▲ 795 points
back
475 comments
I've been using it extensively since the release and the best summary I can give is that it's good enough to use it for (almost) everything and cheap enough that the cost are irrelevant. I'm running it in Oh My Pi with a second instance running as "advisor" and even with 5-6 active sessions (effectively 12 streams) I'm struggling to spend more than 5 bucks per day.
OpenCode Go even has double limits temporarily so for 10 USD you effectively get 140 USD of tokens to spend. It would impress me if someone could burn that amount with "normal" usage. Even when running multiple sessions.
I have a Claude Max subscription but I've barely touched it, it just feels like a step back to have to think about limits and usage even though the models are stronger.
The beauty of intelligence at this cost (even if it's not SOTA) is that it opens a whole bunch of new use cases. Test failure in CI? Have the bot automatically propose a fix, its cheap enough that you can discard it w/h issues. Test coverage too low? Auto generate tests on CI for every pull-requests! Monitoring server logs, continuous security audits and investigating every received exception now becomes possible.
I'm thinking about having it automatically filter and re-rank my social media feeds so I can steer the algorithm instead of the other way around.
Perhaps other people (with enormous budgets) were already doing all of the above but for us this is a really exciting release!
If what you're saying is true and accurate, then US-based AI labs are in big trouble. The only saving grace might be some sort of a 'national security' proclamation banning the use of state-of-the-art Chinese (and non-US) models across US federal and state governments and large enterprises (especially ones with federal government contracts), but even still, US AI labs will probably lose out massively on international market if a smaller model can match SOTA of just a few months ago.
There's no way large companies outside the US will pay the "US AI lab" premium if they can get the same workloads done at a fraction of the cost using open-weight models that they can self-host and optimize/fine-tune on.
How is $5/day irrelevant? In the $150/mo range you can get effectively unlimited usage of GPT 5.6 Sol (Pro plan). Why use a much weaker model for the same price?
But DeepSeek now has a warning they’re going to sharply increase their API pricing sometime in the future.
I kept running into it looping two nights ago or else getting trapped in a reasoning loop it couldn’t escape from. Switching to Pro helped, but I ultimately had to use GLM-5.2 to recover my session. (GPT-5.6-Sol’s cybersecurity guardrails went off since the problem I was trying to fix involved a race condition where it would segfault and the words “stack frame” in my session made it decide I was being naughty. Yet another reason not to use American models…)
Another time, Flash started trying to make tool calls by just calling bash and catting the tool call to stdout. Then it started running echo xx for every two letter UNIX command it could think of: mv, cp, etc and the it dug into uv, ty, and jj
> it's good enough to use it for (almost) everything
which in your case is?
>even if it's not SOTA
And, probably 99.99% of people using LLM probably don't even need SOTA anyway.
Hear hear. IQ tokens to cheap to meter upon us. So many things changed since last week. Now I've had Prime agent session grinding into its 20-th hour still not giving up. Been using opencode-go since Go sub appeared. What made a difference was deepseek-v4-flash and mimo-v2.5 showing. Very similar middling models ~300b so light on the gpu. 1M context and hybrid archs - so one can actually make use of that 1M (don't grind to a halt like others). In OMP I have one the primary (default), the other one as /advisor looking over the shoulder and nagging. On opencode-go in credits counting they are the bottom-2 in cost, cheaper by 200-350 times than than the top-1. Last week with deepseek-v4-flash-0731 another jump - now it's closer to the top models then to the middle. Now I don't even need the /advisor probably. Still left it there it's sometime amusing the models back and forth. :-) DeepSeek offer /v1/responses api now with flash-0731, so setup Codex to use that too. I'm loving this :-)
Hard to make big predictions, but it sure looks like at least this level of capability is going to be available in the open and relatively cheap to run.
The 'floor' has gone up: today's model a bit behind SOTA is like model releases that were blowing people's minds a few months ago. Compared to, say, DS R1, this is far out stuff.
This, Luna, and (if it's good in practice) Laguna S are also fast and light not just cheap. And, as happened before, DeepSeek's first but other open model makers likely follow.
And a small, fast model taking small steps is...fun? More like working with code.
