DeepSeek V4 Flash 0731

499 pointsposted 9 hours ago
by tosh
(arcprize.org)

303 Comments

LaurensBER

9 hours ago

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!

abixb

7 hours ago

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.

FernandoTN

6 hours ago

I think you're overlooking the fact that for long-horizon tasks, even small errors compound over time and can lead to catastrophic outcomes.

For simple queries, we have reached the threshold since the beginning of the year, and models are good enough from every provider to make a meaningful difference between one another. (ChatGPT, Claude, Gemini, Grok, MuseSpark, Kimi, DeepSeek, GLM...)

The real unlock will be, and you can already see it with GPT-5.6 and Fable-5, to delegate complex enough tasks that will take more than 24 hours to get done and they will not lose track. I'm not talking about a loop, but the actual intelligence to recover from these compounding errors that accumulate in dumber models.

We're still a long way from the intelligence needed to let one of these agents go ahead and supervise multiple layers of sub-agents underneath to do complex orchestration. The future looks very promising and exciting. Imagine having the possibility of a Frontier model orchestrating as many sub-agents as needed that are running on cheaper models like DeepSeek.

Cookingboy

4 hours ago

That doesn’t make sense. It’s not like SOTA models are error free, yet we still use them.

You use Fable 5 right? If that’s good enough for you now, why wouldn’t a Chinese model that’s as good as Fable 5 but at 10% the cost be good enough in 6 months?

AussieWog93

3 hours ago

I think we put up with Fable's occasional hiccups because there's nothing better at the moment.

I use Claude Code semi-heavily for my small business, and the $100/mo I pay for that is a rounding error compared to the value it provides.

If I can avoid spending an hour or two "massaging" the output from a lower-end model once, or it avoids introducing one load-bearing (sorry, couldn't resist) bug, then that's the entire $100 right there.

Hell, you could argue that the best "coding model" that we have at the moment is the human brain, and people will gladly pay $10,000/mo for one of them.

Arguing over $20 vs $100 for something that actually puts in work just seems insane to me.

palata

2 hours ago

> I think we put up with Fable's occasional hiccups because there's nothing better at the moment.

Which was an argument for using every less powerful model since the moment they got useful, right?

When was that? Opus 4.5 maybe? Let's say Opus 4.5 for the sake of the argument. So back then we were like "DeepSeek is not good enough, I need Opus 4.5". Now DeepSeek is better than Opus 4.5. So if Opus 4.5 was good enough back then, DeepSeek is better than that now.

Sure, it's always nicer to have a slightly better model. But the price difference starts mattering a lot more when all the models are already sufficiently good.

AussieWog93

2 hours ago

To put actual numbers on it, since using AI to start solving all kinds of bottlenecks/inefficiencies in our small business, we've seen monthly net profit go up by around $4,000 USD. These are semi-permanent fixes, and the tech is only partially deployed. I am the only one using it, and I only use it part time.

We've just spun up our first Hermes agent, with direct API access to our main inventory system and that's expected to find another few grand per month in misallocation/inefficiency.

I wouldn't be surprised if we were doing more like $10k/mo higher in 6-9 months' time.

When you're talking about numbers like this, the fact that one AI is $100/mo and another is $10/mo or $40/mo doesn't matter. They could make GLM-5.2, or any other Opus 4.5-class model free and it still wouldn't make sense to deploy in a commercial context.

The other angle I'd approach things from is that Opus 4.5 (and I'd agree with you that that model was the saddle point) was "good enough" for the types of things we were asking it to do back then, but as the models have become more capable the tasks we're asking them to do have also expanded with it.

I know I've personally gone from "hey can fix this race condition with a Redis mutex" 6 months ago to "Independently redesign this full embedded USB stack and QA it end-to-end, working around a specific Kernel bug in macOS Tahoe that requires decompilation to find the source of, while keeping in mind the constraints of our 8-bit AVR chip from 2011" now.

But that said, yes, maybe in 5 years' time we will reach an "intelligence saturation" where the average person won't be able to even conceive of how to use the new SOTA.

ray_kay777

25 minutes ago

I think we're even starting to reach that saturation point now for a lot of people. In my industry (law) plenty of people have tried CoPilot once or twice, or tried ChatGPT a year ago, and as a result have basically dismissed AI as being useless. The setup required to be able to get it to do end to end tasks to your liking is also substantially more work than most people are willing to put in.

RealWed5

27 minutes ago

"оur first Hermes agent, with direct API access to our main inventory system" – let me assure you that absolutely nothing can go wrong here, mate. /s

godwinson__4-8

2 hours ago

The question low cost models will create: Why would you massage output?

Fable 5 is still going to mess things up at any sufficient complexity. The advantage of low cost models with "good enough" intelligence is they can recursively correct. Why? Because it is cheap. Proper requirements and tests and subagents take away increasing amounts of work, at a cost that is not prohibitive.

If you are reviewing code manually you might consider Fable 5 a worse option. As it articulates itself with higher confidence and you already know it is capable, you are may be more likely to miss a mistake. You know to be on guard with a junior engineer. Reviewing a senior who suddenly makes some weird stochastic mistake can be a lot harder. It would be like if the smartest human engineer you knew was capable of some random brainfart in the middle of their massive diff. Imo, much harder to deal with.

Of course, we should keep in mind Fable 5 is only expensive today. It will be cheaper in the future. Autonomous, recursive prompting and improvement is the clear end state. Especially for entities that will always have the budget for that at the SOTA frontier.

Cookingboy

2 hours ago

Except Fable won’t be costing $100 for enterprises that will be considering the Chinese models.

If $100 Claud Max subscription works for you, then great.

But you have to remember your pricing is subsidized by enterprises that pay hundreds of thousands of dollars each month, if not more, to Anthropic.

For those companies, a Chinese model that can cut their AI spend from $1M/month to $200k suddenly seems attractive.

And unfortunately for the American tech industry, the valuation is based off those enterprise deals, not your $100/month Claude Max subscription.

serf

2 hours ago

> not your $100/month..

This is made brutally obvious by anthropics customer support for people with such accounts.

AussieWog93

2 hours ago

Yeah, fair. If we were talking $2,000/mo vs $200 then the maths starts looking very different.

xyzzy123

4 hours ago

There are a lot of tasks that are hard for organisations to run consistently but require some intelligence - monitoring logs and metrics for anomalies and security events, backup audits, audit processes in general, ensuring document quality and consistency, database advice and tuning, customer experience management, process optimisation - that are not "long horizon" in the classical sense of each step depending on the last, but are the result of consistency and attention over a long period of time and a large amount of data.

For this genre of task execution can run with limited horizon and is independent but would be too expensive to do with "us frontier tokens", I think for these, there is value in availability of cheaper tokens.

svachalek

5 hours ago

These are not 24 hours of inference with floating point errors accumulating; largely the system guards against errors compounding. Tool failures, compile failures, test failures, etc, push back against the model taking a wrong turn and force it to correct.

