours ago
intuxikated | 19 hours ago
schmorptron | 10 hours ago
margorczynski | a day ago
ecommerceguy | a day ago
tencentshill | a day ago
dyauspitr | 21 hours ago
It’s going to be hard to enforce it for most consumers though. It’s only going to apply to large corporations in effect.
That being said for coding and most actual “frontier” purposes the American models leave Deepseek in the dust.
presto8 | 18 hours ago
odie5533 | a day ago
raincole | 23 hours ago
raincole | 17 hours ago
DeepSeek V4 Pro price on OpenRouter:
deepseek: $0.435 / $0.87
baidu/fp8: $1.521 / $3.042
novita/fp8: $1.64 / $3.38
Yup. DeepSeek either has next-generation hardware that somehow no one else has access to, or they're selling at a loss.
surgical_fire | 23 hours ago
DeepSeek likely operates at a loss. How big the loss is anyone's guess.
Meanwhile I am happy using their model. It is really good, to a point I forget I am not using Codex or Claude.
zozbot234 | a day ago
overfeed | 22 hours ago
For a while, US automakers thought the same of Japanese, then Korean car manufacturers, and Musk laughed at Chinese EV makers in an interview >12 years ago. People learn and get better at making things until they catch up with the frontier.
zozbot234 | 21 hours ago
govg | 21 hours ago
dyauspitr | 21 hours ago
overfeed | 20 hours ago
throwa356262 | 22 hours ago
When VC pulls out, some of them may go bankrupt.
zozbot234 | 21 hours ago
jdgoesmarching | a day ago
Deepseek has made some incredible advancements in model efficiency, and more importantly actually publishes those advancements so everyone can benefit from them.
overfeed | 22 hours ago
I suspect American inference providers implement the efficiency gains, and pad their margins rather than pass the savings along to the consumer.
missedthecue | 23 hours ago
kajman | 23 hours ago
dyauspitr | 21 hours ago
try-working | 16 hours ago
guelo | a day ago
guelo | 22 hours ago
ReptileMan | 21 hours ago
beacon294 | 21 hours ago
pzo | 17 hours ago
sourcecodeplz | a day ago
doctoboggan | a day ago
I hesitated to even post this comment as it sounds biased and xenophobic. I would love for someone to convince me I am wrong. Does anyone have any insight into the company behind deepseek hosting, and what their history of respecting data privacy is?
nivekney | a day ago
giwook | a day ago
There are widespread reports about how foreign actors (not limited to China) have infiltrated critical networks across many industries in the US en masse and are simply waiting for the right time to exploit them. Frontier models are simply another attack vector (and much more easily exploitable when you think about it).
The fact is that there is potential for this with any cloud-hosted model, whether it is intentional by the actual company building the models or a malicious actor is able to exploit a vulnerability.
3s | a day ago
If you're interested in trying DeepSeek V4 privately, you can try Tinfoil (tinfoil.sh) where all models are hosted in an attested secure hardware enclave, making the inference end-to-end private. Full disclosure: I'm one of the cofounders.
[1] https://cdn.openai.com/trust-and-transparency/openai-law-enf...
conception | 39 minutes ago
jdgoesmarching | a day ago
The tech bro threat model has always been pure jingoism and xenophobia. Ironically, the worst thing a Chinese company has done with my data is sell Tiktok to an American technofascist.
wkcheng | a day ago
We use it that way and it works great.
rsanek | 22 hours ago
jug | a day ago
opsnooperfax | 23 hours ago
dualvariable | 23 hours ago
If I was working on something that the Chinese government considered of strategic importance, then I would certainly be worried about it. But I don't do that.
I'm much more worried about techbros in this country using their LLMs to extensively profile me and produce something vastly more dystopian in this country than the real or imagined social credit scores in China. The people trying to convince you that the Chinese government are the people you should be worried about (as an individual in the United States) are probably the people you really need to be worried about.
WarmWash | 20 hours ago
Waaay too many people think China is structurally identical to the US with the only difference being the language.
Deepseek servers are CCP servers, there is no functional difference or any form of friction to keep the government "in check". In fact there isn't even a concept of "keeping the government in check".
And for the apologists who love to flood comments like this with whataboutism...look at all the shit Trump has tried to do that has been shot down or derailed. That shit doesn't happen in China. Xi Jinping has never been over ruled because that isn't even a thing that can happen there.
If he wants a team to do a daily read of chosen Americans deepseek conversations, he will have it tomorrow, and all he needs to do is say it.
lofaszvanitt | 20 hours ago
WarmWash | 20 hours ago
velomash | a day ago
bel8 | 17 hours ago
max is really chatty for minimal gain.
dburkland | a day ago
vladgur | a day ago
npilk | 21 hours ago
onlyrealcuzzo | a day ago
RIP.
