Diffusion model papers are always interesting to read but I always feel like they need some mechanism to insert or delete tokens.
In the example in the figure in this post, once it has fixed "British munchkin cats _ _ and ..." you _can't_ get to "British munchkin cats are a new and controversial breed." because there's not the right number of tokens between "cats" and "and".
In a coding context, if your model samples a paren or a comma or something which is entirely plausible at that position, it can still close off an expansion which would be syntactically correct.
OK, but then, in this regard, left to right generation is hardy better:
Once you get to "British cats <next-token-here>" you can't get to "British munchkin cats <next-token-here>"; the tokens to the left are done and dusted.
It's kind of a feature. Diffusion is used for images, right? It's like saying, once the image of a door has started to form right next to a kitchen counter, it cannot insert a refrigerator there any more. Well, maybe it doesn't "want to" because that layout is already settled by that time.
However, I believe this would "only" be able to insert tokens, not to delete tokens again it mistakenly produced before. (The deletion in the title refers to the reverse process during training, where tokens are progressively deleted rather than masked.)
But the "infilling" problem isn't exactly solved for AR LLMs, so it's a strange critique.
Further more, you're applying the logic of AR LLMs to diffusion models. AR LLMs are only seeking the probability of the next token (a chain of conditional probability), but diffusion LLMs are modeling the probability of the entire output at once. Because of this token structures that leads to invalid outputs should be extremely low probability if properly trained.
The cat example is from the section on their block-causal attention mask. I really don't think this fixes the issue. So far as I can see, the block schedule dictates when they sample at each position. It does _not_ change that they basically have an array-of-token-vars representation, and once `t_i` is sampled, nothing can "move" that value left or right.
Early draft yes. But when you write an early draft of prose or code, you leave yourself the ability to insert or remove material in a way that _changes the indexes of the tokens you already put in your draft_. If you write a letter, you may know that it ends with "Yours Truly, <your name>", but not know the absolute number of tokens the letter will use. In this framework, once you say that "Yours Truly, John Hancock" are tokens 501 to 506, infilling the preceding sentences requires that you exactly preserve the number of tokens before that point ... which to me seems silly.
I'm sure it's computationally messy to be able to slide stuff around, but if it meaningfully changes the topology of the search process, it may be worth it.
I think the gap is, if they're building hybrids with _forward_ AR and diffusion, they risk giving up the cool part of diffusion which is reasoning back.
I may be imposing unreasonable human biases on to this, but I really think it would be interesting to have the model engage with the structure of the text, rather than just being either a sequence or an array of tokens.
E.g. "I'm going to _ tomorrow." If the _ is not just a token but an expansion in context, which might be a noun phrase, a verb phrase etc, it could be filled in with "the mall", "practice guitar".
In code "if (_1) { return _2; }", _1 could be an expression whose type is bool, and which makes sense as a check to confirm that some process is finished. I don't care specifically how many tokens either of those is, but I do care that it makes sense in context.
Notice how all the major AI companies (at least the ones that don't do open releases) stopped telling us how many parameters their models have. Parameter count was used as a measure for how great the proprietary models were until GPT3, then it suddenly stopped.
And how inference prices have come down a lot, despite increasing pressure to make money. Opus 4.6 is $25/MTok, Opus 4.1 was $75/MTok, the same as Opus 4 and Opus 3. OpenAI's o1 was $60/MTok, o1 pro $600/MTok, gpt-5.2 is $14/MTok and 5.2-pro is $168/MTok.
Also note how GPT-4 was rumored to be in the 1.8T realm, and now Chinese models in the 1T realm can match or surpass it. And I doubt the Chinese have a monopoly on those efficiency improvements
I doubt frontier models have actually substantially grown in size in the last 1.5 years, and potentially have a lot fewer parameters than the frontier models of old
You're hitting on something really important that barely gets discussed. For instance, notice how opus 4.5's speed essentially doubled, bringing it right in line with the speed of sonnet 4.5? (sonnet 4.6 got a speed bump too, though closer to 25%).