Yes it's much easier to have a smarter model that goes straight to the correct answer first, but it may not be necessary or economical. There's a minimum bar for the model where it understands problems and knows the right step to correct them, and above that newer models give diminishing returns.

mycall

an hour ago

> it's much easier to have a smarter model that goes straight to the correct answer first

That's basically ASI not AGI, if you agree humans are NGI (natural general intelligence) and make mistakes and wrong decisions in solutions all the time. Right steps with some wrong ones is acceptable though for AGI.

copperx

6 hours ago

> these compounding errors that accumulate in dumber models

While SOTAs handle these errors better, they compound in all models and there's a term for that. It starts with cluster and ends with an expletive.

I wish I could, but I don't see the need for human steering going away soon if the task involves anything novel (see Terry Tao's chat).

est

3 hours ago

> even small errors compound over time and can lead to catastrophic outcomes

So, death sentence even to frontier models?

ComplexSystems

6 hours ago

It is true. I don't care about having infinite frontier-level intelligence, and I don't care if Fable can one-shot frobnicate a klaxelzorp with a benchmark performance of 97%. I doubt most people do, in fact. I just want something that meets the baseline level of intelligence needed to be a really, really good pair programming agent. It shouldn't have any silly dealbreaker issues involving laziness or hallucinations, it should be smart enough to bounce ideas off of, and it should automate doing tedious boilerplate. And - most of all - I want to be able to afford using it as much as I want. That's what has happened here.

abixb

6 hours ago

I wonder when we crossed the "99 percentile of intelligence for 99% of the usecases" threshold. At this point, the gains seem to be right at the very edge of bleeding edge for narrow and specialized use cases, and wonder if it'll be a sort of diminishing return from here on.

dukeofdoom

2 hours ago

Probably the best counterexample is the games they are able to design. It's still mostly AI slop, few would want to play.

inciampati

6 hours ago

It is amazing how fast it happened. Right now one of my main projects is fully running on DeepSeek flash. My reason was that I was blocked by both of the main US AI labs from working on it because it involves viruses. DeepSeek flash has been killing it since I switched it on, completing the first phase of the project and setting up an iteration in another application space. It isn't the most brilliant model, but it is reliable and I don't have to manage my weekly token allowance. I just spend freely and end up spending only a few dollars a day. Intelligence is going to become a basic commodity. Only special stuff is going to drive us to use special models. And maybe not even that.

btbuildem

2 hours ago

> 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

In what kind of sad and failed dystopia is this a "saving grace"? For whom?

ekidd

2 hours ago

> If what you're saying is true and accurate, then US-based AI labs are in big trouble.

I've been working with DeepSeek V4 Flash 0731. I'd say that it's maybe not quite as smart as Opus 4.5, but it's willing to think things through carefully and keep going until it gets a good answer. So it's a decent Opus 4.5 replacement. Just let it cook.

It isn't Opus 5 or Fable 5. But it's nearly free on Open Router, and it's self hostable on a Mac Studio with plenty of RAM, or using an RTX Pro 6000 Blackwell or two. Which is chump change for any company that employs programmers.

It would absolutely have been a frontier model last December.

rcpt

3 hours ago

The bet isn't that people will be able to automatically reply on bugs and rack up API charges.

The bet is on using AI to gain competitive advantage. You don't win the stock market or make the deadliest drone by switching to the cheap model

palata

2 hours ago

> You don't win the stock market or make the deadliest drone by switching to the cheap model

Really? How many times a small team has outperformed a much bigger one just because they were "doing it right"?

I have been in software companies where most software produced was bad. Not just the code, the overall design everywhere. So... bad engineers with the most expensive model, or great engineers with cheaper models?

qup

2 hours ago

There are more than two options. What about great engineers with great models?

horsh1

5 hours ago

If programming in the US to become unconditionally 10x more expensive, then the exodus from the US is about to begin.

stockworks

4 hours ago

I couldn't agree more, and think of all the wasted inference across accounts overpaying for their subscriptions.. Need a secondary marketplace for this stuff.

gigatexal

3 hours ago

Yeah this is what I’m curious about. How good are they after the benchmarks. I’ve been told yeah they’re good but they’re just building to show off for benchmarks.

The ByteDance folks are apparently training a mythos level model 10T params apparently. If they do would it still be subsidized at these cheap rates?

anal_reactor

5 hours ago

It's been true for almost every business. "Cheap and good enough" usually trumps "excellent but expensive". Ikea, McDonald's, Ryanair, AliExpress, Aldi - these brands prove that catering to poor people is more profitable than catering to rich people simply because there are so many poor people that their collective spending power outweights the one of rich people.

odo1242

4 hours ago

Well, not universally. It’s a tradeoff. If what you said was universally true Apple wouldn’t exist; Spirit Airlines wouldn’t be bankrupt, etc.

tsunamifury

4 hours ago

Apple sells to half the American population. And by definition many of them are poor.

Spirit was broken by oil prices which everyone pays the same for. (There is no cheaper jet fuel alternative).

Not a good comparison to the point of wrong conclusions.

paxys

8 hours ago

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?

ux266478

7 hours ago

> In the $150/mo range you can get effectively unlimited usage of GPT 5.6 Sol

With 5 active sessions going nonstop? That seems like a pretty important qualifier.

brynnbee

7 hours ago

Others have said similar but I disagree, I'm spending $200/m and I can easily burn through my weekly quota with a few overnight goals using 5.6 medium.

ElijahLynn

5 hours ago

Same here with Claude Opus on 20x Max plan. Easy to burn through with 3-5 parallel sessions, each with their own subs going.

Then, once I go over, API pricing racks up FAST!

brynnbee

an hour ago

I burned through $100 of credits in literally 40 minutes doing the same long running task I always do.

darkwater

7 hours ago

And what do you do with all that?

brynnbee

an hour ago

I've reverse engineered multiple classic games and turned them into popular, browser-based MMO-like experiences.

I'm also creating a free platform that replaces extremely out-of-date software, some of it only available with mutli-million dollar contracts, to help medical physics professionals with cutting-edge radiotherapy devices used to treat cancer.

https://brynnbateman.com/ for a list of projects

pylotlight

an hour ago

Time to make creative software for linux that can replace adobe for Video/photo editing :P

ccakes

an hour ago

An MMO of Full Throttle would be amazing!

brynnbee

an hour ago

It'd be interesting to see what someone could do to turn a 90s era adventure game into an MMO! I never played the game myself but I think basically any game dev project that modernizes stuff is really neat.

baby_souffle

6 hours ago

5.6 medium is pretty good at implementing moderately complicated things as long as you've done a good job specking out the types and the API contracts and acceptance criteria.