Claude literally refuses to finish tasks in auto mode and just keeps saying, now is a good stopping point, when it's 1% done (and doing the EXACT OPPOSITE of what I tell it).
Codex is barely better...
May as well pay 1/20th the price for DeepSeek.
Claude seems to have something that looks at how long you've been a customer and then just massively degrades quality.
When I started my subscription, Claude had none of these problems.
2 months into subscriptions Claude is completely unusable garbage, and Codex is not much better.
eiek | 23 hours ago
China is gonna win long term there’s no doubt. The fact that the American firms haven’t created immense escape velocity despite the disparity in spending is quite telling.
zozbot234 | 23 hours ago
onlyrealcuzzo | 22 hours ago
That's more than good enough if you're actually getting what CC Opus is capable of.
I've never been so excited for the future.
wyre | 20 hours ago
vrganj | 12 hours ago
If the Chinese model of open weights wins, AI will benefit everyone.
If the American model of closed weights wins, AI will benefit a few rich guys and everyone else will be thrown into precarity.
dawnerd | 23 hours ago
onlyrealcuzzo | 23 hours ago
I am completely convinced they just screw over their customers after so much usage or so long of a subscription thinking they have them for life.
I have NEVER been so happy to cancel a subscription.
cassianoleal | 21 hours ago
rightbyte | 10 hours ago
wolttam | a day ago
I'll keep running Flash locally for the stuff I care about data privacy, but the value of Pro through their API is unreal for anything else (and I want to give them my training data as long as they keep putting out open models).
rvz | a day ago
Remember Jevons paradox? [0] It isn't at Anthropic or Microsoft [0], but it is at DeepSeek.
[0] https://www.thelowdownblog.com/2026/05/microsoft-cancels-int...
gertlabs | 23 hours ago
Data at https://gertlabs.com/rankings
dyauspitr | 21 hours ago
tacone | 23 hours ago
vitaflo | 20 hours ago
https://api-docs.deepseek.com/quick_start/agent_integrations...
louiereederson | 23 hours ago
mmastrac | 22 hours ago
ReptileMan | 22 hours ago
minimaxir | 22 hours ago
> (2) For all models, the input cache hit price has been reduced to 1/10 of the launch price. This price adjustment takes effect from 2026/4/26 12:15 UTC.
There is no end date. Currently, it's 2% of the input price for DeepSeek V4 Flash and 0.8% with this new V4 Pro pricing, which is extremely low compared to competitors to the point that it affects the unit economics a bit and I thought it would be temporary.
In the case of V4 Pro, the effective cost is ~$0.04/M input tokens given the caching (based on OpenRouter's metrics: https://openrouter.ai/deepseek/deepseek-v4-pro), which is significantly cheaper than even small models from competitors.
maxdo | 21 hours ago
Gemini 3.5 flash : Cache Read $0.15
maxdo | 21 hours ago
And it's multi modal, and available at whatever you might imagine rates limits.
minimaxir | 21 hours ago
For Gemini 3.5 Flash, it's also 10% of input cost.
Which is why 2%/0.8% change the economics in a meaningful way, given the input/cache-heavy way agents operate.
throwdbaaway | 17 hours ago
Stats from pi:
↑400k ↓438k R432M 71.9%/1.0M
Half a billion tokens, $2.12
kingstnap | 17 hours ago
If you are reading ~8 times (8 total back and forth tool calls) that means that cache reads in some sense cost ~$0.4 / M toks (Amortizing the write surcharge over all reads).
It's really quite ridiculously expensive considering what you are paying for is some residence on a VRAM that sometimes gets offloaded to NVMe.
maxdo | 21 hours ago
Probably the most direct competitor of Flash model :
GPT 5.4 mini
Cache Read $0.075 /M tokens
Gemini 3 flash :
Cache Read $0.05 /M tokens
e.g nothing very magical or ground breaking.
freehorse | 20 hours ago
Have not actually compared it to other models, but I would not consider it in the same price range.
maxdo | 16 hours ago
wolttam | 15 hours ago
Palmik | 10 hours ago
DeepSeek V3.2 which uses DSA only (sparse attention, but without compression from HCA and CSA) is a smaller model but uses 10x more memory at 1M context window compared to DS V4 Pro.