It's very impressive and I'm running it locally on 2x DGX. Non thinking mode is very responsive. Thinking mode has some latency but can be switched on when needed. Both are really good
My Claude account was banned the other day. The only possible cause I can think of is that I tried to authenticate from the AI assistant in a JetBrains IDE and, not thinking, entered the details for my regular subscription rather than an API account. As soon as it became apparent that I needed an API account rather than a subscription, I just closed out of the tab. Nevertheless, about 20 minutes later I got an email saying my account was banned for a violation of the usage policy, and my appeal was rejected.
My initial thought was to sign up for ChatGPT, but I had $20 in OpenRouter so I've been trying out DeepSeek V4 Pro with Pi for the last few days and I gotta say, it's good enough for my use case. And even with paying for API usage rather than Claude's subsidised subscription, and with OpenRouter taking their cut, I will probably end up paying significantly less overall. And I really like the flexibility of being able to use whatever minimalist open source harness I want (and being able to switch providers easily, too).
(My demands probably aren't as high as many others' - I mostly use it for help with some hobbyist coding projects, and I tend to ask it questions about how to approach problems rather than just telling it to go off and code stuff for me.)
Note this is the 07/31 release of DSv4 flash and not the "preview" that they put out a couple months or so ago.
I've been running this model locally for a week, and the preview version before that. This updated one feels like a whole tier up. It's very capable for debugging and analyzing documents/data I upload.
The killer feature, IMO, is the speed. On 2x RTX Pro 6000 Blackwell, its ~8k tok/s prefill and ~250 tok/s on a single stream. I saw 1000 tok/s with ~64 concurrent streams on vLLM.
That's fast enough that you can interactively chat with it without switching tabs while you wait, and its a ~300B (13B active, hence the speed) model so the responses are also very good. It's actually more convenient now for me to direct 95%+ of my day to day usage to my local model, and only use Claude Fable for really big coding tasks.
Until this model was released, I was contemplating spending even more money on hardware to run GLM5.2 (~750B) at reasonable speeds, but I no longer feel that need. This is smart enough, and I think it only gets much better for local models from here.
I strongly recommend trying this for programming tasks.
It is strong (not Fable strong though) with a much better “persona” than Opus, and very different blindspots. If you flip between Claude and this you will find both catch the mistakes of the other before they get out of control.
On balance I actually prefer DeepSeek for programming now, because of the way it talks.
DeepSeek has announced an upcoming "significant increase" in price, so this line may have to move to the right soon. https://api-docs.deepseek.com/quick_start/pricing/
Kimi K3 was an interesting model only a month ago, and now we're looking at the same performance for 1/20th of the price. Wild how fast this is advancing.
Compared to the last Deepseek V4 Flash version I've had tons of issues with it getting in infinite loops and talking to itself without executing tool calls, wasting tons of tokens
This is on Pi agent, nothing fancy at all about my prompts or use case. Anyone else experiencing this?
I've also had it randomly go from talking about Rust to talking about the electric chair, controversies about D&D rules (both irrelevant and something I've never discussed) and it's completely blind to it in future prompts even when its pointed out and referenced directly
All this said its still worth it but the agentic performance has degraded in my experience at least
The recently announced they're raising their prices 10x right?
Which would put them... exactly where everyone else is on this graph.
Edit: I seem to have misunderstood the news. I thought the magical cache read pricing was going away (0.002) and they were going to be on par with everyone else (0.02). But I have no idea.
Edit 2: Apparently, neither do they!
>We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected. Please plan your usage accordingly. The specific pricing plan will be subject to official notice.
It's really amazing to see how the gaps between the self hostable models and the closed models has been shrinking in the last 24 months.
And how this has been accelerating!!
I felt this very hard when I had to travel in the middle of nowhere in south america, with no network, and wanted to keep an LLM model on my macbook pro with 48GB of RAM. That was back in April 2026, a few months ago.
I downloaded Google Gemma 4 (google/gemma-4-26b-a4b) and - Oh boy - I was amazed by it's capacity!
I was able to use it to code simple things, ask it about nature, learn new stuff while traveling and make stories for the kids.
Was really amazing to observe and experiment this!
Seems to me there will be some good chance to run these great LLM locally on our hardware!
Amazing time to be alive
I'm not sure about all these benchmarks, I did some very simple tests (I have my own benchmarks https://upmaru.com/llm-tests) and these models fail, not sure if it's the inference provider or the model. They seem to be optimized for benchmarks more than real use cases. Do anything outside their distribution (even if it's not complex) they fail.