I can pretty easily burn through my weekly quota over several agent coding hours with minimal supervision when tasked with some pretty large but well-planned refactors.

rain_iwakura

7 hours ago

not at all true. if you're truly using it across the board for smaller things (translation of pages, filtering of every individual tweet based on its relevance to you etc), the costs ramp up super quickly.

i used for work where i did less and it quickly reaches thousands if you're not careful. i can already see what some will say: skill issue et cetera - whatever.

paxys

7 hours ago

$100-200/mo is the subscription price. You aren’t going to go over. And you can select smaller models as well. Not everything has to be done by the most expensive one.

janalsncm

5 hours ago

You just get throttled, which interrupts your whole workflow.

whimblepop

7 hours ago

I thought it went without saying that GPT 5.6 Sol is the wrong model to use for things like filtering tweets. Apparently not?

nicoburns

6 hours ago

If you factor in cost then it may well be, but it's definitely the case that the high-end models can get you significantly better results than the cheaper models even for tasks that feel like they should be straightforward.

_aavaa_

7 hours ago

I really doubt that the either of the pro plans are subsidized heavily enough to support you swapping v4Flash for Terra, much less Sol.

$5/days is ~330 Mtok/day, that’s a nontrivial amount of work, and none of the gpts are more efficient than deepseek at $/task if deepseek meets your quality bar.

usef-

6 hours ago

They do give a significant amount more than you would expect from the API pricing. The US provider API pricing has heavy margins by all accounts (and most are short enough of GPUs that there's little incentive to drop).

LaurensBER

5 hours ago

5 USD is at the "raw" API price.

OpenCode currently offers 60 USD API credits at 10 USD per month (OpenCode Go) and have even doubled it temporarily as a promotion.

Effectively you can get Deepseek for 1/12th the already ridiculous cheap API price.

_aavaa_

2 hours ago

Those are not at the same price. Opecode’s 60 USD of deepseek usage is charged at much higher rates than what deepseek themselves charge at.

jboss10

3 hours ago

From here, it looks like opencode is hemorrhaging money. I've got a Opencode Zen free account, and I've been using deepseek-v4-flash-free on Pi for a bit, and I haven't hit a limit yet. Sometimes my request fails, but retrys work. I know this is a very cheap model, but it's being given out for free. I assume they might be training on outputs?

jmathai

7 hours ago

The cost per token is super low. If you're used to paying OpenAI or Anthropic API-based fees then the same workload on DeepSeek feels free.

albedoa

6 hours ago

Ignoring the other side of the equation is a pretty wild thing for you to do here:

> 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)

re-thc

7 hours ago

> In the $150/mo range you can get effectively unlimited usage of GPT 5.6 Sol (Pro plan)

Not true. Sol on XHigh or Max runs out even on the $200/mo plan. It's not close to effectively unlimited. Maybe at 2x the current allowance it can.

surgical_fire

7 hours ago

$5 a day is pretty extreme in DeepSeek. You really have to abuse it to get anywhere close to it. Maybe something in around hundreds of millions of tokens per day, considering cache hits and all.

And to be frank, it is not that much weaker for regular software development work. I use Claude at work and I see no difference in capability. I only notice a dramatic difference in how much more expensive it is.

Aeolun

9 hours ago

But DeepSeek now has a warning they’re going to sharply increase their API pricing sometime in the future.

LaurensBER

9 hours ago

Dax (from Opencode) has tweeted that they can replicate or beat the price with rented GPUs. Deepseeks secret sauce is the incredibly cheap caching (magnitude cheaper than other providers).

vLLM has recently released a similar approach. It's not as effective as what DeepSeek does but still an interesting development.

I have no doubt that in due time other providers will match or perhaps even beat the current DeepSeek prices.

minraws

9 hours ago

As someone who recently tried it on some blackwell cards, it's possible to match the prices especially the input can be even cheaper and output can match the costs so you can easily build a net 20-30% margin business even at current GPU prices.

The entire issue is caching, I tried to write some custom to dump to disk kv-caching using some ideas from their papers and my experience with snapshots and vm checkpoint systems, I must say they must have really squeezed that lemon it's hard.

Atleast me with Sol couldn't figure it out over a couple days, a few hours each day, which isn't much but I did feel a bit stuck with existing solutions and felt like I might have to write something from scratch. But if you are willing to put in the effort into the infra I do think it's doable. But it will be really hard to pull it off.

My congrats to anyone who manages to pull it off, they might be able to kill off most AI labs. Assuming they can find the compute, Deepseek really has killed all models for me other than Sol/Fable/Opus/K3 tier stuff.

minraws

6 hours ago

Mild info dump, since this has a few too many upvotes and some folks might be misunderstanding, 20-30% is assuming a typical agentic workload where input tokens dominate by over 20:1 or at least 10:1, if you are output token heavy then this is going to be a different ball game.

And there is no way in hell anyone can afford caching prices same as what DeepSeek is offering, and DeepSeek keeps the cache available for an insane amount of time most providers will flush it in 5-mins like Claude/Anthropic (some offer customizing it but I am not sure of the pricing, it's load based on some like Fireworks, which means assume a couple minutes at most, they say several minutes god knows what that really means).

There is no way to match DeepSeek's current prices, "profitably" if you are renting a GPU and reselling tokens, unless you have some really amazing caching infra or something.

Deepseek's prices are just insanely cheap, I am not saying it's impossible to get there the overall performance suggests it should be feasible, but I will be damned if any provider could match their tps and caching any time soon at those same prices profitably.

I believe even if Deepseek 2-3x their prices across the board even then they would be cheaper for most long running tasks, that's just how good their caching is.

For one I have managed to hit the cache after over 24 hours on their system it's insane, I honestly didn't care because it was so cheap but it truly made me incredibly happy to think about the engineering that must have taken. TTFT is slightly worse, but it's good enough, for those cache prices I can take a few seconds worth of hit on TTFT.

janalsncm

5 hours ago

From what I understand about deepseek’s pricing, they are only charging what they need to break even.

Karrot_Kream

4 hours ago

Thanks this is a comment with a great amount of useful detail.

twotwotwo

8 hours ago

One read is 1) they're getting a lot of traffic for Flash, 2) they've said they're updating Pro soon and expect that to lead to a traffic spike for Pro, but 3) that would leave them overloaded, so 4) they're going to raise prices to avoid it.

It's interesting that most open models adding 1M context did it in a way that reduces KV cache size (though DeepSeek was the most aggressive, using compressed attention on all layers), but only a couple providers turned it into a discount on cache reads.

onlyrealcuzzo

9 hours ago

> Deepseeks secret sauce is the incredibly cheap caching (magnitude cheaper than other providers).

Can anyone working at one of the main US labs (Google, OpenAI, Anthropic) comment on WTF they haven't even tried MLA - despite the obvious massive advantages?

I know enough to know they aren't completely incompetent. So there must be a quite good reason.

But it remains a mystery to me.

DeepSeek's MLA is like almost 2 years old at this time. They've got thousands of people working on this stuff. They clearly have the ability to at least try it...

aabdi

8 hours ago

They already are?

