Hacker News Reader

Comment by proxysna | Hacker News Reader

I am yet to spend $200 on deepseek this year. Not sure what kind of usage can justify $200/month of either openai or anthropic, i'm not even talking about $500. Deepseek is faster, IMO intelligence difference is negligible and it so much cheaper that i no longer care about how much i use it. I never hit any daily/weekly quota or anything like that while working or tinkering. At this point i am OK with being 6 months behind the "frontier", purely on bang-for-buck basis and who cares which shadowy government gets my data.

View on HN

So I took Deepseek V4.1 Flash for a spin maybe 2 weeks ago now (before Luna 6 and Sol 6 were announced), and I racked up $100+ in about 2-3 days. It was pretty great, but it uses way more tokens (TPS is fast, but it's way more tokens per turn) than Sol 5.6 which I found to be about it's equivalent at the time (on medium or high, with DS on max). My cache rate was around 98-99%.

It would definitely cost me more per month than a x20 ChatGPT or Claude plan, probably around $400+ was my estimate at the time. This was with Fireworks (ZDR) which has since increased their prices (and got slower!).

That being said, very impressed with the model, and looking forward to what comes next. As the frontier models become less subsidized, the open models will become more appealing.

P.S. There are subscription plans for open models, but I've found most of them to be extremely slow, have model throttling (only so much of model X), and also very sketchy about training and data retention. No thanks! If you want to share your data, just use Muse Spark contributor. Seems impossible to beat that on price per task if you don't mind feeding your data to the Meta machine (spoiler: I won't).

This is my experience, too. It's a great model, but it burns tokens if you use it heavily for work on complex domains.

Edit: others have noted the provider and harness matters. My experience is with opencode.

same, used a wrapper around cc and i was spending up to $30 a day with basic stuff

Maybe CC does something that breaks the cache? I cannot recommend Oh My Pi enough. Every default is galaxy brained, and it plays incredibly well with deepseek flash 4.1. My favorite coding harness rn for sure.

I found omp used quite a bit more tokens than my fairly basic pi setup... but most of those tokens would be cached with DS V4.1, so maybe worth if there are gains elsewhere.

Were you using OpenRouter? I've used 1.8bn tokens in the past week from DeepSeek themselves and 99.2% were cache hits. Total cost was $18.13 usd.

For readers wondering, OpenRouter isn’t capable of caching as effectively as DeepSeek is because they will, for instance, switch inference providers in the middle of a session.

The way I use openrouter is I find a model/provider combination I like then pin all requests for that model to that single provider.

How do you do this?

provider: { order: ['deepinfra/turbo'], allowFallbacks: false, },

https://openrouter.ai/docs/guides/routing/provider-selection

What harnesses are capable of doing this provider pinning when doing requests on openrouter? Can DeepSeek harness (dsh) do it? Or opencode etc

I stopped using opencode because it has some issues with caching, so I suppose it doesn't do this

Open code and pi can do it. Unironically, I let Claude configure these things for me.

Deepseek plugins* all the way down. It should be easy to add support for this.

Using this interesting framework called Cordis I've recent discovered.

What pests said. And you can make a preset and pass "model": "@preset/deep-seek"

If you disable all other providers but DeepSeek in your OpenRouter guardrails, is that effectively the same thing?

Sure but you still pay only OpenRouter :)

Interestingly, OpenRouter will serve inference at cost for a lot of models.

No? They charge 5.5% (I believe) on every top-up. Which is the equivalent of charging you 5.5% on every request.

Ah, so that's where the profit comes from. Very payment-processor like move, makes sense that Stripe picked them up.

How can they? If you use openrouter Deepseek fine but can’t you select the model direct from Deepseek you’re basically getting api access models that way.

Fireworks directly. At the time they were the best value of cost, speed, ZDR. They got slower on me though, but I think they are retooling, so maybe things have or will get better again. I think fireworks is primarily for when you want to do your own training on top, which I wasn't doing.

DeepSeek, DigitalOcean, GMICloud, and NovitaAI were the only OpenRouter providers that didn't lead to major performance degradation for me

How on earth you can do 100 dollars in 2-3 days with DeepSeek? I have 7 agents in omp running 24/7 every day. I use maybe 10-15 dollars a day. A rarely see a session going over 2 dollars. My maximum is maybe 3.5 dollars and that session took three days.

What harness you are using?

What are those agents doing? I am out of the loop. Bitcoin mining? Blogging? Reddit bots?

I’m lost when I read these sort of comment chains. Free Gemini works just fine for me. Maybe it’s because I don’t use it for programming? How many programmers really exist out there? Surely it can’t support the weight of investment that exists in AI already. It’s just such a small pool of the human race.

First, they come for the programmers, and next the mathematicians. Then it will be the biologists, lawyers and doctors. Humanities will stake it out a little bit longer because AI isn’t human, but AI companies would be dammed if they don’t try. Eventually, with advancements in robotics, stabs at increasingly more physical sciences will also be attempted. Eventually, AI will have its hand in the pie of all knowledge work, if it is possible. Not to mention all the roles like tech support and customer service. Once they have gotten as far as they think they can go, they will try to turn up the prices. However, they might struggle to do so as models are becoming a commodity. This is why they are arguing for regulation and stating that only they can tame these beasts.

I think as long as we have open weighted models there will always be competition. Sure Anthropic could raise their prices but it wont take long for someone else to undercut them. The quality may not be as good but people would probably prefer to pay 1% of the cost for 75% of the quality.

There's plenty of competition. Of course, Anthropic can raise their own prices, but people will just switch. They've done that in the past, when other companies had better offers.

> First, they come for the programmers, and next the mathematicians. Then it will be the biologists, lawyers and doctors. [...] Eventually, with advancements in robotics, stabs at increasingly more physical sciences will also be attempted. Eventually, AI will have its hand in the pie of all knowledge work, if it is possible.

