I think an interesting question is going to be, if models are a commodity, who is going to want to foot the very expensive bill to train them? I'm sure training cost will drop.. eventually, but I doubt it will happen fast enough for any of these companies.
It would have been an interesting experiment to charge more for it right away and see what the market would bear, rather than tease it for long enough for it to be presumably superseded any time now by whatever is next.
Best marketing in a long time though. They fucking called it Mythos and built mythical claims about it. I mean… fucking hats off to the PR team.
And we can't ignore the power of "good enough". GLM5.2 may not be as good as the SOTA models, but it can be good enough for most, of not all, of our needs.
Are they really the best models? Like take anthropic. Without mythos, it's the what? Third best?
Sure openAI just leapfrogged them but .. seriously to get there it's a giant model that costs insane per token.
Nobody needs that, it's like NVIDIA or Intel claiming they have the best gaming performance, but to achieve that they are using more power per frame than anything else.
Everybody is just judging all of this by vibes anyway. Every week, a new model comes out and there's 500 comments simping for it within the first hour of its release. Both OpenAI and Anthropic have been practically indistinguishable to me.
Yes. I mean, most people agree they are. I've used all of the serious contenders (well not Grok 4.5, because Musk, and not Meta Spark because Zuck, but everything else I've used on at least a couple of projects to get a feel for them). My experience roughly matches the vibes. But, Fable is remarkable when it doesn't refuse to do the work (which it does, quite a lot, since my areas of interest are security and training specialist models).
Anyway, the vibes strongly indicate Fable is the best model, but not by an amount that is noticeable to most people. You could pick any of the top 10 models on this chart and do most of the tasks most people are doing with models:
Think airlines - both passenger and freight. They have never come close to capturing all the economic value they enable.
I think any security-related task triggers it to think about the threat model and thus hit the guardrails
Amazing story. If we make such leap in semiconductor field, it will be bigger than anything we have done till now. And all of that in 10years!
The thing is, it needs demand to drive it. Laptops have been roughly the same spec for the last 10 years because we don't need them to be bigger; there's no demand for a 16Tb RAM laptop because we don't have anything that could possible need that much RAM. Until LLMs came along, and we all want to run them locally, and so now there is a market for 16Tb laptops. So we'll invent the tech to make that happen.
And yes, laptop specs haven't changed much and this is partially because the need for spec changes wasn't present, but also during the last 20 years there has been tremendous pressure for efficiency in datacenters.
Despite that, dennard scaling is dead since 20 years. There are physical limits. Already now, the wear effect of electrons jumping is present, and it will only get worse as things scale towards smaller sizes.
There are some benefits to be had, e.g. one can etch models into chips directly so you can pack them more closely, and run more inference on Tensor like chips, but that gives you maybe one order of magnitude improvement in total, at most. Also, of course nobody does that when each 2-6 months a new model comes out.
I had a friend working in optical computing back in the late 80's that would wax lyrical about how optical computing was vastly superior to silicon back then. But it never took over because silicon worked well enough.
If we've hit the limits of silicon then there are other options. We would need to reinvent huge chunks of our tech stack, and that is incredibly expensive, but if the demand is there, we'll do it. The demand has never been there.
Examples of what exactly you're claiming is precedent for this would be helpful.
Not sure where the 1Gb number comes from? A standard laptop now is ~16Gb of RAM, so 1000x (and 1Gb -> 16Tb would be 16000x not 1600x). We went from Kb to Mb and then Mb to Gb of memory roughly every ten years from ~1990 -> ~2010. Each of those jumps is 1000x
Talking this over with claude, though, it pointed out that the need in dealing with LLMs is bandwidth and read-only storage, since the weights aren't dynamic. So we're not necessarily looking at 1Tb of RAM, we could be looking at 256Gb of faster RAM, and multi-TB of (much cheaper) flash storage, with extensive caching built in at OS level. This is all technically do-able with current tech, so it'll be interesting to see if it happens.