There’s a measurable performance tradeoff versus gqa so there’s reluctance.

For the most part though the new deepseek v4 tech is hca and mhc and people are still catching on like with moe and rl. Wait for 6 12 months, minimum time for next pre train.

ronsor

8 hours ago

Are they not?

The big US labs are opaque and don't publish much of any technical details anymore. We don't know what they are or aren't doing, honestly.

NorwegianDude

9 hours ago

Eh, what are you guys even talking about? Deepseek is not cheapest provider as is, and it's MIT. So deepseek making it more expensive to use is just nonsense, they can only change their own pricing. It's the beauty of MIT license and open weights. If anything, these models are some of the safest in the world to use if you worry about a rug pull.

LaurensBER

9 hours ago

There's more to inference than just the input/output token cost. Caching has a massive impact.

Deepseek charges $0.0028 per cache read on Openrouter. The next cheapest is $0.018.

That's a massive difference and quickly adds up on coding sessions (which often hit 95%+ cached tokens).

akman

9 hours ago

90%+ cache hit rate is common, and so you'll see on places like openrouter that Deepseek cache cost is indeed a magnitude cheaper than the rest.

greenavocado

8 hours ago

My usage thus far from api.deepseek.com

  - input_cache_hit_tokens: 1,265,646,976 x 0.0000000028 = $3.5438115328
  - input_cache_miss_tokens: 18,208,088 x 0.00000014 = $2.54913232
  - output_tokens: 9,615,178 x 0.00000028 = $2.69224984
  - request_count: 10,837 (no price)
Total cost: $8.7851936928 (approximately $8.79)

Cache:

  - Hit: 1,265,646,976
  - Miss: 18,208,088
  - Total input tokens: 1,283,855,064
Hit rate: 98.582% (1,265,646,976 / 1,283,855,064)

natrys

5 hours ago

So your cost would be $40.68 with another provider that has one less zero in the cache hit price.

hagen8

8 hours ago

Cached input tokens are what drives most costs.

re-thc

7 hours ago

> they can replicate or beat the price with rented GPUs

They "can" is the caveat here. Rented GPUs are going up in pricing. I recently got an email that DigitalOcean pricing of GPUs were going up.

So

1. They have to get a hold of them (availability is bad)

2. They have to maintain the pricing

_aavaa_

8 hours ago

I'll believe it when I see it. Their prices are still much higher than deepseek, especially the caching.

retinaros

9 hours ago

any link to this caching tech?

LaurensBER

9 hours ago

[Feat][Core] Add disk offloading support to SimpleCPUOffloadConnector — #49644 https://github.com/vllm-project/vllm/pull/49644

This adds disk as a tier in the HBM → CPU → Disk KV cache hierarchy.

There's also a cluster of related KV-offload FS PRs: #49225 (read/write batching, still open) and #49152 (batch store/load in C, merged Jul 28).

It's hard to say if these are similar to the approach DeepSeek takes but they definitely seem very interesting.

fastball

5 hours ago

But the model is open weight?

eli

9 hours ago

I assume/hope this is about prices going up for the next release of Pro

HSO

9 hours ago

even if they double it it`s from such a low base it is still supercheap

metadat

9 hours ago

Source?

dolebirchwood

9 hours ago

If you're on the DeepSeek Platform, you'd see this:

"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."

amelius

8 hours ago

> it's good enough to use it for (almost) everything

which in your case is?

rpdillon

8 hours ago

I've posted a few times about my project that's a collection of 30k-250k webapps that are served from a WebDAV server. The apps know how to write updated copies of themselves back to the server.

My family uses it. I have gallery apps (yearbooks for each year are a lot of fun!) of us on trips and just living, an outlining app that's a mesh of Workflowy and Org Mode (it's called Fluxtral), a markdown-backed app (it uses marked.min.js, and is called Dextral) that offers documents, logs, calendars, and kanban boards, all parsed from markdown. I have a List app for gear, trips, shopping, etc. that we all can contribute to. There are utilities (world clock, calendar) and games (an oracle for RPGs, a KenKen implementation), and apps (a diagram editor that exports to SVG, a web-launcher that uses pneumonics, a Scheme-based hacking environment, and a spreadsheet that does most of what you'd expect aside from Solver and Pivot tables).

I started these projects before AI, and made slow progress over the years, but the modern versions of all this stuff have been built with Deepseek V4 Flash. I've also used Gemini in the very early days, and Kimi K2.6 later on, but these days, since I can now host Deepseek v4 Flash 0731 in a 2-bit quant on my Strix Halo box (128GB, but only about 250GB/s of memory bandwidth, so 15t/s), I used Deepseek with omp for almost everything. It's a very capable model, and I'm amazed I can run it locally and get good results. It's really revolutionary for my (small) use cases.

indigodaddy

5 hours ago

Sounds fascinating! A blog write-up about your platform would be a fun read, if you're up to it

rpdillon

2 hours ago

For sure! I'll be doing a Show HN at some point, just want to feel a bit more confident about certain aspects first.

indigodaddy

2 hours ago

Makes sense, I'll look out for it! although of course most Show HNs these days get lost in a deluge of submissions.. you might actually be better off omitting the Show HN when you submit it..

podnami

6 hours ago

A collection of 30k-250k apps? Like individual unique apps?

rpdillon

4 hours ago

Sorry, a collection of apps whose size is between 30kb and 250kb.

indigodaddy

4 hours ago

Haha, I misinterpreted as well. That would be a lot of apps!

dan_q

8 hours ago

> which in your case is?

oh, they're mad.

afro88

3 hours ago

> 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 don't think this is the win you think it is. It's amazing that this is possible, but it introduces so much human overhead that you can drown in reviews and it can effectively slow you down more than a quick check and fix yourself.

The models need to get a lot more consistent in what they can and can't do before you can automate this stuff and only check the things you know the model isn't good at

anramon

9 hours ago

>even if it's not SOTA

And, probably 99.99% of people using LLM probably don't even need SOTA anyway.

swiftcoder

9 hours ago

At least on these benchmarks, it seems to be pretty handily scoring up with the SOTA from 6 months ago?

ljosifov

7 hours ago

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 :-)

klardotsh

32 minutes ago

I would not recommend DSV4F (even 0731 edition) without an advisor. On its own it’s an absolute drunk intern in my experience, but with an advisor model watching like a hawk when it gets stuck in loops or goes down boneheaded rabbit holes, it’s fine (and very cheap). I’ve been using GLM-5.2 as my /advisor but might try just a second DSV4F instance.

h14h

4 hours ago

Are you using it via the official DeepSeek API, or via a different model provider? If the former, it's worth noting that their cache read prices are one tenth that of every other provider ($0.0028/M vs $0.028/M), so folks who want to use a sovereign inference provider with a zero data retention policy likely won't see anywhere close to the same value.

mh-

2 hours ago

Worth mentioning also that DeepSeek is the only provider in OpenRouter that was disabled-by-default until I enabled a setting: Allow paid endpoints that train on request data.

swingboy

5 hours ago

Agreed. With less than $10 on the DeepSeek API used, I’m somewhere near half a billion tokens over the past week or however long it’s been since it came out.