You seem to think raising productivity is a bad thing?

> Humanities will stake it out a little bit longer because AI isn’t human, [...]

This is really silly. Many languages other than English don't use related words to describe 'humans' and 'humanities'. Will it be easier in those languages? Should we rename mathematics to 'humathematics' or so, to make it harder for AI to take over?

> You seem to think raising productivity is a bad thing?

Higher productivity means very little to people whose labor plummets in value over the course of a couple of years. One day they won't be needed anymore, maybe that's good for you.

Ever since at least the industrial revolution we had this same song and dance play out again and again. Yet, unemployment rates around the world are the world are fairly steady (and vary mostly in line with how business-friendly local regulations and taxes are, and so far not much with the current state of technology).

I'm a software developer by trade. I welcome the coming brave new world in which machines can do all the software.

At the moment, they ain't quite there yet, alas.

With that argument we'd all be unemployed ever since agriculture (which 90% of the populace was working in a few hundred years ago) was invaded by machines. Turns out, people get other stuff to do and are as overworked as ever.

> How many programmers really exist out there? Surely it can’t support the weight of investment that exists in AI already. It’s just such a small pool of the human race

I dont consider myself a programmer but use LLMs almost exclusively for coding.

The number of people able to create useful software today is much much larger than it used to be and arguably a minority of these people are/were "programmers"

Why did this get downvotes? :o

Free Gemini has at least one subtle mistake on every scientific or programming task I ask it. It's usually in the right ballpark but I guess I do stuff outside its training locus.

US has 1.4-4.4M programmers and devops, but considering they are more costly than many other professions, it’s quite meaningful to economy. Plus now you have non programmers doing agentic programming to speedup their work - i know nondev project managers, financial advisors, marketers and lawyers rolling their own mini apps now.

Also even if your Gemini is giving you nonprogramming output, underneath the model is most likely generating code for certain tasks.

I have deepseek agents doing email responses, with real tools (think running quotes, gathering info, scheduling things) and running business processes that used to be done by $35/hr administrative type people. And the capabilities are expanding every day as I learn how to build scaffolding around the model.

I am a tech lead for a useful but non-essential PaaS my company subscribes to. Their product manager occasionally sends me clearly AI-authored emails. I ignore them. I am a believer in the usefulness of AI, but nothing says I don’t care more clearly than sending me slop.

I can’t overstate how bad of an idea I think using an AI for customer interaction is.

I think it depends on the industry. In my company (insurtech) a large percentage of our sales come from AI engagements (phone and text). The customers know they're talking to an AI and they can elect to talk to a human at any time but many times they don't.

But yeah, send me a non solicited AI slop email or worse, political ad, and you dont get the dignity of me saying stop to unsubscribe. Straight to spam for you.

People have been trained to "press or say" a number while going through an AI voice menu on phones for over a decade. So a chatbot looks great by comparison in that context.

I wonder if various companies have recordings of me screaming "OPERATOR!!!!" or pressing 0000000000 ****

Optimum is probably using a middle ground: Use AI to do all preparations like suggesting quotes, gathering background about the customers query etc but then have the human write the email for personal touch. Still much faster than doing it all by hand.

> I have deepseek agents doing email responses

So, spam?

Not all communications done by a non-human is necessarily spam.

Spam is unsolicited, unwanted email.

If someone is asking you for a quote, and your tool replies, it's not spam.

It might be slop for all I know (or might not be!), but it's not spam.

Similarly for arranging meetings with parties that you already have a relationship with.

You have an AI handing out potentially legally binding quotes for you??

Work for my company. Research, code, analysis.

pi. I wasn't even going that hard. I checked the logs for Sep 18 and I did just shy of 3b input with approx 98.5% cache and 5.8m output, which cost around $35. Most of the was a Rust code review exercise with 1 driving agent and a varying number of subagents (up to 6 some times). I do the same with with Sol med/high driving and Luna x-high reviewing and get at least as much done if not more in a day, but I'd use up two x20 weekly allowances for the week. Worth noting that token cost isn't super meaningfull on it's own, because DS is super token heavy (but also great at caching) compared to Sol. (my stats show DS uses 3x the tokens as Sol)

The shape of my work changes obviously, so it'll vary, sometimes more, sometimes less. For example, fixing all of the bugs and defects I found that week was 2-3 times the effort and chewed through my ChatGPT allowance, but I had banked resets...

Also worth noting that codex models have been kind of all over the place recently with their usage... and it looks like costs are changing again.

You might be overusing subagents. Especially with a chatty model like DS, you’ll be wasting millions of tokens on re-discovering the project and facts instead of actual reasoning.

I agree, and I started sending well defined review packages to Luna x-high instead, which is how I have this setup when using Codex (Sol drives, Luna async reviews, and Sol keeps moving). Same process with DS really dropped my DS usage by a lot and I think that might actually be the secret sauce. Especially if you use Luna through a subscription (probably the entry pro level would be enough). I'm not sure what a Luna equivalent open model is though. 5.6 Luna max was catching a lot of issues while my implementer keeps rolling. Last couple of days I had 3-4 Sol Mediums running with Luna x-high reviewers (async reviews) and one Sol medium orchestrator and Astra X-high to plan everything out.

Gotta be honest though, I don't love fiddling with this all the time. I would rather be working on my projects than evaluating my usage. Having DS in my back pocket should i need it is a relief. The providers get fiddly though too.

Such a high cache may be a sign that your agents are using tools inefficiently, thus taking too many turns. Or, like someone else said - you may use tools inefficiently many agents that rediscover stuff etc.

Pi out if the box tries to optimise system prompt size, which is not necessarily good and will cause exactly this effect for all but the most simple tasks.

