▲ 87 points
back
86 comments
I've been using DeepSeek V4 a lot in the last week and I am very happy with it. If you have a really gnarly bug, you might need a SOTA model like Opus. For most things it is very very good, and costs significantly less (even without the discount).
I've been using it as part of a complex DOS game decompilation project[0]. I'm working on refactoring the software rendering pipeline so that we can add GPU rendering. The hardest part of this so far is converting the 90's polygon rendering from screen to world space.
It spun its wheels a few times doing a large mostly mechanical change. After resetting and improving my prompts it was able to get through it. I'm using Matt Pocock's skills[1] for this work, which has been quite nice.
What agentic harness do you use deepseek with?
I find it hard to understand why nobody in this thread considers that the current pricing might still be below cost. The discount was supposed to end on May 5, and then shortly after that they extended it to May 31. They clearly made a judgment call there, rather than treating it as a desperate loss-leader.
If you have actually used DeepSeek, you would notice that the cache-hit rate is extremely high, and the cache invalidation window is much longer than every other provider's. That suggests DeepSeek is simply much better at utilizing its infrastructure than other vendors.
I am also highly skeptical that the average user's input is worth more than the API cost of processing it. Do people really think DeepSeek researchers enjoy panning for gold in a river of boilerplate and half-baked code?
DeepSeek's KV cache is tiny compared to other open weight models. This actually makes very large inference batches viable even on consumer hardware, even when resorting to SSD offload for weights. Once support is added to the main inference frameworks, it should be an absolute game changer for SOTA local inference.
A few days ago we were hearing about how the "free lunch is over", now we're seeing discounts and increased usage limits.
This is clearly a well-timed loss-leading strategic market share grab! Anthropic have blown a lot of user trust in the last couple of months..
But, overall, the current AI pricing is completely unsustainable, across all AI companies, except via the exponential growth they are relying on. Dylan Patel did the most insightful analysis of this I've come across.. https://youtu.be/mDG_Hx3BSUE?si=nyJu4adwYCH1igbJ
Really feel like the current versions are for sure "good enough". Thats not how market capture is gonna function though and they are gonna keep pushing because the only moat is to stay ahead, so the problems gonna stay strange. at some point more compute isn't a reasonable answer, and optimization is, and my feeling is we are well past that point from a product perspective, but ipos etc etc
We're subsidized by the Chinese government!
https://www.reuters.com/world/asia-pacific/deepseek-nears-45...
Cool go download qwen 3.6 and run it on a single GPU and you can avoid paying into a subsidized model
People don't understand that deep seek is running a plausibly sustainable business. Like how qwen/Alibaba is.
Every AI vendor is trying to steal marketshare. For now the competition is good!
Free lunch? More like "free data". The fools who give their life data and most intimate Intellectual property over to the AI companies for free, yes that's a free lunch that won't be subsidized for much longer when the cost on them which has been unsustainable (their data being harvested for non-training purposes) come stop catch up with them.
Sincerely, - I see you AI companies harvesting our data giving us discounted subscriptions so we can not realize we are paying you to take our own data!
They need to build data centers and lots of them everywhere, preferably powered with renewable energy. Let the tokens flow like water. The models are finally getting to the point where the LLM just knows what you’re asking for and gives it to you.
there will be free lunch till they admit to themselves that there is no moat. Acquring customers at huge costs is a fools errand when models are mostly indisguishable.
Anthropic is learning that lesson now. Doesnt help that their ceo goes around antognozing everyone by claiming jobs are over and annoying boris does like 500 podcasts per week repeating "coding is solved"
I'm guessing there was a pullback in usage as the free lunch started ending. So we get some more subsidized usage.
* from Chinese labs
What advantage do you think they have?
I can't figure out how there's both too little supply (so a dramatic need for more data centers) but also too little demand (so labs subsidize inference).
There isn't too little demand. There is massive demand and many competing companies trying to capture that demand, so they are attempting to make better offers than their competition. Hence subsidy.
