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!
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.
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
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.
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!
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.
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).
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.
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).
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.
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.
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
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.
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.
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
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?
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.
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.
Note they double the price if you use during peak time. However, they define peak time with respect to China, not Europe or the USA...so if you are out of Asia, I guess Australians might be impacted, and its still cheap anyways.
I'm really excited for where the open weight models go from here. I've had fun with just CPU inference on old servers that only have AVX1; here's hoping for commoditized TPU-like hardware!
Perhaps I am doing it wrong, but the deepseek llms fail so hard for any complex task and don't compare to the paid models. It keeps forgetting basic instructions after two responses
They did recently announce they're increasing prices though (got a mail yesterday I think), so not sure this analysis showing it as price outlier will last
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.
I love DeepSeek V4 Flash since the pre-0731, now even more. It is the first model that is truly too cheap to meter.
But I find it having a pretty significant problem with tool calling - no idea why, but tool calling with it is SLOW. As long as the model is reasoning, all good. But give it a bunch of tools and it becomes extremely slow.
One thing that popped into my head is that this shows how committed they are to building something that scales across the world
China has zero energy concerns in terms of energy production - not literally zero, but they’d be able to prioritize other dimensions and not necessarily worry about efficiency
Here they are though releasing models that sip resources
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.
Price is confounded by VC subsidies, economies of scale, and inference optimizations. I think a more interesting chart would be ARC AGI vs forwards pass flops or ARC AGI vs training tokens. Of course we don't have those numbers for the closed source models or even some of the open weight ones.
DeepSeek V4 Flash 0731
(arcprize.org)770 points by tosh 7 August 2026 | 462 comments
Comments
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!
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.)
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.
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.
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
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.
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 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.
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.
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).
https://x.com/vijucat/status/2085415745144672492?s=20
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.
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.
From here on, it's going to become all about harnesses that best situate and organize swarm intelligence at scale.
And a meaningful chunk of the comments are saying "this piece of garbage isn’t even at the level of gpt-oss 20B".
[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.
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.
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?
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.
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.
ARC-AGI II:
- GPT-5.2 (medium) %26.7 ($0.759)
- DSV4-Flash (max) %61.4 ($0.04)
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?
Does no thinking emissions for context saving.
promising!
https://news.ycombinator.com/item?id=49198661
What secret sauce do they have?
But I find it having a pretty significant problem with tool calling - no idea why, but tool calling with it is SLOW. As long as the model is reasoning, all good. But give it a bunch of tools and it becomes extremely slow.
Am I the only one experiencing this?
China has zero energy concerns in terms of energy production - not literally zero, but they’d be able to prioritize other dimensions and not necessarily worry about efficiency
Here they are though releasing models that sip resources
Imagine if they had GPU resources of western labs.