It was the very first thing I noticed: it looks suspiciously like they just rebranded sonnet as opus and raised the price.
I don't know why more people aren't talking about this. Even on X, where the owner directly competes in this market, it's rarely brought up. I strongly suspect there is a sort of tacit collusion between competitors in
this space. They all share a strong motivation to kill any deep discussion of token economics, even about each other because transparency only arms the customers.
By keeping the underlying mechanics nebulous, they can all justify higher prices. Just look at the subscription tiers: every single major player has settled on the exact same pricing model, a $20 floor and a $200 cap, no exceptions.
These AI companies are all in the same boat. At current operating costs and profit margins they can't hope to pay back the investment, so they have to pull tricks like rebranding models and downgrading offerings silently.
There's no oversight of this industry. The consumer protection dept in the US was literally shut down by the administration, and even if they had not been, this technology is too opaque for anyone to really be able to tell if today they're giving you a lower model than what you paid for yesterday.
I'm convinced they're all doing everything they can in the background to cut costs and increase profits.
I can't prove that Gemini 3 is dumber than when it came out because of the non deterministic nature of this technology, but it sure feels like it.
opus 4.6 was going to be sonet 5 up until week of release. The price bump is even bigger than you realize because they don't let you run opus 4.6 at full speed unless you pay them an extra 10x for the new "fast mode"
If that's true, it would be surprising; the current Sonnet 4.6 is not in the same league as either Opus 4.5 or 4.6, either anecdotally or on benchmarks.
Because Opus 4.6 is better than 4.5. So if it's true that Sonnet 5 was so good they gave it the Opus name, does that mean there was an Opus upgrade that didn't pan out? And what is Sonnet 4.6? An upgraded Haiku? Just trying to follow the red yarn in the conspiracy board here.
I don't know whether there was an opus that ran into trouble or if they just looked at the model they had and decided that they could charge more than originally intended.
sonet 4.6 presumably is either a version of sonet 4.5 with optimizations for cost instead of perf (or a haiku that also got upscaled).
Anthropic is preparing for IPO this year, so it's not exactly a stretch to suggest that they might be trying to decrease their losses and increase inference margin.
It's quite plausible to me that the difference is inference configuration. This could be done through configurable depth, Moe experts, layers etc. Even beam decoding changes can make substantial performance changes.
Train one large model, then down configure it for different pricing tiers.
I dont think thats plausible because they also just launched a high-speed variant which presumably has the inference optimization and smaller batching and costs about 10x
also, if you have inference optimizations why not apply them to all models?
It kind of makes sense, at least a year or so ago, I know $20.00 unlimited plans were costing these companies ~$250.00 averaged out, they're still lighting money on fire with $200.00 but probably not nearly as bad, however, I'm not sure if costs have gone up with changes in models, seems like the agentic tooling is more expensive for them (hence why they're pushing anyone they can to pay per token).
Cite a source. Your concrete claim is that, on average, for every $1 of subscription revenue on a monthly subscription, OpenAI and Anthropic were losing $11.50?
It seems completely implausible.
I could believe that if a $20 sub used every possible token granted, it would cost $250. But certainly almost no one was completely milking their subscription. In the same way that no one is streaming netflix literally 24/7.
From what I've gathered, they've been mostly training limited. Better training methods and cleaner training data allows smaller models to rival or outperform larger models training with older methods and lower-quality training data.
For example, the Qwen3 technical report[1] says that the Qwen3 models are architecturally very similar to Qwen2.5, with the main change being a tweak in the attention layers to stabilize training. And if you compare table 1 in Qwen3 paper with table 1 in Qwen 2.5 technical report[2], the layer count, attention configuration and such is very similar. Yet Qwen3 was widely regarded as a significant upgrade to Qwen2.5.