What you want is to give enough context to the agent to minimize the amount of searching within the codebase etc.

If you want to track cache, what you should do, imho, is to check if you have cache expirations mid sessions (ideally you should not), and if you don’t then lower cache use is actually better - it means that your model doesn’t reread what it just wrote.

There's a bunch of skepticism in the replies but I ran over 100 tasks against DeepSeek 4.1 Flash and Sol (among others) and I can confirm, it is in fact a little smarter than Sol and a little more expensive than Luna. https://slopcop.com/power-ranking?pricing=api

I also spent $280 on DeepSeek doing the tests (direct to DS, not OpenRouter). I suggest that if you can't conceive of anyone spending $200 on DeepSeek, you're not being ambitious enough!

Is slopcop your domain, because that is awesome. Wishing you great success with it.

It is. TY!

So if money isn't a problem, opus is the best?

Yes. And on subscription pricing it's even cheaper per task solved than DeepSeek.

wouldnt taht be technically no? Fable should be best if money _really_ isnt a factor?

tbf, i barely feel the difference with opus 5.5 anymore either.

How have you spent hundreds of dollars? I’ve only spent 11 and I’ve been using it for four months!

Same experience. I often see people say how little they spend on DeepSeek v4.1 flash, but when I put 60 bucks into my account, it was gone in a few days of non-exclusive use. I'm actually curious what the difference is. I used it through pi and opencode, but the harness seemed to have no obvious impact on usage.

Maybe they just use it less. If you code all day you can go through a billion tokens.

check your cache its maybe?

Note that $400 a month is a typical car payment. For the same money -- well, OK, probably a little more -- you could be paying off a set of 4 RTX 6000 cards that will run 4.1 Flash all you want, all day and all night, at hundreds of tokens per second.

Electricity costs money.

And there's a lot of labour and effort involved in setting these things up and maintaining them. People don't even run their own email servers, even though the hardware side of that is trivial.

Electricity costs money

Not $400 a month, it doesn't. At least not around here (we average $0.12/kWh and I don't personally run my cards over 300W.)

And there's a lot of labour and effort involved...

Theoretically, the people who hang around HN are more likely than most to be capable of the labour and effort of setting these things up and maintaining them.

People don't even run their own email servers...

People don't run their own email servers because a convenient coalition of spammers, standards bodies, and large email providers have done their best to make running one's own email server almost impossible.

4x RTX 6000 running at full load with relatively cheap electricity ($0.25/kWh) is around $450/month.

I did the math and decided it’d better to pay for tokens than to buy the hardware and generate them myself.

> Theoretically, the people who hang around HN are more likely than most to be capable of the labour and effort of setting these things up and maintaining them.

'Capable' doesn't mean your labour has no opportunity costs.

I think my argument is easier to attack by noting that you can use AI to substitute for much of that labour.

That's what I did personally, not being a Linux guru. But to your point, I still had to spend a lot of time and effort dorking around with the hardware.

It's worth it, knowing that there are no rugs Sam or Dario or anyone else can pull.

> It's worth it, knowing that there are no rugs Sam or Dario or anyone else can pull.

At the moment, it's still very easy to switch from one open weights model to another, and even between the closed models. So the 'no rug pull' property is nice to have, but not as big of a deal.

Power averages $0.30 to $0.45 for a business in the UK if you convert it to dollars. Smaller businesses without load shedding agreements are at the middle / top of that. (there are lots of people complaining about it). Maybe I just have expensive power, but $0.12 per kWh is pretty cheap.

> Theoretically, the people who hang around HN are more likely than most to be capable of the labour and effort of setting these things up and maintaining them.

Theoretically, the people who hang around HN have immense opportunity cost when doing this. Being capable does not mean it takes no time. Time they could be using for something more profitable (and fun?).

That's one way to look at it. Another way to look at it is to point out that the potential cost of delegating cognitive resources to companies like OpenAI and Anthropic is unbounded.

The AI models we currently have still maintain their working state with a finite and laughably-small context window, but that is already starting to change. My .claude directory contains over 300 .md files that I didn't put there myself. Their contents are eye-opening. The question of who owns, stores, maintains, and can access that data is going to become insanely important over the next couple of years.

If you thought LLMs themselves were disruptive and contentious, just wait until the fight over object permanence gets under way. That's when owning your own box full of graphics cards is going to become important. My own bet is that I won't care too much about the electric bill or my opportunity cost when we all find out what the AI labs really have in mind, and what they're going to have to do in order to justify the valuations they're seeking.

TL,DR: it's not about the tokens, IMHO.

Ok, but I have two of those cards plugged into fiber internet, and they consistently rent out for $1500+/mo on vast.ai. So your example would be $3000, minus a couple hundred for business internet and electricity… that still buys a LOT of $200/mo frontier model subscriptions.

Seriously? People are paying you $1500/month to rent your 2x RTX 6000 Blackwell cards?

I think the point was this was the spot price he/she would be paying if he/she had to rent and didn't own them

You can buy 2 of them from Central Computer any day of the week for $30K US, and that's not the lowest price I've heard lately.

So if you can spend $30K and immediately start mining $1500/month out of thin air, that's a pretty nice investment even if the electricity costs $200/month. Two years later the cards will have paid for themselves entirely and (I suspect) will still be pretty useful.

It does argue in favor of just paying OpenAI or Anthropic for tokens, though.

You do need fairly high end workstations to put them in, adequate cooling, power and very solid business internet as well. I've put more work into keeping them online than I had expected. But overall the economics are fantastic for it right now, the big question is how long it will continue.

If you break down the pure profit per work hour you put in (after all the expenses), how much is that?