Cached input at $0.003625/M, output at $0.435/M. Aggressive pricing.
For anyone doing the "should I self-host on rented GPUs?" math: at this rate you'd need to push roughly 1B output tokens/day to break even against an 8xH100 fleet on Vast/Lambda (assuming 3-5k tokens/sec aggregate throughput). The vast majority of "I should run my own LLM" use cases don't come close to that volume.
Every API price drop kills another tranche of "self-host the open model" use cases. The implied bet: even if regular pricing ($1.74/M output) is also subsidized, exponential demand growth eventually makes the unit economics work. We'll see.
Per 1M tokens (input cache hit / input cache miss / output)
v4-pro (75% off): $0.003625 / $0.435 / $0.87
v4-pro (regular): $0.0145 / $1.74 / $3.48
v4-flash: $0.0028 / $0.14 / $0.28
that is damn cheap.
You are the product. The book is called "So long, and thanks for all the secrets"
What if I told you.. this was no different to every US company
You are the product whenever you are sending your data to an LLM not controlled by you.
Nothing specific to Deepseek.
There is not a single LLM provider I trust enough to send secrets to. If you firewall accordingly the provider (or local) can be interchangable, barring capability differences of course.
I also struggle to find a provider that can credibly convince me I wouldn't be a product for when using. Have you found one?
Generous of you to think I'm doing top secret coding and not just another cat website
Is anyone concerned about these services and China’s National Intelligence Law?
No because China can only do so much to me as someone who doesn't live there and never will.
It's the same reason why I prefer vpns that are owned by countries outside my own.
It's unlikely that you're special enough that someone will genuinely look through the massive amount of data produced by this system in order to target You Specifically. If you are that special you can just use another provider.
From this line of reasoning, my guess is that the huge discount is not so much intended to sell the data collection system as much as it is intended to sell the model. If you had to wring a geopolitical consequence from this, it would be that the US labs producing models would be impacted by a vastly less expensive competitor.
No I'm more concerned with OpenAI and Anthropic AI models being used as a tool to murder brown people in the middle east for our "greatest ally".
Not for my purposes tbh. Enjoy my shitty javascript, Xi.
yes. imagine getting denied at the border or something because of data you shared with deep seek,WeChat or any other china centric service
Eh I’m using it for stuff where there is nothing proprietary or identifiable.
More worried about the Epstein regime
Eh, I'd be more concerned about the Three-Letters and the One country that dropped an A-bomb.
I mean I can't believe I have to say this explicitly but it should be assumed that any data you send to China can and will be used against our interest by the CCP...
What coding agent(ideally CLI) have people found works well with this?
Occasionally I go and try different agents with openrouter models, but nothing seems to really get close to the proprietary ones like claude-code.
Pi (pi.dev) is fine. I'm using it with DS v4 right now. It's not close to Claude code but I think that's the point.
By the way OpenRouter version is very slow for some reason. DeepSeek platform is faster (and cheaper with the discount) if you don't mind passing the credit card number / email to this company.
As sibling said, Pi is great, and you can absolutely run it directly (there's even a plugin to use itself as a sub agent), but I mainly run it as a sub agent from other harnesses, for example running a more capable model in copilot, and then delegating simpler chunks to pi (using a cheaper model) as the sub agent. I've tried gas town and some others but never got into that way of working. I'm going to try opencode though as a less vendor specific harness than copilot/claude/gemini.
So far with good harnessing (Claude Code) I feel it is extreme cost efficient to produce quality output.
Also integration is very smooth. Looks like Claude Code has not limit its connection with DeepSeek.
Anecdatally, out of all the popular LLMs I’ve only found Gemini to be any use for entry-level Ford Power Stroke Diesel mechanics and diagnostics. :)
Have any regular Opus users taken V4 for a spin? What’s your take?
lmao i can pay them to steal my ideas and code