I Compared Deepseek V4 Flash 0731 (low) to Gemini 3.5 Flash Lite (minimal) and GPT 5.6 Luna (no reasoning) and Deepseek V4 Flash 0731 gets it wrong alot, where as Gemini and 5.6 Luna just gets it done.
Last weeks's discussion (591 points): https://news.ycombinator.com/item?id=49120299
It's not frontier, but it's far past what we had at the beginning of the year. It's very usable. I get great instruction compliance, tool calling, and with a trivial workflows flow it has very good long-running performance as well.
Oke of the great advantages of v4 flash 0731 is that even in the largest size unsloth quantized gguf, Q8 K XL, it will fit well within the resources of a 256GB DRAM server. If you have no gpu at all and are okay with setting up a workflow that handles slow token per second rate, give it a task and check back in 4-6 hours, it works great. And remember to give it more lengthy tasks to run overnight. Whatever workflow you set up, the idea is to keep it busy 24x7 doing different things in parallel.
Well, I'm impressed.
Someone else here said we could get the model via OpenCode Go for $10/mo and get about $120 worth of credit, so I decided to give it a whirl. My first month is actually $5.
In the 3 hours I've been using it I've burned 3% of my 5-hr, 1% of my weekly and 0% of my monthly.
It's fixed 3 or 4 issues in my C++ game, despite not having visual capabilities to see the screenshots I was trying to give it. One-shotted them too.
Luna struggled with what I thought was an easy task (had to replace a few ASCII chars with the correct unicode char but kept choosing incorrectly).
I'll keep using it.
we got this running on 4 RTX Pro 6000's and for single request we're getting around 250 tok/s we can support about 48 concurrent requests we're seeing around 2400 agg tok/s peaking around 24-31 concurrent users. Model performance feels like gpt 5.4 - mostly using it with pi agent. the only thing i'm missing with this model is vision and i see some folks have done some work like https://huggingface.co/webbrain-one/DeepSeek-V4-Flash-0731-V... but have not yet tried it out.
Seeing everyone spend like 200USD a month seems kind of mad.
I have £20/month Gemini and £20 a month claude for a bunch of personal projects.
Yes I have to wait sometimes, it's probably a good thing.
For the last 3 months I've been using V4 Flash Free with Hermes through Opencode Zen both personally and at my company and I've been having a great experience so far. It's my go-to model for terminal work, managing my entire ubuntu server, Cloudpanel, managing static websites, doing SEO audits, network tests, DNS troubleshooting, e-mail deliverability troubleshooting...
Furthermore, in my company we are using MCPs for Google Ads (it manages our ads), Analytics, Search Console, Zoho CRM, Microsoft Clarity... We use it to crawl specific websites and send daily summaries to our sales team in MS Teams channel. We use it to send daily summaries on marketing statistics and analytics... All with a FREE model. We are rarely hitting any limits so far and in case we need more tokens - we use NOUS or openrouter to pick between Flash or Pro for specific tasks that require more churning.
AMA.
Flash is a delightful model and the start of intelligence at effectively insignificant cost.
From here on, it's going to become all about harnesses that best situate and organize swarm intelligence at scale.
This reminds me of those pareto-style speedrun record charts when a new glitch is discovered.
[0] https://taylor.town/silver-landmines
When I see dramatic leaps like this, it tells me that the important hacks haven't yet been discovered.
Caching makes a huge difference to cost. On Fireworks AI, for example, if it hits the cache, you pay only 20%. And uncached is just $0.14/M tokens for DSV4-0731! I get entire re-architecture projects (with new tests and documentation) done for mere dollars. DSV4-0731 is a daily driver for me.
But note that you have to use Cline (or other harness) if using vscode. I was shocked at how poor the recent versions of GitHub Copilot are at using the cache (with Fireworks AI, but I believe it's a more generic problem).
I always find it confusing that a meaningful volume of the comments are saying "this reached parity with SOTA models. Best $/task."
And a meaningful chunk of the comments are saying "this piece of garbage isn’t even at the level of gpt-oss 20B".