How important is it that you maintain a good availability record? I have a few RTX 6000s in a server that I use occasionally for NDA'ed contract work and for experimenting with open models, and now I feel stupid for just letting it idle at ~1 kWh when it could be earning revenue. I didn't realize that anyone would care to rent prosumer-grade GPUs.

At the same time, when I need to use the hardware for something, whoever is renting it from me at the moment is going to get unceremoniously booted, and I imagine they are not going to be happy about that. I assume that vast.ai's providers get uptime ratings that drive their work allocation, right?

Yeah - that is the issue. Doing it that way is going to nuke your reliability rating and you will have to cut prices drastically to get rented.

What I do is… rent on the same platform when I actually need to use a card. A benefit is that if I need a burst of more power, that’s not an issue since it’s available from other hosts. But obviously there’s an inflection point of first party utilization where it makes more sense to own.

Yes, they have been near 100% utilization on vast.ai at $1 to $1.25 depending on demand for the past two months. They're in two separate workstations, which I purchased for around $12k each before GPU and RAM prices went nuts this year.

Now, I have absolutely no clue how long this situation is going to last! But the economics don't really work out for local models while it does.

which year are they from?

I bought them direct from PNY the week the RTX Pro 6000 was released. Prices are almost double now what they were then.

Doesn't a RTX 6000 run at about $100/month for electricity? Do that x4 and I don't see where your savings would be coming from.

claude 5.5 with price discounts is a step up in terms of quotas, as well as gpt 6.1 sol, i still feel like for complex things claude is way better vs anything else. also for the topic started, opus 5.5 and sonnet 5.5 are much faster too. I was using grok for that reasons just to do faster changes, but now i'm considering to cancel my cursor subscription, clade is much better and it's also as fast as anything else especially with workflows.

> if you don't mind feeding your data to the Meta machine (spoiler: I won't).

How is that different from feeding into the OpenAI or Anthropic machines?

I think you may have a problem with caching, can you check how much cache % you have?

Whats your harness?

It’s easy to hit your quota. “Speed up the compilation time of this C++ codebase. Feel free to use several subagents to search through the files in parallel.” That’ll cost you about $200 for a codebase of ~1,000 files.

Subagents are like trading derivatives. You can lose as much as you want.

What bothers me about this whole AI tokenomics situation is the lack of transparency. OpenAI and Anthropic have to perhaps be the most opaque companies in existence wrt their offerings. There's like a thousand variables that they can change on the backend at the push of a button which can wildly swing API spends within the same model (partly also due to the non-deterministic nature of LxMs, but still), and there's no objective way to measure them other than vibes.

When the regulations do arrive, I think they should really focus on AI companies and API providers being more transparent wrt how they're billing their customers. Because right now, it's a totally vibes-dependent and a mess.

Yeah, you might get away with a singular 'wrt' with some consternation, but two?

It's all Gacha for business.

And it's all measured in "intelligence", a completely meaningless term. For coding i'd be much more interested in how much context actually works, what the complexity of algorithms it can understand and create is, for what languages. How much it manages to follow existing structures or that is just adds ad-hoc machinery to pass the test, etc etc.

A smaller model in the same generation will never be the same as a bigger one, assuming this is a smaller model, and the same generation, as naming implies, it will not be comparable, it might be on the benchmarks, even on the benchmarks that matter, but the whole story should also give the drawbacks.

It's still insane that they stopped showing you all the tokens you pay for. They could inflate the billed reasoning token amount by a lot before it would raise any eyebrows.

Internet Ad business has been like that for a long time with bot click & Co.

Ain't that the truth.

In Search Advertising, the amount you pay (under GSP Auction) is a function of your pCTR. And guess who determines your pCTR? The Search Engine itself! :-D

Totally hand-wavy and non-objective ... just like how employees are billing their employers.

you REALLY need to start working on your own local models, no not because your 4080 on a pi harness with qwen 3.8 will be mediocre today, but because in a year when OpenAI and Anthropic have fully enshittified there will be NO regulations and they'll have your chestnuts in a vice and you'll be glad you can tell them to rm -rf

Why use subagents at all

Time is money. Parallelism is very helpful optimising one to get the other.

Money is also money, which anyone who makes heavy use of parallel subagents will quickly learn.

This truism is intuitive to everyone but always funny to me how everyone never has any time, needs to save time, needs to hire staff workers for every mundane job and robots can't come soon enough… all so we can binge watch Game of Thrones and 90 day Fiancé.

And watch 10 hours of football on Sunday for our DraftKings bets.

Money is money too. Increasing contexts costs non zero money, even with cache hits. Also context rot is a problem that subagents help with

> Time is money. Parallelism is very helpful optimising one to get the other.

Parallelism is fantastic when it actually speeds up the entire pipeline, but in my experience most people's jobs (at least the ones for which AI is currently relevant) involve a lot of overlapping "hurry up and wait" branches that drastically blunt the real benefits of that sort of parallelism.

There may be specific situations where it makes sense to do it, but just immediately going full gastown on anything AI related seems like such a giant waste to me, of both money and finite world resources.

Preserve context in the lead chat - let the subagents fill up their own contexts then only return the necessary information.

Why not just have an agent that can branch its context?

Forking conversations have been a thing for a long time, and it's not the same thing (e.g. you may have 50% of your context used up at fork time that is carried into "both agents" afterwards)

Also sometimes _not_ having the context is what you want.

Eg when I have the AI do a self-code-review before bothering a human, I also want to the AI to give the draft PR to a sub-agent that only has the publicly available context that a viewer of the final PR would have; and not all the accumulated reasoning that lead to the writing of the code in the PR.

It is exactly the same, you’re just doing it wrong.

Because two agents are faster than one.

Why hire a junior developer if you have a perfectly competent senior developer already on your team?