I personally don't think the weight efficiency will improve that much; if anything big does happen, I expect it to be about scalable architectures and continual learning
I hope everyone reads these LLM threads like your post, complete shots in the dark because nobody here will get close to predicting what the environment will look like, even the "insiders".
I don't know how to apply those lessons to AI, as long as AI requires so much hardware to operate. If small models actually get capable enough, the shape of the industry changes drastically.
There are little models that are very good for their size. I say nice things about Gemma 4 damned near every day. But, I'm not writing code with it. I am using it for finding security bugs, though, as the 31b variant is outrageously good at it for its size: https://swelljoe.com/post/gemma-4-exceeds-expectations/ and I'm also using it as a base for my own training experiments, specifically the 12b which is small enough to train a LoRA for on my local hardware so I don't have to rent cloud GPUs. The 12b QAT can run on your Pixel 10 Pro today and is frightfully smart for its size, and has great vision capabilities.
But, I keep saying "for its size". You have to be realistic about what tasks these self-hosted models can do. They are getting better though. Gemma 4 31b is competitive with models 10 times its size from a year ago. That's remarkable, and indicates where things are going.
Short term, we can compare this to 2-3 recent (mini) revolutions: internet, mobile, cloud. Then the answer is somewhat predictable and (somewhat sad personally). Companies owning the main distribution of intelligence (big labs) or distribution of the app/cloud layer (Google, MSFT, AWS) will make most of the money. In fact Google looks well positioned that way with owning intelligence, cloud (and even hardware, if they can get TPUs right as commercial product).
Long term view is interesting and somewhat satisfying (again, personally). We can compare this to industrial revolution, but for intelligence instead of physical labour. I hope, to borrow from Alan Kay's words, the total value generated will be more than what few big labs can capture. Though we will also see normal market dynamics of boom and bust in play. Companies building something useful, patiently will keep winning the markets. But only to get challenged by newer modes of the technology emerging.
In this long term view, the technology per se doesn't offer monopolistic profits to big labs. I think Anthropic is well aware of this and they are trying to extract as much cash from white collar work automation as they can before things are democratised. Contrary to popular opinion, they are also trying to seek a regulatory capture here by to maintain monopolistic position in the US market by scare mongering about China and open source. Its a case study how they managed to keep the good boy image of themselves while doing this.
In the end, I hope the technology emerges as electricity or combustion engine cars. Yes early pioneers (e.g. Ford) were perhaps able to make lot of money. But eventually, the technology was too important to allow one party to have monopoly and we had an abundance market which enabled jobs and money for a lot more people.
Edit, postscript : Dario, Sam and even Jensen will end up looking like the new the John D. Rockefeller's of this era. I'm personally hoping Demis Hassabis actually solves something much more important (problems in diseases, biology etc) with AI.
I would not put "cloud" at the same level as Internet and mobile. Cloud is just a layer on top of hardware that in the end makes almost no difference for Internet to exist and operates. Said differently: I doubt the world would be different nowadays without "the cloud". But it would definitely be different without Internet or mobile (smart)phones.
It's not comparable because OpenAI caps thinking to High in the ChatGPT "Chat" interface (and the "Work" thing where it actually does let us use Extra or Max is fucking shit).
GPT 5.6 Sol (High) is almost certainly worse than Opus 4.8 (Extra), and nowhere close to Fable (Extra).
I literally got a refund for my $20 OpenAI subscription after playing around with 5.6 Sol for a couple of hours (yes even on Codex) because it's so unusable and I'd rather just use Fable today and 4.8 Extra starting tomorrow, still within my $20 Anthropic plan. And I'm not even poor.
It is quite literally unusable to me.
Is different than
> quite literally unusable
That's not sign of commodity actor, just the opposite.
You didn't switch to 'Random Corner Store Token Seller' down the street, did you?
There 2-3 top players, that is not commodity.
Commodity is when there are enough that none of them have market power or can set prices.
'Commodity' means you buy your tokens from the Grocery Store on their loan plan. Like consumer credit is a commodity.
FYI predict this is roughly the way it will stay.
We will develop a lot of use for Chinese Open-ish models etc. but the SOTA's will maintain their place for a lot of things.