I’ve found it to be very capable. I’m using it with pi as well and some custom extensions I’ve put together over the past few months and it’s pretty crazy having it do what I need it to a vast majority of the time, do it fast, and see that it’s used like $0.12.

m101

5 hours ago

Could you go into how you run two instances that speak to each other in an implementer / advisor role in parallel? I’ve been looking for this sort of orchestrator / worker solution where there’s constant feedback and nudging between the two.

LaurensBER

4 hours ago

Sure! It's https://omp.sh/ with the advisor mode enabled.

You can probably implement something similar as a plugin for your preferred harness. From a technical perspective I think it just sends the output w/h the thinking and tool trace to another model and asks it to double check everything (exact prompt must be somewhere in the OMP repo).

m101

4 hours ago

Thanks I’ll give that a go.

Would you run a less costly model as the supervisor given it’s consuming a lot of text and may have a simpler task to do like “make sure the implementing model doesn’t start over-engineering things”?

frogperson

3 hours ago

The pricing was awesome, but deepseek just sent out emails warning of a large price increase.

LPisGood

6 hours ago

Does auto generating tests even do anything helpful? Don’t they just sort of tautologically say the code does what it does at best or do something completely ridiculous like test and implementation that only exists in the test file at worst?

LaurensBER

6 hours ago

We have an extensive description of _how_ tests should be written and they're reviewed by a human. All the AI does is fill in the boring middle part.

sfifs

6 hours ago

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

jojohack

7 hours ago

Running DeepSeek with Pi as well, any plugins you recommend running it with ( e.g. native browser for snapshots, etc. )

ycui7

2 hours ago

it is funny when people say i am struggling to spend money.

electroglyph

5 hours ago

my experience is the same, but deepseek is planning on increasing prices soon, which will make it a lot less attractive

meetingthrower

8 hours ago

What's the best harness to use with it?

LaurensBER

8 hours ago

I've enjoyed using https://omp.sh/

rpdillon

7 hours ago

Seconded. I love OpenCode and Pi, but omp is my daily driver.

stavros

4 hours ago

What's good about it? I use OpenCode and it does what I need, basically.

rpdillon

2 hours ago

Spawning subagents via orchestrate and using /advisor are both super valuable. Haven't really unlocked the full capability of omp yet, but hoping to over time.

ljosifov

7 hours ago

omp - current top, after using codex claude opencode pi that I still use too

dominotw

8 hours ago

> Auto generate tests on CI for every pull-requests!

this seems like such a bad idea

EchoVoicy

8 hours ago

Depends on the prompt I think. If it's just "Generate tests plz" then I agree, but if its

"If this PR adds any new endpoints, ensure that there are functional and integration tests. If there are not, please investigate the feasibility and appropriateness, and create functional tests using the guide found on our wiki for guidance https://www.ourdevwiki.site/how-to-make-functional-tests" then maybe it could add some value.

But that very much depends on the specific system. Some tests are obvious, some not so much.

tcp_handshaker

7 hours ago

And software keeps getting worst.

The analogy I like is that building software is running a Michelin restaurant. The moment you scale, the chef is just writing cooking books and is absent, and you move into franchising, you will be amazed at the bottom line revenue scaling, while customers will be progressively appalled with the food...

palata

2 hours ago

Not that I disagree, but the average software before AI was more like a McDonald's. I have genuinely seen companies who were writing code a lot worse than what AIs produce nowadays. Doesn't necessarily mean that their software is better now, but my point is that before AI, I don't think that software compared to Michelin chefs.

jmyeet

8 hours ago

> 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.

I hadn't really thought about this but AI may well be the technology that disrupts and ultimately destroys social media.

The value proposition of something like FB or IG is, as we know, the network effect. The platform gets to extract value from user generated content. I believe that users should own the platform, a bit like the Wikimedia Foundation, because they're the ones that create value. Federation is a popular belief on HN and I've come to believe that's simply the wrong solution to the right problem.

Anyway, how these social media companies make money is by optimizing the feed for engagement. People know it too so you see people trying to build an audience by rage baiting. And then more time spent equals more advertising revenue.

But what happens when the AI can simply slurp all the posts and then filter and rank them? It destroys the engagement and advertising model. And I'm not opposed to that, honestly. It may be on eof the few good thing sto come out of AI.

catigula

8 hours ago

>Test coverage too low? Auto generate tests on CI for every pull-requests!

Terrible use-case.

_s_a_m_

8 hours ago

These posts have to be Chinese bots, these models are all trash. Used it via OpenCode for an hour, cost me one hour of my life. It is for anything complete trash.

r14c

7 hours ago

I've gotten a lot of good work done with deepseek models. Like with any generic harness there's some tuning that has to happen. I used open code for a while, but I've landed on pi.dev as my go to since its easier to tune and has better deepseek integration. iirc open code is quite bad at utilizing cache and doesn't have a lot of ways to specifically tune the harness for a particular model.

apitman

7 hours ago

I've found it to be pretty good so far.

greenavocado

8 hours ago

(1) you used opencode (2) what provider did you use. openrouter is trash because they shit up the model serving. no max effort and horrific cache utilization, on the order of 50-75%, absolutely garbage. beware

alex0015

8 hours ago

What should we be running deepseek on besides opencode? I chose it because I heard good things. Also provider is directly through deepseek credits.

stavros

4 hours ago

I use the Deepseek API and pay peanuts. Very satisfied.

greenavocado

8 hours ago

oh you used opencode go?

harness: omp.sh

dan_q

8 hours ago

You're mad.

EchoVoicy

7 hours ago

Point 1 finger out, and you point 4 back.

ak_t

7 hours ago

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.

ComputerGuru

6 hours ago

What quantization level is that? Because official endpoints are slow.

ak_t

6 hours ago

It doesn't need extra quantization. The official weights are natively mixed precision FP4/FP8, so it fits in ~160GB. The API slowness is probably from being batched with other concurrent user requests. The provider's aggregate throughput gets higher but per-stream speed slows down.

zargon

5 hours ago

V4 Flash fits entirely in two RTX Pro 6000s without any quantization at all.

bel8

6 hours ago

From opencode go $10/mo plan I get between 60 t/s and 100 token/s even with large contexts of 150k+ tokens.

I wouldn't call 80 t/s slow.

ponyous

6 hours ago

You are right, relatively to other llm providers this is not slow. But if you think what is possible when you have 1000t/s a sec you might find it slow.

apitman

5 hours ago

I'm getting like 25 tok/s on 2x RTX Pro 6000. This is with llama.cpp, but I had GPT tune it for me. I was under the impression vLLM was at most ~2x faster, and usually for highly parallel loads. Any tips on where I should look first for an obvious blunder?