You can do a code review on a "less capable" model that costs less, and the key model gets its output / summary, then you can have that model build a plan, and feed it to cheaper models. It's a more efficient approach than just running everything through Opus, and now that Sonnet is a lot better I'll probably use them more frequently, one thing to note is don't ask it to spin up endless subagents, I'd cap it to 2 or 3 at a time, otherwise, yeah you'll hit your limit extremely quickly.

Code reviews also work better in sub agents because the reviewer agent didn’t write the code being reviewed.

(More precisely, doesn't have the reasoning from when the code was written)

The models get dumb as context fills. Subagents allow them to accomplish a task with minimal context rot. You can also use cheaper models for subagent tasks

here's a simple task. Collect all info from unstructured email signatures, 3 years of small business emails.

Opus directs and starts haiku/sonnet subagents. Much more efficient than opus reading it all.

Without this explicit direction, Opus burned through usage to create complicated regex parsers & 8 excel sheets.

Afaik there is just pay-as-use with Deepseek

> Subagents are like trading derivatives. You can lose as much as you want.

Excellent pithy warning.

Sadly this is true - for individual folks on the lower end of the spend spectrum.

But there’s a point on that spectrum where the ability to run multiple experiments in parallel, even with a significant amount of (one time) wastage, is overall more cost effective than the alternative.

I have done something similar on a larger repo on claude code max and it didnt even fully exhaust the 5 hour limit.

I spent the weekend trying Deepseek 4 Pro on a Linux porting project and it led me down a complete rabbit hole where Linux wouldn't even boot by the end of the weekend. Waste of $120. Switched back to GPT 6 on Monday and Linux is booting again and I'm making progress.

The only thing I've found Deepseek and Kimi good for are security tasks that GPT refuses to do.

This is a summary of what Deepseek did and got wrong:

Lost the proven baseline: changed kernel source, configuration, compiler, RAM geometry, MMC width, and peripherals together. Matching an upstream commit did not preserve local boot fixes, making failures difficult to isolate. Misidentified an image: a file labelled “r18-known-good” actually contained the r23 parent bootloader. Filename-based reasoning replaced verification of the artifact’s identity and provenance. Shipped inconsistent boot contracts: flash-16b’s loader read too few kernel blocks. Fresh2 changed the device tree without updating the loader’s expected length and CRC, creating deterministic rejection before normal Linux handoff. Patched binaries without maintaining reproducible source: loader constants diverged from source, a separately compiled cache-flush length remained stale, and assembly used an oversized stage-two slot. Their causal contribution to hangs was not established. Overstated diagnosis: claimed failures were definitively in U-Boot, blamed compiler or IPU changes without controlled isolation, converted noisy observations into confirmed hangs, and neglected persistent journals as an alternative explanation. Mistook compilation for integration: framebuffer registration was incomplete, timing success handling was inverted, BT.656 selection was unreachable, encoder overrides were missing, and audio lacked software clock configuration. Misread hardware evidence: asserted interrupt-free PMIC operation, assigned RF to the wrong SPI controller, confused regulator identifiers with register addresses, and described repeated encoder writes as unique registers. Overclaimed results: treated kernel/probe indications as userspace success, presented earlier discoveries as new progress, and omitted failed flashing attempts from the final narrative.

You mean 4.1 Flash which is the first great Deepseek? The one that actually surpasses Opus in my books now.

> Deepseek 4 Pro

There's your problem, 4.1 Flash is significantly better and cheaper, to the point where the official DeepSeek API is going to (or already has, I forget) redirect requests for Pro to 4.1 Flash, and adjust billing accordingly too.

4 Pro is still offered by providers I'm sure, since it's open weight, so I can understand making that mistake.

That's an unfortunate experience. Think of v4.1 flash as actually v5.0 flash. It's night and day compared to the 4.0 flash (and 4.0 flash was unintuitively better than 4.0 pro). I would re-evaluate with v4.1 flash. I'm not saying it better than Sol or anything, but it's in the ballpark.

There's a difference between "write this function for me" coding agents and "build this prototype from end-to-end". If you're doing the former, deepseek is fine. If you're doing the latter, it's not gonna work, and that's where the extra intelligence is most valuable.

build this prototype from end-to-end, is this how people build serious software with AI?

For a prototype or a 1 or 2 use tool, yes, this is exactly how serious people are building software.

This is how ChatGPT, Cursor apps are built. They spent so much money on PR stunts, but "thousands of agents" can't make an app that doesn't freeze on each keystroke. Not even talking about user-friendly ui

Weird, ChatGPT has always worked really well for me.

Early on their software was like this, the original ChatGPT desktop app was basically unusable and full of memory leaks that would tank the software. They’ve long since fixed that though

The new ChatGPT desktop app is a dumpster fire though, it can’t even scroll properly when streaming the answer.

Such issues are pretty common among people I talk to at least. I hear comments about it pretty often, both when it comes to ChatGPT and Claude.

For myself, with ChatGPT for example, it regularly gets extremely slow if there is a lot of text in a single conversation. Especially if I try to scroll up.

Some pages have been strangely just broken for a while now as well. Usage analytics just renders lots of these duplicate "Usage history" components, where the data just never loads: https://imgur.com/a/vpcQIiw.png

I can totally imagine a scenario where some agent built it, another tested and approved it, and nobody at OpenAI even looked at it once or knows that it's like this.

We use codex in vscode. Use agentic development but, agents only do things that We agree upon. Like We create impact study, proposals, feasibility and then implementation plan with full regresison, test coverage and end to end smoke test. A new fetaure will be given to front end only if it satisfy all our requirement in real API smoke test. It is not fast. but still we think we get 10X productivity gains and clean code, consistant quality and maintainability. And we dont use AI for front end. We have also found C# is more effective for agentic development. With Roslyn APIs, strong typing, structured namespaces, and compiler level semantic information, agents can navigate and reason about the codebase with relatively little context. In our experience, C# is surprisingly token efficient for large codebases cmpared to Go or rust.