However, for training, they doubled the pre-training token count, and tripled the number of languages. It's been shown that training on more languages can actually help LLMs generalize better. They used Qwen2.5 VL and Qwen 2.5 to generate additional training data by parsing a large number PDFs and turning them into high quality training tokens. They improved their annotation so they could more effectively provide diverse training tokens to the model, improving training efficiency.
They continued this trend with Qwen3.5, where even more and better training data[3] made their Qwen3.5-397B-A17B model match the 1T-parameter Qwen3-Max-Base.
That said there's also been a lot of work on model architecture[4], getting more speed and quality per parameter. In the case of Qwen3-Next architecture which 3.5 is based on, that means such things as hybrid attention for faster long-context operation, and sparse MoE and multi-token prediction for less compute per output token.
I used Qwen as an example here, from what I gather they're just an example of the general trend.
Similar trend in open text-to-image models: Flux.1 was 12B but now we have 6B models with much better quality. Qwen Image goes from 20B to 7B while merging the edit line and improving quality. Now that the cost of spot H200s at 140GB came down to A100 levels, you can finally try larger scale finetuning/distillation/rl with these models. Very promising direction for open tools and models if the trend continues.
> Parameter count was used as a measure for how great the proprietary models were until GPT3, then it suddenly stopped.
AFAICT that's mostly because what you're getting when you select a "model" from most of these cloud chat model providers today, isn't a specific concrete model, but rather is a model family, where your inference request is being routed to varying models within the family during the request. There's thus no one number of weights for "the model", since several entirely-independent models can be involved in generating each response.
And to be clear, I'm not just talking about how selecting e.g. "ChatGPT 5.2" sometimes gets you a thinking model and sometimes doesn't, etc.
I'm rather saying that, even when specifically requesting the strongest / most intelligent "thinking" models, there are architectural reasons that the workload could be (and probably is) routed to several component "sub-models", that handle inference during different parts of the high-level response "lifecycle"; with the inference framework detecting transition points in the response stream, and "handing off" the context + response stream from one of these "sub-models" to another.
(Why? Well, imagine how much "smarter" a model could be if it had a lot more of its layers available for deliberation, because it didn't have to spend so many layers on full-fat NLP parsing of input or full-fat NLP generation of output. Split a model into a pipeline of three sub-models, where the first one is trained to "just understand" — i.e. deliberate by rephrasing whatever you say to it into simpler terms; the second one is trained to "just think" — i.e. assuming pre-"understood" input and doing deep scratch work in some arbitrary grammar to eventually write out a plan for a response; and the third one is trained to "just speak" — i.e. attend almost purely to the response plan and whatever context-tokens that plan attends to, to NLP-generate styled prose, in a given language, with whatever constraints the prompt required. Each of these sub-models can be far smaller and hotter in VRAM than a naive monolithic thinking model. And these sub-models can make a fixed assumption about which phase they're operating in, rather than having to spend precious layers just to make that determination, over and over again, on every single token generation step.)
And, presuming they're doing this, the cloud provider can then choose to route each response lifecycle phase to a different weight-complexity-variant for that lifecycle phase's sub-model. (Probably using a very cheap initial classifier model before each phase: context => scalar nextPhaseComplexityDemand.) Why? Because even if you choose the highest-intelligence model from the selector, and you give it a prompt that really depends on that intelligence for a response... your response will only require a complex understanding-phase sub-model if your input prose contained the high-NLP-complexity tokens that would confuse a lesser understanding-phase sub-model; and your response will only require a complex responding-phase sub-model if the thinking-phase model's emitted response plan specifies complex NLP or prompt-instruction-following requirements that only a more-complex responding-phase sub-model knows how to manage.
Which is great, because it means that now even when using the "thinking" model, most people with most requests are only holding a reservation on a GPU holding a copy of the (probably still hundreds-of-billions-of-weights) high-complexity-variant thinking-phase sub-model weights, for the limited part of that response generation lifecycle where the thinking phase is actually occurring. During the "understanding" and "responding" phases, that reservation can be released for someone else to use! And for the vast majority of requests, the "thinking" phase is the shortest phase. So users end up sitting around waiting for the "understanding" and "responding" phases to complete before triggering another inference request. Which brings the per-user duty cycle of thinking-phase sub-model use way down.