Also, I have to say, DeepSeek's API has a very good cache hit rate. With the same workload, I see ~80% KV cache hit rate with the DS API vs ~50% with the major western inference providers for open weight models.
dyauspitr | 21 hours ago
g023 | 21 hours ago
picardo | 20 hours ago
freakynit | 11 hours ago
picardo | 7 hours ago
freakynit | 7 hours ago
zmmmmm | 20 hours ago
jorl17 | 19 hours ago
We've been working on a project which can be thought of as an agent, just not for coding. So we've been building everything: agents, sub-agents, RAG, dynamic intent detection, changing models based on what's being done, etc. In our tests, DeepSeek V4-flash is the cheapest model with acceptable replies (few hallucinations, while finding the right information). It's not the cheapest one we run overall (we're actually surviving with 3B models for some tasks), but it's definitely the one powering the system and driving the main "agent".
spudlyo | 19 hours ago
jijji | 18 hours ago
nelox | 18 hours ago
maltalex | 17 hours ago
The same model hosted by other providers is much more expensive [0]. So either DeepSeek can host it much cheaper than anyone else, or their business model is different. I suspect the latter, especially since their privacy policy [1] says personal data, including “User Input,” can be used "To improve and develop the Services and to train and improve our technology".
[0]: https://openrouter.ai/deepseek/deepseek-v4-pro/providers
[1]: https://cdn.deepseek.com/policies/en-US/deepseek-privacy-pol...
raincole | 17 hours ago
But why not? Gaining market share at a loss isn't the US's patent.
missedthecue | 15 hours ago
Loss leading only works when
- it leads to a situation that allows you to prevent competitors from selling to your customers (gilded age railroad and pipeline industries are great examples). Then you can eventually raise prices and not lose back any market share.
- or when it allows you to remarket to customers and make back the difference (selling a single console at a loss to sell a whole library of high margin videos games, or selling jet engines at a loss to lock in 30-year maintenance contracts).
raincole | 15 hours ago
Also, in case of LLM, market share = more people uploading their whole codebase/legal documents/unfinished books/literally everything to your servers for you to use in future training. So the incentive to sell at a loss is much stronger than other kinds of service.
freakynit | 12 hours ago
Once they cross a certain threshold, nVidia can say goodbye to it's monopolisitic profit margins of over 70%.
GPU infra capex is the biggest spend for the inference providers as of now, power, second biggest.
China has already cracked the power part, they are now close to cracking the GPU part.
WithinReason | 10 hours ago
oceansweep | 2 hours ago
Before DeepSeek, no one sold cheap tokens anyways and then DS showed the profit margins.
throwburn202605 | 9 hours ago
So their strategy now is to try get as much raw content for their inference. You're being "paid", via discount, for your use
amazingamazing | 9 hours ago
d4ust | 15 hours ago
Palmik | 10 hours ago
Inference stack efficiency: Many of these providers take off the shelf sglang / vllm / trtllm and hope for the best. Meanwhile DeepSeek team is known for pushing the boundary of optimizations.
Now, sglang and vllm are great pieces of software, but take DeepSeek's Sparse Attention (DSA). Introduced 1.5 years ago (https://arxiv.org/abs/2512.02556), used by DeepSeek 3.2, GLM 5, DeepSeek V4. Only now is it slowly strating to get optimized in the major inference engines: (https://github.com/sgl-project/sglang/issues/19380 https://github.com/sgl-project/sglang/pull/22851 etc.). Of course, DS V4 adds extra optimizations into the model architecture on top of DSA, and those will take more time to be taken full advantage of by the open source inference engines.
Privacy: Betting that people will pay extra for inference hosted outside China. This is especially true with DeepSeek, because DeepSeek is transparent about using API data for model improvements.
And few other things (scale (matters a lot for MoEs), reliability, soft enterprise lock in, etc.)
---
There is also, likely, tacit collusion at play here. Look at GLM 5 and GLM 5.1 prices. GLM 5 and 5.1 cost the same to run, but providers decided to charge much more for 5.1 because it is much better model, and because Z.AI raised their price as well.
gpugreg | 9 hours ago
But I agree that the main driver is that they are really good at optimizing. They will have chosen their architecture in such a way that it will be as efficient as possible on their own infrastructure, so they have a massive head start. Inference framework developers still have to catch up.
SyneRyder | 6 hours ago
I'd love to give these models a try, but I'd rather not use a provider that trains on or stores my data (beyond standard legal requirements of course).
polski-g | 47 minutes ago
SyneRyder | 40 minutes ago
Though to be honest, I'm not sure I want to trust business workflows to a website where the only contact is a Gmail address and no physical contact address. That site looks incredibly dodgy.
keithfawcett | 17 hours ago
sidcool | 14 hours ago
ReptileMan | 11 hours ago
The western models ideological bent is both heavy handed and stupidly implemented.
sidcool | 7 hours ago
Palmik | 10 hours ago
smallerfish | 6 hours ago
China sell lithium at a loss to make it unprofitable for Australian/US miners, for example (https://www.miningweekly.com/article/china-is-oversupplying-...).
ascotan | 5 hours ago
US companies dont sell AI services in China (as far as I know) but deepseek markets to US companies and customers.