I'm guessing tensor parallelism or similar?

lcampbell

4 hours ago

You're gonna want a custom vLLM build.

Here's a runbook: https://github.com/local-inference-lab/rtx6kpro/blob/master/...

If the newer builds aren't working, you might try running the old v6 build (based on the eldritch-enlightenment image). gilded-gnosis gave me some problems that I haven't bothered to track down, the old builds are still gonna blow away llama-server performance. And that's before you get hooked on vLLM's PagedAttention and can run multiple sequences without a ton of extra overhead.

NoboruWataya

7 hours ago

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.)

nodja

7 hours ago

I'm the same way, I have a very low/sporadic usage of any subscription I've tried. I now just use openrouter with DS4 pro/flash. It also gets rid of usage anxiety where I would try to justify the $20/month by forcing myself to use the tokens for projects as the weekly limit deadline neared.

andai

7 hours ago

I've been asking about psyops and bioweapons and I'm still going strong. I did get a Sonnet session shut down the other day though which feels like some kind of achievement.

ignoramous

7 hours ago

> 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

If you prefer subscriptions, OpenCode Go ($10/mo), Cline Pass ($10/mo), Atlas Code ($20/mo), and CommandCode ($1/mo) serve some of the best open weights with generous limits. OpenCode Go currently offers $120 for $10 on DeepSeek Flash v4 (if you're okay with data retention).

apitman

an hour ago

> OpenCode Go currently offers $120 for $10 on DeepSeek Flash v4

At DeepSeek's absurdly low rates or market rates?

kzrdude

5 hours ago

Just trying to understand, https://opencode.ai/docs/go/#privacy currently says DeepSeek V4 Flash has 0 days data retention.

> DeepSeek V4 Flash: ZDR agreement is renewed monthly. The current agreement is valid through August 31, 2026.

Is there other info I should be aware of w.r.t data retention with opencode go? It's hosted in China, so other middlemen may be active (I doubt it, but possible)?

WarmWash

5 hours ago

If it's hosted in China, they can tell you whatever you want to hear and do whatever they want to do.

What are you going to do? Take a CCP company in front of a CCP judge?

jacquesm

22 minutes ago

They can do the same thing in the US. What are you going to do, sue OpenAI or Anthropic?

anon373839

5 hours ago

Yes. There is a big question mark about whether Anomaly itself (as the OpenCode Go middleman) retains data. The docs were completely silent on this the last time I checked.

0xbadcafebee

5 hours ago

According to this code (https://github.com/anomalyco/opencode/blob/dev/packages/cons...) only Grok and Luna have 30-day retention, the rest have 0-day retention. Of their legal documents, none mention data retention for anything other than personal data. The exception is if you '/share' which explicitly gives them your session to share over the web with others.

0xbadcafebee

5 hours ago

OpenCode Go does not send any inference to China unless you go into settings and manually select 'Enable models hosted in China'.

ignoramous

5 hours ago

That's a recent addition I wasn't aware of. Thanks. Previously, OpenCode docs claimed they had no agreement with DeepSeek on data retention for training. Even more previously, OpenCode said they used providers based in US/EU/Singapore (which they no longer do so).

nylonstrung

6 hours ago

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

johnnyApplePRNG

an hour ago

What quantization are you using? Which infra provider?

Baseten.co's version got into a loop rather rapidly... I've since added loop detection and adjusted some other settings on the pi coding agent and have yet to notice it again. I also switched to DeepInfra ... who serves an fp4 version admittedly, but I've had no issues with it as of yet and it's the top provider on openrouter.ai volume wise.

natrys

5 hours ago

I think this one requires a bit of strong prompting. I am also normally a Pi user, but my experience in OpenCode with this model has been drastically better than in Pi, where it overthinks a lot and gets distracted by random things.

It might be even better in Codex or Oh My Pi according to this bench I saw earlier: https://nitter.net/composio/status/2085330847951970801

bel8

6 hours ago

I've been using for work, from opencode $10/mo subscription plan, on high effort (which is better than max imo), and haven't had any issue.

When it was first available in opencode, it was kinda slow for me, I guess because everyone wanted to try the new shiny. But now it's back to being screamingly fast and Opus 4.8 level of smart, for penies.

the_duke

6 hours ago

Yeah, I saw the same thing - quite annoying. It can be mitigated through the prompt.

modeless

8 hours ago

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/

vb-8448

8 hours ago

Why? It's open weight, there are plenty providers on open router that are serving the latest v4 flash at 0.14/0.28 $.

LorenDB

8 hours ago

Yes, but even the cheapest providers on OpenRouter are charging at least 10x what DeepSeek does for cached input tokens, which is where DeepSeek gets most of the cheapness.

petesergeant

7 hours ago

(nevermind, I was reading DeepInfra as Deepseek. My bad)

VulgarExigency

7 hours ago

You are missing a 0 to the left of the 2 on Deepseek's number

modeless

8 hours ago

This would be more convincing if those providers had converged on a number that was not the exact pricing of DeepSeek themselves. Clearly DeepSeek is setting the price here and without them holding it down I expect increases.

vb-8448

6 hours ago

Non necessarily, they can easily increase market share by staying where they are.

nicce

8 hours ago

The most expensive defines the price. Others need to be just slightly cheaper.

542458

9 hours ago

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.

MarkLowenstein

8 hours ago

Real question: is there anybody that is both maintaining alpha-dev capability by keeping abreast of all these daily changes, while also reserving enough time to actually work?

Seems like we've reached the event horizon of whether AI advances are worth paying attention to.

naught0

6 hours ago

I enjoy using opencode go to play around with a lot of different models. I wind up using deepseek v4 flash for most everything, stepping up to minimax m3 if that doesn't cut it, finally preferring GLM for complex tasks or important planning I want to go right the first time

I recommend opencode or something akin to it to play with models. Any big model updates or hot new ones will naturally run across your desk that way

becquerel

7 hours ago

I think the play now is to just try out whatever the best new model is every time you see a headline that fundamentally reorganizes your conception of what's possible.

cyanydeez

8 hours ago

Are you saying we've reached peak Bike shedding?

bee_rider

7 hours ago

How about: The yaks have started shaving themselves, who can keep track of how good a job they are doing?

petesergeant

7 hours ago

I don't think you need to be keeping abreast of them really, you just need to be using the best model you can get enough tokens from, which for many people is Fable 5 @ $200ish, ideally fanning out implementation to cheaper models

whinvik

9 hours ago

Yeah either the benchmark isn't very useful anymore or V4 Flash is a really, really good model.

fallingbananna

8 hours ago

GPT 5.6 Luna is an extremely cheap and still very capable model.