You're not gonna get something that's ready to ship, but as a first pass to get something running yes. Let's you explore far more ideas with only a few hours of agent time.

And you can’t use deepseek with a loop approach for creating such PoCs?

You absolutely can. On a large codebase. Written in multiple languages. I also use gpt models and for most tasks I prefer ds. It can orchestrate really well, and it stays on task without going off to reinvent the w̵h̵e̵e̵l̵ state machine.

I started KeenLore (an emotive audiobook creator) that way. I gave it software specifications, languages, JSON schema definitions, container requirements, hardware configuration (8GB NVIDIA T1000 GPU, 96GB RAM), and zero user interface mockups. For the second round, I asked it to build a completely independent, re-entrant, and data isolated demo system on top of the web application. The demo application included voice generation using one of its voice designs. Here's the output:

https://www.youtube.com/watch?v=WAeHgE94rVo

The system performs quotation attribution on my local hardware for my near-future, hard sci-fi novel (having nearly 500 quotations) with over 97% accuracy.

The initial prototype was developed quite quickly, but numerous successive iterations were required to fix numerous gaffs by Opus 5 (because it doesn't actually _understand_ what it takes to make general-purpose audiobook narration software).

I don't think most prototypes are serious software.

Oh buddy you have no idea what a good plan and agent harness can do with deepseek…

I am doing mostly hardware drivers recently, it works fine for complex work.

This was my experience 3 months ago. I had an Android app that interacted with a Bluetooth device that I wanted to reverse engineer and build my own Linux app for it. DeepSeek was struggling really hard. Claude did it end to end after 3 or 4 prompts. To be fair, I was using a web interface for DeepSeek and the CLI for Claude; maybe that makes a large difference.

"build this prototype from end-to-end" works fine with DeekSeek V4.1 Flash, the problem occurs if you're not only building a prototype but want a finished product.

I'm mostly using deepseek 4.1 flash via openrouter, the way I'm doing stuff is:

1. write a sketch of a spec by hand

2. have the llm review the document and question me until it can generate a spec

3. review the spec and revise where needed

4. have it write an implementation plan

5. another round or revision/review

6. executing the plan step by step through the plan, plausing between each step to see if we are still on course and if the decisions it made track with my understanding of what we are doing.

I've been working for a couple of hours tonight, the total cost of the session is €0.6.

it's not the build this thing end to end, but also not quite write function x for me. It is still a lot of manual review, but I find I really need it to even discover what I actually want to build. I just cannot imagine building something in a single shot and getting something that actually has value (unless it is basically a clone of an existing thing). To me the whole value of ai right now is that it's now very cheap to build custom software that exactly matches your preferences.

One of the big differences using the better models is that you don’t need to hold their hand anywhere near as much. Fable is crazy expensive, but I’ve seen it just one-shot some remarkably complex projects. The code it produces is much better, too.

yeah i just think that if I gave the initial pitch to Fable it would have successfully built something that wasn't exactly what I wanted. maybe it is because i went to school for design, but how something works is so important to me and it is hard to get to that point without actually thinking about it step by step and visualizing how interactions would work. I just cant imagine a single one-shot prompt containing enough information to for the model to know what to build. I guess if you have Frontier model money like these silicon valley freaks you could just iterate by asking it to tweak what you didn't like until it is right.

I do the same but with gpt-6-sol (mostly via github copilot cli).

The one-shot capabilities of frontier models are nice for demonstration purposes, but not actually all that useful to me - the result is often an amalgamation of ad-hoc ideas the agent came up with, poor UI and a hodge-podge of a data model, but at least demonstrating clearly where the specs are lacking.

I agree that in practice, iterative design with lots of review and hand-holding are needed to get quality results. Generated code (I mostly program C++) is often bloated and not succinct or "simple" enough to my tastes, so it requires multiple rounds of cleanup.

For core logic, I find having the agent review code is often faster and more valuable than having it implement it in the first place - it can find and fill in the things I missed.

For my own sanity it is very important that I remain in control and understand the generated code when quality and maintainability are goals of the project.

I've been trying to use DeepSeek V4.1 Flash more and been very impressed. My current (very rough) rule of thumb is that an Artificial Analysis score of ~40 is the crossover point for "good enough" for most of the things I need to do with coding agents.

47 is my crossover for serious things (e.g. Grok 4.7 is below the line and GPT 6 Sol is above the line). I mean, Opus 5.5 is way better, but GPT 6 Sol still gets the job done for anything that doesn't require design thinking.

Although I do think Luna 6 max is ok for some basic things, would never use it for coding myself.

What's the most important task you would/wouldn't trust with an agent below the line?

Had the same experience. I got rid of my pro sub of OpenAI. I’m really freaking impressed.

Where are you inferencing DS4.1 flash reliably ?

I use OpenCode Go. I've also used the DeepSeek provider through OpenRouter a bit in the past and it seemed solid.

The intelligence difference between models like DS4.1 and Sol/Opus is NOT negligible.

If both DS4.1 and Opus can complete the tasks you throw at it at good enough quality the differences are negligible.

Who cares if your car can go 200mph if all you need is 60. If my requirement is 60mph, I want a faster 0-60, not a higher top speed.

Or just a cheaper way to reach 0-60

Specially if using the 200mph car when you need it is just a /model away.

A car that feels safe to be driving at 200 mph is going to feel more comfortable at 60 mph, compared to one for which 60 mph is at the very limits of its abilities. Analogies only go so far so I'm not sure there's anything to be learned from that though.

Yes, even for crud apps there's a huge variability in how nicely you can make them, and how many mistakes or footguns models make along the way. I'm very curious to what level these people are building to.

