$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.

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.

Real work. $200 looks good on the outside until the essence of it, e.g. the models lie. I gave a list of spec to Sol and Sol decided some items didn't need to be done and the reason was "unproven", "not enough evidence", etc.

They all come up with amazing ways to lie (or be lazy). Often times what you get isn't what you asked for (only on the surface). E.g. I ran it to iteratively bench and optimize a better data structure for the project. It spent hours and finally came up with something. When I check it out -- it benchmarked the wrong criteria and was way off. So here we go again. Most AI work looks good on the surface. There are infinite edge cases.

So to do real work and gate it you need to:

1. Plan

2. Get it to do the work

3. Get independent agents to check from different angles

4. Take that feedback and get it to fix those gaps

5. Match against the plan and redo parts if needed

Every task is easily 4-5x the estimated amount of tokens.

p.s. well I did burn some banked resets building a compiler for some language AND it is still NOT done. Every time it says done I say check it says ok we still have bugs...

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.

> Not everything has to be done by the most expensive one.

Ok then.

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

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

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I'm also creating a free platform that replaces extremely out-of-date software, some of it only available with mutli-million dollar contracts, to help medical physics professionals with cutting-edge radiotherapy devices used to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

And out of curiosity, how do you automate testing the porting in the browser that's actually playable etc? And aren't you a bit scared of hosting and serving the "hairy bits" such as full assets? Nice job anyway!

I automate testing playable parts in browser by adding a dev mode that allows text commands for everything instead of having to rely on clicking UI or 3D elements. It can see the full game state in JSON and interact in any way via commands.

And re: IP - I just accept that I might get a C&D any day and have to take it all down. I'm careful to not accept a single penny for any reason and don't even have Patreon. Usually monetizing is what makes IP owners unhappy. And for the Pokemon MMO I just don't advertise it anywhere meaningful since they'll C&D the second they see it. Largely made it for my nephew and we play it together.

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.

> 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)

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.

Per the open code zen pricing page[1], it appears that the token prices are the same, but their cache is 10x more expensive?

[1]: https://opencode.ai/docs/zen/#pricing

Deepseek 0.14/0.28/0.0028

Pro Opencode 1.74/3.48/0.145

Deepseek 0.44/0.87/0.0036

For flash the input/output is the same, but the cache difference is big, you're paying 10x on >95% of your tokens.

For Pro, it's even worse, input/output is 4x and cache is 40x. The price different is really brutal. Yes you will still come out ahead by spending your first 10$/month on opencode go, but you will be saving a lot less than initially appears from their (60 USD for 10 USD pitch).

[0]: Deepseek: https://api-docs.deepseek.com/quick_start/pricing/ [1]: Opencode: https://opencode.ai/docs/zen/#pricing

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.

With unsloth's Q3_S quant + kyuz0 'llama-vulkan-radv-performance' toolbox, I am getting 280+tps (batch and ubatch at 2048) for PP and 18+tps for TG. I really only need 256k context so it all fits.

If I go down to the Q3_XXS quant + dpsark + 'llama-vulkan-radv-performance', I can get about the same PP and 25+tps for TG with draft set to 2 or 3. Fits about the same as above.

Edit: I did notice the 25+tps quickly degrades down to 20+ after the first few hundred tokens.

oh, they're mad.

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.

  - 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

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).

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.

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.

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

[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."

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

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.

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.

Terrible use-case.

harness: omp.sh

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...

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.

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.

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.

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.

Surely they are using read-only access.

This is such a ridiculous objection for how beloved it is. Wide swathes of the public can't cope with any adversity or risk.

With an agent (especially incompetently employed), the danger of unwittingly destroying your company (or at least, the crucial data/reputation) is rather higher. We are notoriously bad at estimating the downside risks in complex systems.

The most obvious case is the downside risks in complex financial constructs... things look great for a while ... until a sudden surprising collapse arrives and totally destroys all the upside you think you have created.

In any case, GP's post was not such a balanced consideration; it was just parroting a beloved risk-aversion meme that can easily be deployed against building anything (what if the building falls on top of someone?) or even leaving home to go to work ("travelling in a hunk of steel at lethal speeds – let me assure you that absolutely nothing can go wrong here, mate.")

What I find tiresome about that meme is the presumption that "something can go wrong" is useful input on its own. It's not. Mistakes are made all the time, the only way to avoid that is to stop breathing. Even in the process of me standing up and going to the loo, something can go wrong.

If the guy wants to make a case that it's too dangerous for the expected benefits, he has to actually make that case. Saying "risk exists" with no elaboration is a waste of HTML. "something can go wrong" every time he swallows food, yet mysteriously he still does it.

(the suicide analogies may seem mean-spirited, but I kind of mean it. If you consider every action primarily from a standpoint of "what harm or irreversible change can result from this", the only permissible path is to do nothing. To be moral is to be as close as possible to a rock or another inanimate object.)