DeepSeek is my cheap and cheerful Chinese model of choice for API use. Has been for a while, but now it's Flash instead of Pro. Even cheaper, and now better then Pro. I feel like most of the major Chinese models are benchmaxxed, they have weird quirks every time I use them (Qwen 3.8 Max doesn't check its work and leaves stuff broken, doesn't write tests unless prompted, etc., Kimi ends up being quite expensive and rarely better than GPT Sol or Opus 5), while DeepSeek models seem to be generally as good as the benchmarks indicate: Not the best, but stronger across the board than any model within an order of magnitude of its price.
It's always fun when Max reasoning is cheaper than High reasoning.
These are very interesting results, and honestly hard to believe, even as a big 0731 fan.
If I'm reading the chart correctly, a couple observations:
* deepseek-v4-flash-0731 max is better than kimi-k3 max
* glm-5.2 is dumber than a box of rocks (this must be on low reasoning or something, right?)
This is way more extreme than other results I'm seeing, like those from Artificial Analysis.
Perhaps it might be interesting: a latent thinking version is here https://huggingface.co/nmitchko/DeepSeek-V4-Flash-0731-Laten...
Does no thinking emissions for context saving.
I use DeepSeek on a daily basis and I didn't spend $10 in the whole month of July delivering eight fully functional apps.
Btw if you need an app I may deliver it to you in ten minutes for just five cents if I'm in the mood. Just let me know.
wow. i remember when GPT-5.2 (medium) was everyone's favorite.
ARC-AGI II:
- GPT-5.2 (medium) %26.7 ($0.759)
- DSV4-Flash (max) %61.4 ($0.04)
Did anyone else experience a change in verbosity? I've been playing around with an agent that holds your hand in a Jupiter notebook and it felt like it started writing essays versus nice, concise, helpful paragraphs like before. My gut was correct because I checked my Deepinfra usage and it was almost a 2x out-token usage for every in-token.
Not a huge deal since it's still cents per session, but my bigger issue was the weird change in tone. It became a lot more pretentious and over-explanatory.
Heavy prompt reworking helped but maybe that's just the cost of being better at coding and ARC-AGI?
I have been using deepseek v4 pro almost exclusively. I was using Kimi a lot but it just nose dived. The decline started with the release of 2.7 and accelerated with the release of 3.
When I need vision capabilities I use GPT 5.3 codex and if deepseek can’t figure something out after a few goes I switch to GTP 5.5 or 5.6 (I’ve been giving Terra first bite recently and it does pretty well, and have used Sol a couple of times).
Using this regimen means I spend under $100 per month on inference and I work all day everyday with multiple agents running simultaneously all on API token spend not subscriptions.
That page needs a Pareto frontier display. But wow, it absolutely demolishes.
This latest DeepSeek is almost at the "too cheap to meter" level. That's going to be a larger unlock than models like Fable/Mythos that are way too expensive to justify, IMO.
What secret sauce do they have?
Got this up and running at 250 tok/s about 2500 agg tok/s for 16 users and now this morning got https://huggingface.co/webbrain-one/DeepSeek-V4-Flash-0731-V... working - patched it to support multiple images... should post this soon to huggingface - having about gpt-5.6 terra high locally feels pretty good
One of the best things about this version is that it is trained in the codex harness. It feels just as good as OpenAI models in using codex tools, but extremely cheap and with 1M context
I’ve been refreshing hacker news constantly for a week now waiting for v4 pro, after they stated it would follow «soon». I have learnt «soon» is a matter of definition.
it's great but we need a multi-modal model of this quality and price to truly declare victory.
But it makes me quite curious, how a text-only model can do so well on ARC-AGI-2 being a set of visual puzzles? It would have to solve it entirely using text-only spatial reasoning about the grid (or maybe writing code?). I am curious if this is normal or do other models use their vision capabilities to solve the puzzles?
The interesting part for me is whether DeepSeek V4's reasoning gains hold up on long-running tasks rather than short benchmarks.
Price raise incoming
Finally something that is breaking away from the pack. Interesting that max costs less than high. I still think, currently, TPS is more important than near frontier intelligence. Likely for reasons that LeCun outlined, maybe out of a billion prompts you will get value from that intelligence. When we have very fast models abstraction will work as that filter.
This is the best model to come out since the beginning of open weights models for those working with classified data that you can not use hosted services for. I've been using it pretty much day and night since it landed and I'm nothing short of amazed. You'll need some pretty good hardware to run it though.
Looking at the caching price of deepseek compared to its competitors, does it have a secret sauce or is it just subsidizing?