Opus 5.5: $4/$20

Deepseek 4.1: $0.02/$0.60

Just to illustrate how cheap the Corolla is in your analogy. Also Opus output would be $50 without competition.

I suspect most people in this position would be using the subscriptions, though.

A $20 Anthropic subscription is $500+ equivalent of API credit. You can build quite a lot on the $20 plan and get to use the best model.

Their API pricing has healthy margins built in.

The premium price is only worth fo the hardest problems.

For CRUD shoveling, models like DS4.1 are enough.

And the intelligence gap between cheap and premium is closing, as can be seen from the title of this post.

The issue is once you solve the hard problems, the lower models start messing things up that were working and reverting all fixes for the hard problems. They'll just go off and do dumb stuff.

In the Artificial Analysis index, MiMo 2.6 Pro is smarter than GPT-Sol 6.1 Low at the same cost, and only slightly dumber than Medium. MiMo 2.6 Flash is marginally cheaper and smarter than GPT-Luna 6 Max. (There is no GPT 6+ Terra, which would otherwise be in that range.) These are not negligible or trivial results.

I've been using Claude Code at work and OpenCode for side projects for a few months. Every OpenCode model I've tried always felt subpar compared to Claude, but good enough. But it changed with DeepSeek 4.1 Flash, I've been using it for the past few days and I've come to forget I was not using Claude, it's a really good model and it's basically free for my usage (I used it almost all the weekend and spent ~$5)

What do you do? I’m 200 bucks deepseek flash in about 2 months and it’s increasing

I use it as main Hermes model that orchestrates codex/droid harnesses with subscriptions for heavy dev work

I do have ChatGPT as main assistant that sets direction and delegation of projects to Hermes

At my increasing usage, kind of 200 usd subscriptions makes sense and max out on Luna max

I put 5 dollars at deepseek a long time ago, and somehow it has never been fully spent, can’t imagine anyway to spend 200$ on that thing

Recently, drivers for a bunch of obscure hardware. Lots of c\c++, that i am ok with but not enough to make hardware drivers (i am just impatient). Just using pi agent with a few plugins.

How complex are they? Drivers vs an entire application would be a big difference in token usage, and it could also depend on the type of work being done.

Complex enough where performance matters. I've done entire applications, client, server, infra and ci as an experiment with deepseek v4 pro a few months ago. Works for that too.

I've spent $200+ on deepseek and this is for making a multiplayer FPS game. Trust me there are use-cases.

And no it did not deliver. A lot of it was re-done by Astra

> I've spent $200+ on deepseek and this is for making a multiplayer FPS game.

Why do you expect that $200 will give you that on ANY model? Multiplayer FPS games are very difficult to make, no AI will deliver that today.

Deepseek v4 pro was able to create procedural dungeons, placeholder 3d assets, 12 weapons, and menus. Struggled with ragdoll effects, and gun modeling in blender.

Astra was able to model low poly enemies, rig them, do simple animations, and greatly improve procedural generation. I have it modeling assets in blender every day, which I often have to go in and fix

Nice, that's all single player stuff.

I have a roadmap for server authoritative coop with rollback and most physics disabled. I can barely think of things to throw at 6.1 Sol to use up my weekly credits to be honest.

I think the only answer to this is you're just not using agents enough, because even with the very cheap pricing, it's still easy to rack up a large bill.

>At this point i am OK with being 6 months behind the "frontier",

Indeed. A few people [1] pointed on twitter that current best of Open Source Open Weight models are better than all the frontier model 6 - 8 months ago. And we heard Deepseek v5 and MiMO v3 are quite a leap.

I am not understanding the value proposition of these frontier model. Or at least not to a point they are looking at $1T or $2T market cap. Especially MiMO which may be ( correct me if I am wrong ) the most transparent model we currently have?

Or did I overlook or missing something? It is quite scary if true.

[1] https://x.com/samsaffron/status/2104903300298469877

Sonnet 5.5 on low is leagues ahead of deepseek v4.1 flash in price per _result_ in my testing. I say this as someone who has the same attitude as you and really wants open weight models to be competitive and succeed.

The biggest problem with DeepSeek 4.1 Flash is all that it hallucinates a lot.

[1] - https://www.youtube.com/watch?v=P4dTq4X8bqk

Everyone uses it differently. Not sure what the intent is of these kinds of posts.

Some people are able to take advantage of the increased capabilities that frontier models give us. For others, working with open weight models (that are extremely good in their own right) is good enough.

Deepseek models are very impressive. No one's debating that.

But frontier models are definitively ahead. Perhaps not by much in some areas, but there are clear gaps.

And that is kind of the point I think. Open source models add a much needed source of competition to the world of LLMs and are very much an important counterweight to the dominance of Anthropic and OpenAI.

A lot of people having different pricing experience. I think it’s important to understand that caching can differ, than if the agents spend waiting on code, or consume a lot of content. It depends on how you structure you codebase and how explorable it is, how much effort you set and probably some other issues.

For raw productivity most of what works is best and switching will cost you getting on use parity with other models, as you need to learn what they good at, potentially how the tool works and how to prompt it best.

For tasks that you implement in code, you should have benchmarks and evals.

That said for me was Luna a huge leap and 500+ of cost savings a month

> Not sure what kind of usage can justify $200/month of either openai or anthropic,

Making more than $200/month with the results?

A bit tired of spending $200 out-of-pocket for openai. What do you use as harness? (for me the harness if half of the benefit... controlling my PC, working from phone, etc.)

OpenCode works nicely for me. You can connect it to the DeepSeek platform with an API key.

I use Claude Code + eternal terminal + tailscale + tmux + some custom skills to get notifications through nfty.sh.

I get the same UX on every platform, works perfectly on very low bandwith environments such as in a cabin, in the subway or in the middle of nowhere.