Are you saying we shouldn't care about the future of affordability and access because at this moment we have seemingly endless access?

Sounds extremely short sighted.

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.

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.

It's that good. They are far from capacity limited, and even if they were, you can rent a single MI300X from somewhere like Hot Aisle and get more tk/s than you'll be able to use.

That one is dirt cheap at API pricing, I can't imagine quota is going to be a concern on the $200 subscription, which in my opinion easily supports full time use of 5.6 Sol on xhigh.

A couple of very talented friends were uttering curses upon the entire bloodline of whoever convinced them to try letting sol xhigh do serious work. Deepseek cleaned it up for a fraction of $20.

The cost per task was $0.03 with DeepSeek, $0.05 with Luna. $1.23 for Sol.

Tokens per second 132, 202, 70 respectively.

They also previously said prices will go down significantly once they get a hold of the upcoming Huawei chips (later this year).

Prices are going up just because they can. It can easily come back down. They aren't strained by some IPO / VCs requiring them to 1000x their earnings.

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

Think about it this way.

Let’s say you could buy an LLM that gets things right 98% of the time. But there’s another LLM that’s 100x the price but gets things right 99.9% of the time. To the lay person this sounds trivial but to a serious business this intelligence gap could represent millions, or billions of dollars.

But that’s simply not the case. It’s very clear that vast majority of the business do not generate additional value from incremental intelligence gain from these models.

There is a reason why Chinese open weight models are now popular even in American enterprises, because CTOs realize that they are indeed good enough.

The Chinese models are not good enough for anything other than pair programming, which is just a very last-gen way of using agents.

And when the big US models get better we will move with them. Until we stop seeing returns there is no "good enough", I don't know why this is so hard for HN to understand.

And even it isn't "enough". I can very clearly see myself using more advanced agents to move up the abstraction ladder.

For businesses that have actual problems to solve, I see them investing in the frontier for a good bit longer, probably until we have AGI that can replace employees, maybe even a bit after.

This is why I find the "good enough" arguments silly. Like, the usefulness of an AI tops out to you when you can pair program with it? Seriously? You cannot envision ways in which more advanced AI enables you to do more, better? That's bizarre to me. I don't ever see myself running out of problems to solve.

Is this what you are looking forward to?

This matches my experience with DeepSeek V4 Pro at Max reasoning, the preview version of the model kept regularly messing things up. About 30-60% of additional time to fix the output was needed.

On similar tasks, GLM 5.2 at Max reasoning screwed up maybe 20-30% of the time, while it still definitely made noticeable mistakes, they were far fewer in total and less egregious.

Kimi K3 at Max reasoning drops that value to below 10%, it's about as good as Opus or approaches Fable in some tasks. At High reasoning it also seems to be pretty close to Opus 4.8, not sure about the latest Opus model yet, but it's up there.

Only problem is that K3 is nowhere near as cheap as DeepSeek models, despite me personally liking the writing tone more (less Anthropic slop) and finding that it doesn't block my cybersecurity prompts, recently reproduced SQLi with a proof of context so I could justify fixing it.

My overall thoughts (released over some time):

https://blog.kronis.dev/blog/ai-slop-is-a-self-inflicted-tra...

https://blog.kronis.dev/blog/kimi-k3-is-out-is-anthropic-don...

https://blog.kronis.dev/blog/z-ai-s-glm-5-2-is-a-great-model...

I'd say as Chinese models get better, whatever moat Anthropic and OpenAI have dissipates. Currently the main things keeping me with Anthropic are their performance (tokens/second) and the fact that their visualization abilities within the app are pretty good.

That was ages ago (in LLM release timelines). DeepSeek V4 Flash beats it now and a lot cheaper.

> On similar tasks, GLM 5.2 at Max reasoning screwed up maybe 20-30% of the time,

GLM 5.3 bridges this gap.

> I'd say as Chinese models get better, whatever moat Anthropic and OpenAI have dissipates.

Their moat, especially OpenAI is funding and hardware resources. They gain train models 10x as large and also serve at large scale. That's it.

I’m sure the next models will only get better, when they’re released. Also super curious about what Moonshot will achieve and the full DeepSeek V4 Pro release!

> Their moat, especially OpenAI is funding and hardware resources. They gain train models 10x as large and also serve at large scale. That's it.

I’ve seen how much slower Kimi K3 can be and that part seems correct, their own GPU production still has ways to go and export restrictions definitely limit what they can do.

Not sure about the size part, if Kimi K3 achieves SOTA performance at 2.8T parameters, western models being >2x that size would be insanely bad in regards to efficiency. I bet they’re all within the same order of magnitude and below 10T and won’t really have a reason to go even that high for the foreseeable future.