A chinese model being in the same ballpark of capability at half the price sounds believable to me.

ignoramous

9 hours ago

In my use, DeepSeek v4 Flash (which replaced the quite excellent MiniMax M3) lags behind GLM 5.2 & Muse Spark 1.2 (let alone Kimi K3). Also, K3 is a much bigger multi-modal model, while Flash is text-only and likely optimised for coding tasks.

nwienert

8 hours ago

Yep, and the v4 flash final is about 2.5x slower than preview making it no longer a fast model, in fact slower than Luna and bigger models in many cases.

Spark is actually the interesting one imo. It's significantly better, also significantly faster. If you are ok with letting Meta soak up your data (which DS does too) it's also the same price.

thehamkercat

9 hours ago

And now nobody seems interested in it because the price hasn't gone down

it's still $3/$15 for all providers on openrouter

because of some Kimi license

https://openrouter.ai/moonshotai/kimi-k3#providers

johnnyApplePRNG

8 hours ago

Morph has it for a slight discount, apparently.

Uptime looks crap, though.

thehamkercat

6 hours ago

I believe it's because they are below $20 Million revenue limit (which Kimi K3's license has)

So we won't see any price decrease unless Kimi changes the license of K3

dyauspitr

8 hours ago

Not for long, Deepseek is saying they will have a significant price jump soon. They really shouldn’t do it because they are on the cusp of capturing the scalable API market.

telotortium

7 hours ago

They need to be able to serve their market. The price increase is partly load shedding. If they improve their ability to serve their load, they can always drop it again, as OpenAI did with Luna recently.

ignoramous

7 hours ago

> as OpenAI did with Luna recently

My read is, OpenAI is neither able to claw b2b money (away from Ant) nor are they able to stave off open weights on the other. In short, they're struggling to hold onto their distant #2 position in the coding market, and these pricing changes reflect a (desperate) change in strategy.

lukewarm707

6 hours ago

and i still won't use it, because they log and spy on your prompts XD.

the private endpoint costs 10x (azure).

private endpoints for deepseek (lots of providers) also cost about 10x more.

but 10x more for deepseek is $0.028 cached input, and 10x more for luna is $0.10.

mosura

9 hours ago

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.

chorizo

8 hours ago

This also reflects my experience and should put to bed the distillation rumours. This model feels nothing like the Claude models, including tone and blindspots.

andai

7 hours ago

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.

guilamu

7 hours ago

Where does this "10x" comes from?

surgical_fire

6 hours ago

> The recently announced they're raising their prices 10x right?

No.

They sent an email to customers saying that they will raise prices "significantly".

How much that will be is speculation.

My guess is that they will just remove the 75% discount they gave when they released V4 preview. It will still be relatively cheap even at 4x the current price.

gizmodo59

4 hours ago

It will be comparable to Luna then.

stuaxo

2 hours ago

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.

Petersipoi

2 hours ago

I max out my $200/month Claude plan. You are obviously just not taking advantage of it to the same level as others. Which is fine. Don't pay for something you don't need. But I would definitely take a massive productivity hit if I had 1/20th the usage.

therealdrag0

2 hours ago

Just depends on commitment, time, and scope. Personally I also use the 20$ plans. But people spend 200$ for a gym membership. So shrug.

system2

an hour ago

People use Claude Code for professional reasons. Waiting is not a luxury when there is a deadline.

zacksiri

24 minutes ago

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.

theogravity

10 minutes ago

You're using it on low, that's why. There's a huge difference in performance from low to max effort.

jacquesm

an hour ago

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.

LeBit

2 hours ago

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".

ddxv

4 minutes ago

I guess both are true and for everyone at some point. All models, even SOTA, fail. When they fail, it is quite frustrating. Additionally, some models are very cheap to run and use. When Deepseek fails the cost was minutes and pennies.

arjie

8 hours ago

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.

walrus01

8 hours ago

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.

kromem

6 hours ago

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.

SwellJoe

8 hours ago

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.

eli

8 hours ago

Qwen 3.8 Max is very strong at troubleshooting and code review.

SwellJoe

8 hours ago

I'll grant it's very thorough when assigned a troubleshooting task. I'm not as confident of its code review though it is very good at security vulnerability auditing, and isn't hobbled for that work like Fable, and even Opus refuses some work in that area now.

apitman

5 hours ago

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.

dools

5 hours ago

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.

g023

an hour ago

It says 'yes' where the others say 'no'. Good enough for me.

zmmmmm

5 hours ago

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?

ghosty141

5 hours ago

xiaomi mimo is very cheap and not bad.

surprisetalk

9 hours ago

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.

Almondsetat

4 hours ago

It's crazy to think V4 Pro still hasn't finished the post processing.

xyzsparetimexyz

8 hours ago

That page needs a Pareto frontier display. But wow, it absolutely demolishes.

tarruda

4 hours ago

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

djc404

4 hours ago

Do you have any sense how using it with codex compares to OpenCode?

It’s always a bit tricky picking the right harness (when you have options). Sometimes the differences are subtle but meaningful. But who has the time to run everything twice and compare all the time!

tarruda

3 hours ago

I don't have experience with opencode, so I couldn't tell you.

Codex is really good in my experience, especially due to its native sandboxing. Deepseek seems really well versed in its tools, including update_plan and knowing when to request sandbox escalation.

seanmcdirmid

5 hours ago

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.

minimaxir

9 hours ago

It's always fun when Max reasoning is cheaper than High reasoning.

Terretta

8 hours ago

Rework is expensive.

Tell your PjM who should tell your PgM who should tell your PdM, all the PMs...

Maybe if "the business" sees it is true of LLMs, they might believe it's true of giving better context to engineers up front then giving them time to think and prototype (thinking tokens are an answer prototype).

sourcecodeplz

8 hours ago

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)

KolmogorovComp

7 hours ago

Looking at the caching price of deepseek compared to its competitors, does it have a secret sauce or is it just subsidizing?

ignoramous

6 hours ago

That's DeepSeek's way of selling "token plans", yes. But without the downsides like daily or weekly limits and guaranteed upfront/fixed spend.

gentlewater

9 hours ago

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.

w2seraph

5 hours ago

There's no reason that LLMs should cost beyond grave digging sums when this one topples the charts it'll be over.

harisamin

8 hours ago

I'm curious... is anyone using DeepSeek V4 Flash from HugginFace? Is the cost around the same as directly form DeepSeek or from Openrouter?

tosh

9 hours ago

results comparable to gpt 5.6 luna but cheaper

promising!

minimaxir

9 hours ago

Since the x-axis is log-scaled, DeepSeek is much cheaper than visually implied (mousing over the raw values, it's 1/4th the cost of Luna).