You're assuming that SOTA never hits a hard ceiling, letting local models catch up and achieve parity
It seems unlike that the frontier labs are going to be keeping ahead forever, they'll hit some kind of ceiling eventually
I think a few things are going to happen:
1) The Open Weights never really fully catch up, because there's too much Engineering and integration going now. It's way more than 'weights'
2) The Commodity Chinese models never quite catch up for the same reason every other product they make does not catch up - while they will shine in some areas, it won't land fully.
3) Horizontal integration, supply chains, availability, SLA, security, branding, regulatory requirements - all of this will add up to something competitively maintainable.
Can you name a product category that has truly hit a ceiling? Cars, computers, phones, airplanes ... always seems to be a way to nudge forward.
Good point though.
I think basically every product category has hit ceilings by now, honestly. Do you think vacuum cleaners are significantly better at vacuuming than 10 years ago?
Not really. But they pivoted to doing autonomous vacuums instead. The actual vacuum tech doesn't seem like it's getting much better though?
Same with a lot of appliances. Fridges aren't really better at keeping food cold than they were 30 years ago, are they? They just have "smart home" stuff now, and they are probably much more energy efficient
I guess you can look at that as "not reaching a ceiling" as an overall appliance but the actual discrete technology is not changing or improving much imo
Go and use a fridge from 40 years ago and compare to a modern one - granted, their essential function has not really improved that much. They were much more durable before, but rudimentary.
Most product categories in tech have evolved, and it's why there are leaders in most categories.
Fridges have made huge leaps in energy efficiency, they’re easily 3-4 times better at cooling your food.
Energy efficiency is nice, it's a good improvement, but from an end user perspective the 30 year old fridge still keeps your food just as cold.
If you were from 2025 and got trapped in the 1970s and needed to keep some milk from going sour for a day, you wouldn't be thinking "damn if only these old 1970s fridges worked more efficiently. I could easily accomplish this goal with a modern fridge!"
The 1970s fridge would do just fine
It's about ease of access, price, noise, convenience, durability, features.
My folks have this fancy 2 door thing, perfectly quiet, makes the best ice you can imagine, it's hidden into the cuppboards, it's energy efficient, has these crisper things, you can see in and reach around easy, lots of space. It's a better product.
It's all relative. In computing we're used to Moore's law driving most of the innovation (including this AI boom, which was at least partially due to availability of high powered GPUs), if improvements become less "exponential" and more "linear" it would feel like stagnation.
Right now in AI, we're talking about leaps of capabilities in months. When improvements come along every other decade, it's not "hitting a ceiling" but effectively it's plateauing and stagnating compared to this period of high growth.
For example, Chinese models are said to be roughly 6 months behind. If this remains constant and frontier AI models gets a break through every couple years, this isn't "hitting a ceiling" but it would erode away the competitive edge they have over the Chinese models.
"Commodity" doesn't mean there are no luxury goods in the space, it just means there are many options that will work for most people. I'll pay more for the best model, right up until the best model stops being a good deal. But, switching won't be all that painful. Even last week before Meta and xAI released competitive models, I was already using DeepSeek for tasks best served by an API and where the smartest model isn't critical. It's just so cheap, I can send it 10x more tasks for the same money. I haven't even mentioned several others that aren't competitive today, but likely will be.
I think predicting this market will stay like it is, with clear dominance by Anthropic and OpenAI seems like it requires ignoring a lot of countering evidence.
There are two Tier 1 platforms today.
Meta, Google and XAi are formidable Tier 1.5 place, any one of which could rise to the fore.
My belief is that it will be Google and that probably only one of them will keep up in the long run.
There is a 'breaking point' when you start to get past 4-ish players - it really does start to introduce competitive pressures.
Your second point about Tier 2 substitution is valid, but a few things:
1) Tier 1 models are not a 'luxury good' - that has a different economic definition. They are for most applications today actually just the quality, rational choice.
2) Substitution will have different effects for different people, and you're right that AI for many tasks will be commiditized.
All of the profits in Mobile Phones go to Apple even though they are not the biggest player.
