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.
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.
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?
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.
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.
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.
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.
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.
Right now the US dominates everyone else in actual chips in data centers. So even if deepseek etc tries to undercut, they’re very capacity limited.
This is made brutally obvious by anthropics customer support for people with such accounts.
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.
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.
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.
Another one I did was a printer data stream translator from an obscure format to PostScript/PDF (or just PNGs), complete with cups support, etc so these old apps can easily be hooked up.
Flash is capable now of running long range defined-goal tasks like this.
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).
So, death sentence even to frontier models?
Well... I would think that the whole AI industry in the US are working towards public bailouts... Which I guess they'll get under the current administration... So they'll be fine... Nobody there really seems interested in actually creating a profitable business anyway...
In what kind of sad and failed dystopia is this a "saving grace"? For whom?
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.
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
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?
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?
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.
At least here in Germany Aldi isn't even really limited to the poor, it's famously a place where you can run into anyone. Where I used to live in Berlin close to the government district I literally on occasion ran into the chancellor (and her bodyguards). Aspirational shopping where you buy premium goods to pretend to have higher social status honestly seems a bit on its way out. Even middle class people seem to consciously shop more utilitarian now.
Another time, Flash started trying to make tool calls by just calling bash and catting the tool call to stdout. Then it started running echo xx for every two letter UNIX command it could think of: mv, cp, etc and the it dug into uv, ty, and jj
With 5 active sessions going nonstop? That seems like a pretty important qualifier.
Then, once I go over, API pricing racks up FAST!
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
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.
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.
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.
$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.
> 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)
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.
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.
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.
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.
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.
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.
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...
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.
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.
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).
- 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)
Cache:
- Hit: 1,265,646,976
- Miss: 18,208,088
- Total input tokens: 1,283,855,064
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
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.
[0] https://openrouter.ai/deepseek/deepseek-v4-flash-0731#provid...
"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."
What's normal usage? I mean, Kimi is already really keen to spin of lots of subagents, and DeepSeep can probably do the same?
> 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.
Yes, though I did that even with Claude (on my employer's token budget). The agents are great at doing the gruntwork of chasing down the reproduction of flaky tests, too. They need some hand holding at first, but the guidelines are usually re-usable per project. (Claude specifically needs to be told to really concentrate on reproduction, and not eagerly start fixing the flake: if you don't have a reliable reproduction, you have no clue whether your fix actually fixes anything.)
which in your case is?
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.
oh, they're mad.
And, probably 99.99% of people using LLM probably don't even need SOTA anyway.
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
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.
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).
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”?
this seems like such a bad idea
"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.
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...
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.
Terrible use-case.
harness: omp.sh