LUmBULtERA

9 hours ago

Is it still cheaper than Luna if using an OpenAI subscription? My gut is no, but I have not done the math.

swiftcoder

9 hours ago

You'd have to compare against something like the OpenCode Go subscription, and I'm fairly sure deepseek napkins out cheaper in that scenario

LUmBULtERA

8 hours ago

I'm still not sure, there's a promo going on now, but generally Go gives $60 of API credit and right now it might be $120 with deepseek. But $20/month OpenAI subscription I believe gives you many hundreds of API-equivalent usage? I've heard $100/month giving many thousands API-equivalent per month.

minimaxir

9 hours ago

Everything is cheaper if using a subscription, but some applications require API usage.

jrflo

8 hours ago

Might not actually be that much cheaper, we don't know what margin OpenAI is charging on Luna API. Open models likely have much less margin.

evanjrowley

7 hours ago

The benchmark performance tells me DeepSeek v4 Flash could be very cost-effective at playing SNES/Gameboy games.

andai

7 hours ago

It won't be long before I can just stay home, and have my robot ride my bike for me.

qup

20 minutes ago

I'm going to send mine to visit my mom. It's so hot in August.

luyu_wu

9 hours ago

It is wild that this a log scale of cost to me!

simonw

8 hours ago

That's a pretty great score for a model you can run on as (expensive) laptop.

CrosswordPuzzle

8 hours ago

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!

nikp123

8 hours ago

I just used it for some Kubernetes + FluxCD tasks and oh my is it good.

mycall

7 hours ago

I'm curious how much worse the 0731 quantizations do.

johnmlussier

7 hours ago

Been running it using Prime Agent and absolutely love it.

clayhacks

9 hours ago

Why wasn’t this run against ARC-AGI-3? Or did it fail to solve anything?

chorizo

8 hours ago

They tweeted that ARC-AGI-3 results take longer to run, so we’ll need wait a bit longer.

m3kw9

3 hours ago

The token price seem to be jigged, how do you know if it's subsidized or temporary. Anyone can just lower the token price to get to the left.

jacquesm

19 minutes ago

Almost every service provider in the AI field is subsidizing their token cost to some degree, they're all shooting for marketshare and lock-in (and they're not really achieving the latter).

steadyw0

6 hours ago

Should try it sometime

Havoc

9 hours ago

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

minimaxir

9 hours ago

That is only when using the DeepSeek API directly. OpenRouter has 24 different providers serving it at existing prices.

system2

4 hours ago

How can OpenAI or Anthropic fight against these prices?! $0.14 input, $0.28 output. For 1M tokens...

gxs

4 hours ago

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

dcchambers

9 hours ago

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?

est

2 hours ago

> What secret sauce do they have?

Quant company usually squeezing every penny.

throwaway_95283

9 hours ago

limited resources, no modern GPUs, no $10 billion dev budgets.

pair it with codewhale, 50 agents, 200 MB of ram.

pama

7 hours ago

No secrets—all published. Very efficient attention. Excellent kernels. Great caching subsystem. Small and well trained model.

iagooar

8 hours ago

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.

Am I the only one experiencing this?

casey2

6 hours ago

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.

esafak

9 hours ago

It's serviceable but, like many Chinese models, it uses a lot of tokens to get work done.

gruez

8 hours ago

>it uses a lot of tokens to get work done.

That's irrelevant when you use $/task as the metric, which the OP does use.

esafak

8 hours ago

It also affects the time.

cyanydeez

7 hours ago

If I had the GPU size, hook it up to llama.cpp and setup the --reasoning-budget and reasoning-message; Most of that additional reasoning is a lot of garbage and you can redirect it to useful output.

That's how I handle the Qwen27B and 35B

nomel

6 hours ago

> Most of that additional reasoning is a lot of garbage and you can redirect it to useful output.

What do you mean by "redirect it to useful output"? Could you give an example? This sounds interesting.

cyanydeez

4 hours ago

It's specific to the harness. Using dynamic context pruning, the budget cuts it off after a select amount of tokens and the budget message tells the model to use subgents to finish whatever it's thinking about

nomel

3 hours ago

Nice. Does it use a summarization, or a hard cutoff?

ttkciar

2 hours ago

llama.cpp uses a hard cutoff. The agent then does "something" that is specific to the agent's implementation and configuration. It might summarize and then "finish the thought" with a different model, and then resubmit the prompt to the llama.cpp API endpoint with <think>..</think> prefilled. The primary model then infers the remainder of the reply.

leizhou

8 hours ago

so cool. does it mean it can understand the verificated code

WhitneyLand

8 hours ago

The DeepSeek team is so strong, very impressive.

Imagine if they had GPU resources of western labs.

mosura

8 hours ago

Necessity is the mother of invention.

SV companies get way too comfortable when they have enough in the bank to stay running more than three months.

artursapek

2 hours ago

These prices are not real. They already said so.

addozhang

an hour ago

what's the real you mean?

antirez

9 hours ago

Price is not a good meter. Active parameters per token are. Joule would be even better.

orbital-decay

9 hours ago

It's an excellent metric, the amount of applications not viable now due to cost/latency/throughput is vastly bigger than the amount of current use cases. Even current ones do benefit, e.g. it's a great executor subagent.

Energy and intelligence are good too, sure.

fallingbananna

7 hours ago

What if we used 100% of the brain all the time?

As an end consumer, I don't care about the number of active parameters. I really do care only about the tracked metric (how well does it do the job, and how much does it cost... ideally also with time included, but that wouldn't fit on a 2D chart)

minimaxir

9 hours ago

Price accounts for computational/architectural efficiency improvements whereas active parameters does not.

polytely

9 hours ago

for someone with a limited budget it is actually very important because it makes me less scared to experiment.

muricula

9 hours ago

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.

minimaxir

9 hours ago

DeepSeek V4 Flash 0731 is an open-weights model which means price is determined by competition/invisible hand of the marketplace: https://openrouter.ai/deepseek/deepseek-v4-flash-0731

With the exception of cache costs, all providers have similar input/output costs.

kennywinker

9 hours ago

Not counting the cost of making the model, which is subsidized by… someone? The chinese gov i think?

segfault99

4 hours ago

DS comes out (one of, or) the most successful quant fund in China.

They don't strictly need any kind of subsidies.

FWIW they have a funding round planned (kerfuffle about leaks from CEO presentation few weeks back) -- presumably because infrastructure needs have ballooned.

Naturally there will be some PRC government interest in one of their flagship AI companies. From what is visible seems to be more along the lines of ensuring that DS gets its fair share of resources -- e.g. Xi Jinping meeting founder and positive comments about success of DS means that (hypothetically) Alibaba can't screw DS too much on infra charges to kill off a 'competitor'. Also would imagine that DS's top guys have been clearly identified and will have been 'discouraged' from going to work for one of the SV polycules. But even here as much carrot as stick -- none of the DS top guys will ever need to work again except for love of the job.

_aavaa_

8 hours ago

Subsidized by inference profits and volume.

npn

9 hours ago

weak argument. deepseek v4 flash is open weight, you can easily find other providers with competitive price with Deepseek (except for input caching), some even half as cheap.