Almost all of the profits in Silicon go to the leading edge chips - even though there are a zillion fabs that make legacy chips.
If I run an oil refinery, my fractional distillation system needs to be reworked depending on the exact mixture of crude I'm taking as input. So there are still switching costs even in the textbook example of a commodity.
Crucially though the exact upstream I use has minimal impact on the downstream. Closer equivalents, say, another barrel of WTI grade crude from a nearby regional supplier, require extremely minimal reworking. Oil from a different region, that might require more retooling, so I might be willing to sustain a longer shock in market conditions before making that switch. The important thing is the output broadly remains the same, but even this is broad, e.g. a different mix of inputs yields different ratios of output.
LLMs are quite similar no? Maybe switching to another SOTA model has minimal reworking, as you can delegate at the same level of abstraction to the model, whereas switching to a slightly-behind-frontier model you need to do more hand holding. Switching costs being nonzero does not preclude them from being broadly an interchangeable input in the production process. Any non-frontier use (99% of SWE) will be delivered on pretty much the exact same timeline irrespective of which model was used, so my requisition process looks more like buying barrels of oil than e.g. shopping for a new phone.
There are 3 SOTA model makers, and they have pricing power.
The 'switching costs' is not the issue so much as the inherent control over the commodity.
Think OPEC - when they acted as a cohort - they raised prices dramatically by having enough control to 'set prices'.
When OPEC lost it's pricing power ... nobody could set prices.
Fable is considerably better than GLM5 and it will have a strategic input - there is just hardly any substitute for it.
If these were cars - we'd just use whatever fuel.
But these are 'F1 races' - if you have some low grade 'dirty fuel' you will lose the race. You must have the 'top fuel'. There are 3 provides who implicitly collude and set prices.
Has nothing to do with market power. Market power can persist irrespective of the existence of a commodity. Market power only goes away in a perfectly competitive market.
Does such a thing exist? Hahaha. No
By the same measure, NVDA is Cisco, providing the backbone and capturing a ton of the early benefits, but soon becomes furniture while the excitement moves further up the chain.
I think we can start to see the outlines of this happening with LLMs as well. Local models have gone from being proofs of concept to something that is at least remotely comparable to contemporary models, and the gap there is closing faster than SOTA models are pushing it forward. Local models still suffer from high hardware requirements, but so did early image gen models where typical consumer hardware was insufficient for efficiently running them.
Ok, I completely missed that one. Can you point me in the right direction?
[1] - https://github.com/lllyasviel/Fooocus
The US government thinks they can dictate who can access "Mythos-level" (whatever that is) LLMs. But what will happen when this can be run on consumer hardware?
I guess this will be yet another vector too attack open computing and the idea that people can a) own computers, and b) choose what software they run on their own computers
That's a little late now, they should have tried to figure that out at the beginning. But then, it wasn't their own money they were burning in the meantime.
That is a very astute and concise way to explain everything about how the frontier labs are behaving and how they're trying to push more people to pay token rates for the best models. At the current subscription prices ($100 or $200 a month for a generous, though bounded, amount of tokens), frontier models are a no-brainer, most folks and companies will use them. But, at token rates, 10x or 100x the cost of open models or what I was spending on the frontier models a month ago? That is a harder question to answer "yes" to. I certainly wouldn't spend $1000 a month for the best model, much less $10,000; my employer might pay $1000/month, but definitely not $10,000. The frontier labs need everyone to answer "yes" to spending 100x what they currently spend to justify the valuations, and it's just not going to happen as long as everyone knows how to make these models.
Both OpenAI and Anthropic are trying to figure that out now. Anthropic, in particular, has their finger on the trigger...they want to push people to usage-based billing for Fable. But, OpenAI released 5.6 Sol, competitive with Fable (or close enough), and it's available via subscription (even the $20 subscription!), and there's no moat keeping someone from switching. If Anthropic really does end Fable access on the subscription plans in a few days, I predict a large market move back toward OpenAI.
The market isn't going to bear the cost of making the frontiers investment make sense.