Anecdotally it seems like the majority of continued uplift comes from models doing more and exploring more of the search space, and we are getting better at giving them embarrassingly parallel problems to Hoover up with infinite token spend.
1. Intelligence per dollar spent/token output seems to be increasing rapidly. Opus 4 was $15/75 per million input/output tokens when released. Opus 5.5, which is significantly better, is $4/15. Sure it cost more money upfront to train Opus 5.5, or the bigger model (Fable 5.5) which tuned it, but once it is reached, there are massive diffusive effects.
2. The capital equation is still favoring more spend. Anthropic's revenue is growing rapidly and they are profitable on inference. Huge upfront spend for realized gains later on is standard for R&D. As long as serving the models is profitable and generates more value than the energy/compute consumed
3. The search space is too large to assume that improvement in AI assisted research is just models dumbly trying every combination until something works. For every additional bit searched, the space doubles, so even a trivially small combinatorial exploration incurs prohibitively large costs. Intelligent searching which continuously prunes that space along dimensions of effectiveness is not a trivial skill and is among the emergent accoutrements of these smarter models.
But they are essentially fighting for smaller and smaller part of the pie. The better open models are the less you need frontier models, so the frontier AI companies are essentially fighting for market that will shrink the better AI is.
Sure there is inertia and the usual corporate enterprise nonsense of "we only buy stuff from other big companies regardless the cost" that keeps the licenses going but we're quickly getting into point where you don't need any of the frontier models to do majority of the tasks
I mainly think this because frontier labs already try to compete at this tier. For example, 6.1 Sol has identical pricing to Kimi K3.1, but with a lower price per task due to Kimi's verbosity. It seems to score roughly the same on most benchmarks, with some (like Artificial Analysis) putting 6.1 Sol significantly higher. If you need to step down, 6.1 Terra is also cheaper than K3.1 or K3 and performs well relative to their benchmarks. Opus 5.5 and much of Anthropic's recent push seems to confirm this as a major area of focus as well.
I still use a lot of open models and harnesses, but I don't know if cost competitiveness (at least not with the current crop of open models) will ultimately drive a huge chunk of average users away.
Things could change, but fundamentally, large labs aren't precluded from training more cost effective models if there is significant demand for them, which is what we're seeing now.
6.1 Sol is not a “frontier model” and does not command the pricing of one. The reason it exists is because of competition from non-frontier models that are “good enough”.
In fact, you kind of made the same argument. As “good enough” models commoditized, the use cases requiring frontier models and frontier model pricing will shrink.
And, even if performance is not as good, many will prefer open models.
"The better open models are the less you need frontier models, so the frontier AI companies are essentially fighting for market that will shrink the better AI is."
The implication in this is that frontier labs' market is exclusively people who need the absolute frontier, i.e. Astra or Fable. And the point I made was that frontier labs are clearly prioritizing the broader market of people who want more cost effective models, and that if OpenAI/Anthropic offers models that are price competitive with the top open models (before factoring in subscription discounts and corporate lock-in), then I don't see how that leads to the average user walking away from the labs in favor of self-hosting or relying OpenRouter/Together/etc.
I also went out of my way to clarify that I personally prefer and use open models wherever possible. I'm not saying that there isn't a real group of people who prefer them. I'm specifically saying that frontier labs also sell "good enough" models and that there's nothing preventing them from training models to fit the demands of users. It's not as if training a frontier model boxes them in such that they can't release an Opus 5.5 or a 6.1 Terra.
The premise is this: frontier models seem to advance, and OSS models also seem to advance in terms of cost efficiency etc. It's like a race with say a 6 month (WLOG) gap between frontier and OSS models.
For frontier models to win perpetually against OSS, they will need to have a perpetual 6-month lead, or develop an edge OSS can't replicate for some architectural reasons (say MacOS vs Linux). If say a frontier lab trips because of a bad quarter or some approach temporarily maxxing out, they may not be able to maintain an edge vs OSS any more. It is a structural argument, not an empirical description of the advantages today.