I tried using other harness such as Pi and opencode but I did not like them. If Claude Code gets weird I can swap in an instant.

You just need to follow this guide and disable artifacts in Claude Code's config: https://api-docs.deepseek.com/quick_start/agent_integrations...

I’m curious about your use of Tailscale, is that for you to reach a local LLM remotely from anywhere?

I have a big beefy desktop at home which I ssh (using EternalTerminal instead of raw port 22) into. I built it last summer right before the prices got very expensive. It's headless so I use a cheap Macbook Air to connect to it at home and I use Termux on my phone to continue working from everywhere.

Dude prompt your agent to set up always on remote connections via systemd… then you can drive from Claude or codex mobile apps natively over their native hookup. Works great.

I don't want to rely on Claude of Codex mobile features at all. Also none of this works when you use third party LLMs which is what the current comment tree is about.

This. There's so many better ways to drive a fleet of agents this way than using "remote" features of which I don't trust, anyway.

What makes it run well on low bandwidth environments? Is it eternal terminal? (I am not familiar with that).

I found that ssh is pretty bad in these scenarios, so my usual herdr over a phone doesn't always work properly.

Now I'm using paseo which in principle solves my issues properly, but unfortunately it's "reconnection" and state sync is pretty slow (probably going over their servers).

Yes Eternal Terminal has been a godsend for bad connectivity and switching networks.

The other part is that since SSH only sends the changes on the screen, it uses very little bandwidth.

Try Eternal Terminal it's very easy to setup and is just a simple layer over SSH.

My main gripe is latency. Typing out letters with bad connectivity is very error prone - a native text box is all I would need, but I will give it a try.

I use crush (https://github.com/charmbracelet/crush) with the following patches:

  curl https://tg.st/u/0001-fix-unblock-all-commands-in-bash-tool.patch | git am
  curl https://tg.st/u/0002-feat-add-light-theme-with-auto-detection-for-white-b.patch | git am
  curl https://tg.st/u/0003-feat-enable-yolo-mode-by-default.patch | git am
  curl https://tg.st/u/0004-fix-disable-mouse-grabbing-to-restore-native-termina.patch | git am
  curl https://tg.st/u/0005-feat-skip-project-init-prompt-and-quit-immediately-o.patch | git am
  curl https://tg.st/u/0006-feat-remove-scrambled-rune-animation-from-waiting-sp.patch | git am
  curl https://tg.st/u/0007-feat-remove-quit-banner-and-thank-you-message.patch | git am
  curl https://tg.st/u/0008-feat-show-output-in-full-instead-of-collapsing-trunc.patch | git am
  curl https://tg.st/u/0009-fix-discover-map-model-features-advertised-by-v1-mod.patch | git am
  curl https://tg.st/u/0010-feat-keep-large-and-small-model-selections-in-sync.patch | git am

I use pi.dev, it is pretty minimal, but you can extend it however you want since agent has access to it's own documentation.

I have a server living in my home office, always on. I have a tmux session on it with vanilla Codex and Claude Code CLI, I can via my Ubiquiti network stack wiregaurd in to this box anywhere on the globe with just my laptop. Works super well for me. I also have some cheap shelley power plugs that I can use to cycle my PC’s power state if needed.

This is basically my setup but I'm using tailscale and zellij. I don't have any contingency plan in place for my power or home internet going down though..

I used to do the same but being unable to copy/paste easily, follow links, upload files, etc led me to write a small bridge so I can turn Github issues into my harness frontend:

https://github.com/skorokithakis/symphony

Now I chat with OpenCode/Pi over Github issues, which is a miles better experience.

https://omp.sh/ has amazing defaults and it sips tokens. Works really well with DeepSeek V4.1 Flash.

Use the model through a fast and reliable provider such as Fireworks directly, skip OpenRouter.

Did you ever try pi by itself? For those new to the pi ecosystem - any rationale to go with pi vs omp?

It's like choosing between vim and helix. I started my career with vim in the early 2000's, customized the whole thing and had my config in a version control.

Then I installed helix and I just use it without config.

If you like configuring things take pi, if not omp is pretty much great defaults.

You must not be a software engineer shipping features that span multiple systems constantly as fast as possible in order to differentiate your company from competitors.

90% of software engineers are not, so they don't need a premium model.

The speed of deepseek is insane to experience after using claude code with opus for so long. Not only is the tps roughly 3x faster, but the round trip times are magnitudes faster.

$200/month is for Navier-Stoker grade problem.

Where do you do your inference?

> which shadowy government

A bit of false equivalency there

Abundance is a value all its own

Are you just using it directly from them?

yeah

Kind of ambiguous without saying token amounts and cost.

API pricing?

Deepseek isn't as good as Kimi, let alone the frontier models.

It's a very noticeable hit. It's like programming with mid-2025-era models: ignored instructions, dead code, mistakes.

It works...but anyone going from Sol to Deepseek is going to have a rough transition.

Opus 5.5.is crazy good. I ask it to do things and it just writes the code to do it.

I just ran a huge text/image extraction grudgematch against all the current inexpensive models except gpt-5.5/5.6/6 (due to some issues with openrouter and bugs in my code) and DS4 ranked very poorly. Accuracy winner was Gemini 3.8 flash with minimax M3 and qwen 3.8 placing, and the chinese models beat the incumbent (Gemini 2.5 Flash) on cost whilst keeping like 95% of the accuracy.

I haven't used deepseek for anything else but the above results make me question its overall capability. Meanwhile qwen3.8 has continued to impress.

It's trivial to hit that kind of quota if you're trying to execute on major projects. Especially as you start having dozens or hundreds of subagents investigating, prototyping, and working on different things.

You clearly haven’t used opus or astra or only had simple tasks. Such a difference maker

Reply on news.ycombinator.com