As investors will start squeezing them for profitability, I suspect focusing more on efficiency will be commonplace.

They're a lot larger e.g. Fable. It is insanely bad. Do you know how much more resources "Western" companies have? Most in China don't have random GPUs to "play with" like every "frontier lab" employee does.

> As investors will start squeezing them for profitability, I suspect focusing more on efficiency will be commonplace.

They're born lucky though. Efficiency is "free". The next generation hardware e.g. Nvidia claims Blackwell -> Rubin is 10x efficiency (verified by Neoclouds apparently).

I do have a way of correcting through redundancy, though. If you are just vibe coding, you need to use the most capable model you can find and even then it might not be good enough.

I mean, this proves my point. Better models enable you to get more done. With Fable, 80% of the time, I no longer have chat with an agent over the details of a PR. I give it an outcome and it gets done. This means I can work on much more with the limited time I have.

And I don't see this ending. When better models come out that take that from 80% to 99.x%, I will have that better model manage teams of other models and move up the abstraction layer.

If models get even better than that, perhaps I stop reviewing PRs entirely. Maybe normies can start using agents to build real things.

Unless your business doesn't have many problems to solve and isn't in a competitive environment, it will benefit from using the best models.

My point is you don’t need the best model if you just put in QA processes that can be done by models also. And if you don’t have that, the model is probably not going to be good enough.

Fable has only been out for a month but somehow everyone is supposed to have moved to a completely different way of working that supposedly only works for Fable and nothing else…

This kind of takes makes me cringe. Why don't you go back to LinkedIn?

I have no idea what you mean by “pair program with an agent”, but Opus have been able of autonomous coding since last November, and with any half-decent harness even local Qwen3.5 was able to do so 6 months ago.

Fable is a stronger model, which means it can solve harder tasks but it's also over-hyped, because only a small fraction of task is hard enough to be Fable-worthy.

Fable is the only one that reliably one shots complex changes and makes the right design choices. Everything else requires handholding.

I can let Fable loose on a 12+ hour (for AI) task and it will have performed it flawlessly when I come back the next day. K3 and Opus are not like this.

And no, our harness is not the limiting factor here.

This is not appropriate for HN. Please review the guidelines: https://news.ycombinator.com/newsguidelines.html

There are businesses other than FAANG. I don't know why this is so hard for FAANG employees to understand.

It's been 2 months since Fable was released to the general, man. Nobody knows what's going on inside of these companies except the people at the coal face.

It's funny to see that Anthopic shills have been saying the exact same thing for the past two years now (and it was OpenAI fans before). It's amazing to see that Claude 3 Sonnet was "great" but now that even Qwen 9B is better than this version of Sonnet DeepSeek V4 is still not good enough despite being stronger than Opus 4.7 was.

> Let’s say you could buy an LLM that gets things right 98% of the time. But there’s another LLM that’s 100x the price but gets things right 99.9% of the time

If you think Fable makes 20 times fewer mistakes than DS4 you're delusional. It doesn't even do 20 fewer mistake than Gemma 4…

Same reason it makes sense to assign a team of humans that cost $100k/mo to a product that brings in $5M/mo, rather than one human with 5 Claude Max subs.

The cost is a rounding error.

This is why I quite like Kimi K3 - close to the same performance (definitely like Opus, approaching Fable), noticeably cheaper, generally good enough for me to daily drive. Only problem is that their official provider (on the Vivace plan) feels kinda slow, I'd say close to 2x slower than Opus on Max reasoning on average (probably more relatable than Fable).

You can say this about literally every product we buy. And yet...

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.

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?

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.

At this point, I don't even know if its possible

- You describe the breakdown in terms of time but it's more accurately a function of reasoning complexity.

- You seem to assume that no intermediate evaluation is possible.

- Often it is (e.g. the build breaks or tests start failing), allowing for course correction. There's definitely a cost to that but it can still be cost effective if the accuracy is "good enough" and the price difference significant.

- There are numerous tasks that don't require Fable or GPT5.6 level reasoning to improve efficiency by an order of magnitude.

Such as Decision Making. /s

Many devs who have never tried from either side build it all up in their head but it's almost always been a matter of thorough tedium, which LLMs are excellent at churning through, especially when there's api docs/headers/code comments.

If you have the space, try mirroring your port at the switch level and capturing every packet then making it go through them all to look for whatever. We have NSA at home lol

Like Europe?

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.

Spirit was broken by oil prices because they target the low end customer with their ticket prices. Oil prices went up and they had no pricing headroom to charge more on tickets so they simply went kaput. Ryanair also suffered a fair bit. Other airlines did (comparatively) fine because they had the ability to increase prices since their customers are less price sensitive.