I'm just talking about the idea that frontier labs will lose market share to providers that sell open model inference at sub-Fable prices. And I'm not convinced that this specifically is a threat to frontier labs, because they also train and sell models that fit that demand.
True in theory, but in practice hardware availability and cost will keep open source impractical for years to come.
I would argue this is more due to market economics than actual inherent cost. Opus 4 was when the open models started catching up. The R1 bootstrapping chain of thought paper via RL paper was released a couple of months before Opus came out and the open models started to significantly catch up after that. The duopoly was effectively broken and they lost their price setting power.
I would also like to point out that “profitable on inference” is similarly meaningless, because it’s based on a figure that arbitrarily discards whatever bits of their operating costs they need to to get to the claim that they’re profitable on inference. Potentially including ones that can’t be ignored because they’re every bit as much a part of COGS as the actual inference is.
It's going to depend on the task (and this is prone to cheating), but you can get a decent idea by looking at task benchmarks and the cost of completing them.
I was inclined to agree but Anthropic's latest S-1 statement only includes a paper loss due to their value going up. They're objectively making money now, the only question is can they keep it up and can they grow enough to be able to pay the bills they've signed up for in future?
Which is a really important distinction. One measures specific points on the jagged boundary. The other is a word that blithely paints it all with a single brush, and sweeps the jagged boundary problem under the rug in the process.
It really is important to keep track of that. Whenever we talk about “intelligence” as if it were a well-defined (let alone quantifiable) thing, we fall into the kind of wooly thinking that lets anyone say pretty much anything. We abandon the realm of grounded statements in favor of pure vibes.
> Anthropic has made roughly two years of historical progress over the last year [...] We are confident that the acceleration was not due to an acceleration in the scaling of training compute.
Note that Anthropic has essentially heavily censored this report, and the authors themselves say "the public has to rely on METR’s judgment much more than would be ideal".
What information are you trying to either ask for or share with us?
The original article is pretty clear, agent swarms and token spend are AI automating AI, and the RSP attribution condition specifically excludes your capital thrown at training concern?
We're getting the Claude efficiency talks at work. We're looking at all sort of optimizations, including local models and graphify, to reduce spend. There are people dedicating their nights and weekends to dropping token consumption and cache rebuilding. A few of these will likely catch on and be brought into common use. That will relieve at least a bit of the pressure, and enable some of the scaling to continue.
I'm not sure I follow. Are you saying a hardware shortage is going to restrict AI development? Because it seems to me like it would just push us towards an even greater focus on efficiency, which is still pushing the Pareto Frontier.
Not at all. Have you tried local AI lately? Qwen 3.8 is legitimately much better than SoTA from 1 year ago (Opus 4.1). You can run it on a consumer GPU.
Qwen Next Flash runs on a DGX spark well and is as good as some things you need datacenter class GPU's to run even today.
Check out Artificial Analysis if you haven't. Sometimes you don't realize how fast these things are moving, but to compare things from a year ago to today is just wild.
They aren't just getting better, they're getting more efficient by many orders of magnitude.
But, then as now, what they didn’t have was profits.
But that didn’t stop the AI winter from coming. Because a technology being transformative does not protect it from over-investment. History perhaps suggests the opposite is true.
Nor did it prevent the AI companies of the 80s from going bankrupt when they got undercut by the commoditization that naturally came with the maturation of the technology.
Given how "cost per capability " is still exponentially decreasing and only needs to stay on trend for a few months longer before nobody needs to pay anyone more than a few cents a month for something close to general intelligence - I am with you there.
I just hope it isn't the USA's Darien Scheme. If the entire world basically gets what is now frontier AI capability for very little cost, at the US private equity and taxpayer's expense, and given the state of the US current account right now, then that is pretty much it for the Pax Americana.
We're coming up on four years since ChatGPT launched, and the only thing LLMs seem to excel at is the movement of imaginary little green pieces of paper.
the same thing you are:
> if I may ask?
asking.