It’s a classical business lesson that being a “cost-sensitive” vs a “value-sensitive” business (what this tradeoff is called) is a tradeoff. It’s notably recommended that startups don’t target the lower end in prices since you can’t compete on cost with a business that has more economy of scale than you; you have to compete on features. And being a cost-sensitive business means that you are affected much more than other businesses by changes in material/component prices, because a 13-cent increase in the cost of a component matters more the more product you sell, and if you increase prices too much customers will start to wonder if the “budget” brand is really a good value proposition over the mid-end or high-end ones anymore.

Literally everything you're saying doesn't line up with reality. Apple's most popular product today are two lower end (neo and mini).

You make the mistake of thinking your theory defines reality, when in fact reality says quite the opposite here.

I'm saying this because your logic is the classic logic that everyting thinks makes sense until you do it. Everyone can't target the high end, there are limited customers with many choices, making it far harder actually to win.

You target the demand/pain/need regardless of market.

https://en.wikipedia.org/wiki/Category:Defunct_low-cost_airl...

https://en.wikipedia.org/wiki/List_of_defunct_airlines_of_th...

https://en.wikipedia.org/wiki/List_of_defunct_airlines_of_th...

https://en.wikipedia.org/wiki/List_of_defunct_airlines_of_th...

Seems more like a overall industry problem, not limited to low cost carriers

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.

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?

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?

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

The question is not "can you win with the best model?", it is "can you not win without the best model?".

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

For Anthropic and OpenAI, presumably. And the rather large economic distortion field around them, that may or may not go very badly for all our retirement funds if those firms become insolvent...

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.

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...

this has already happened with manufacturing, so it isn't surprising that other industries follow.

The US premium in engineering and scientific endeavors have been lacking for the past 30-40 years, and if it werent for tech and silicon valley, the US would have nothing state of the art. Even on that front, the US is falling behind given how much effort in tech has been diverted into privacy invading, and advertising.

The US has been riding momentum, but eventually that momentum will stop. It will take half a century to get back up to speed, and by then, the US will have fallen behind so far that catching back up will seem impossible.

The telling evidence would be if china has the first moonbase before the US does. I think this is highly likely looking at today's US administration.

I joined a company that is an Anthropic shop and I am genuinely shocked.

Sonnet 5 is a little better on long horizon tasks and headless unsupervised agent workflows - but for in-IDE workflows, it's virtually unusable.

I am so used to flipping around my codebase at warp speed with DeepSeek flash. It's so fast and accurate I don't have time for parallel agents. It's a really rewarding workflow.

Moving to Sonnet, you ask is something simple like "split this into a seperate file" "implement this method" "this is my schema, implement a repository for it". It'll spend 30 minutes thinking and charge like $12. And no token caching, what are you even doing Anthropic?

It's unusable.

DeepSeek are in a league of their own

As it is now clear by behaviour of companies and US govt, all these investments will be backstopped by US govt. No US AI company will go hungry, they are national champions.

I dont see how the outlook is any better for the open weight companies. They’re in the exact same situation as the closed weight companies except they have had much less revenue, and built up less of a brand, leading up the the point where they are equal in terms of model quality.

I downgraded my Claude subscription and delegated my Claude Opus access to serve the role of an Architect to brainstorm and plan every step of development.

I leave the development to Deepseek.

Claude gets to review at many layers. It is often just as good as if I let Opus develop it by itself(the architect session will find similar number/level of gaps).

Deepseek flash as an architect and problem solver is not as thorough as Opus5 + high. Codex sol+ high is even better than Opus 5 at this moment for my needs.

First of all, no one knows the "true cost" of any of this, yet, but we know it's expensive. To what extent are the Chinese labs being subsidized? Are they real businesses?

Second, the Chinese labs aren't some "super geniuses", while the American labs are full of clowns. As of today, like the past 3 years, American labs are SOTA. That might change, but let's not act like the American labs don't know what they are doing.

The idea that people are going to use cheaper models for cheaper work isn't some novel revelation, it's completely obvious. People are doing it already, eschewing Fable.

The point is it takes money to keep developing models. Everyone is playing by the same rules. At this point, the US labs are trying to build businesses. I'm not really sure what the Chinese labs goals are. But I do know they aren't doing charity work.

And who gives a flying fuck. I am a "real" business and I count my money. It is not my life goal to prop some fat cats crying crocodile tears

I have been using it since it got released and its as good for scoped coding tasks, as the other big models I use, but just soo much cheaper.

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”?

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

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.)

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

I am not sure why you wouldn't want to use the SOTA models unless speed is a concern. Otherwise you are leaving quality on the table.

The 'floor' has gone up: today's model a bit behind SOTA is like model releases that were blowing people's minds a few months ago. Compared to, say, DS R1, this is far out stuff.

This, Luna, and (if it's good in practice) Laguna S are also fast and light not just cheap. And, as happened before, DeepSeek's first but other open model makers likely follow.

And a small, fast model taking small steps is...fun? More like working with code.