I'd suggest that a measure like 'density[1]/parameter' as you put it will asymptotically rise to a hard theoretical limit (that probably isn't much higher than what we have already). So quite unlike Moore's Law.
Obviously, there’s a limit to how much you can squeeze into a single parameter. I guess the low-hanging fruit will be picked up soon, and scaling will continue with algorithmic improvements in training, like [1], to keep the training compute feasible.
I take "you can't have human-level intelligence without roughly the same number of parameters (hundreds of trillions)" as a null hypothesis: true until proven otherwise.
Why don't we need them? If I need to run a hundred small models to get a given level of quality, what's the difference to me between that and running one large model?
You can run smaller models on smaller compute hardware and split the compute. For large models you need to be able to fit the whole model in memory to get any decent throughput.
> There is not one person on the planet, who wouldn't prefer a doctor who is deeply considerate of the complexities and feedback-loops of the human body, over a doctor who is simply not smart enough to do so and, thus, can't. He can learn texts all he wants, but the memorization of text does not require deeper understanding.
But a smart person who hasn’t read all the texts won’t be a good doctor, either.
Chess players spend enormous amounts of time studying openings for a reason.
> Multiple small models, specifically trained for high reasoning/cognitive capabilities, given access to relevant texts
So, even assuming that one can train a model on reasoning/cognitive abilities, how does one pick the relevant texts for a desired outcome?
Bitter Lesson is about exploration and learning from experience. So RL (Sutton's own field) and meta learning. Specialized models are fine from Bitter Lesson standpoint if the specialization mixture is meta learned / searched / dynamically learned&routed.
The corollary to the bitter lesson is that in any market meaningful time scale a human crafted solution will outperform one which relies on compute and data. It's only on time scales over 5 years that your bespoke solution will be over taken. By which point you can hand craft a new system which uses the brute force model as part of it.
Repeat ad-nauseam.
I wish the people who quote the blog post actually read it.
It's the same thing. Quantize your parameters? "Bigger" model runs faster. MOE base model distillation? "Bigger" model runs as smaller model.
There is no gain for anyone anywhere by reducing parameter count overall if that's what you mean. That sounds more like you don't like transformer models than a real performance desire
Is anyone doing any form of diffusion language models that are actually practical to run today on the actual machine under my desk? There's loads of more "traditional" .gguf options (well, quants) that are practical even on shockingly weak hardware, and I've been seeing things that give me hope that diffusion is the next step forward, but so far it's all been early research prototypes.
Because diffusion models have a substantially different refining process, most current software isn't built to support it. So I've also been struggling to find a way to play with these models on my machine. I might see if I can cook something up myself before someone else does...
Based on my experience running diffusion image models I really hope this isn't going to take over anytime soon. Parallel decoding may be great if you have a nice parallel gpu or npu but is dog slow for cpus
Nothing to do with each other. This is a general optimization. Taalas' is an ASIC that runs a tiny 8B model on SRAM.
But I wonder how Taalas' product can scale. Making a custom chip for one single tiny model is different than running any model trillions in size for a billion users.
Roughly, 53B transistors for every 8B params. For a 2T param model, you'd need 13 trillion transistor assuming scale is linear. One chip uses 2.5 kW of power? That's 4x H100 GPUs. How does it draw so much power?
If you assume that the frontier model is 1.5 trillion models, you'd need an entire N5 wafer chip to run it. And then if you need to change something in the model, you can't since it's physically printed on the chip. So this is something you do if you know you're going to use this exact model without changing anything for years.
Very interesting tech for edge inference though. Robots and self driving can make use of these in the distant future if power draw comes down drastically. 2.4kW chip running inside a robot is not realistic. Maybe a 150w chip.
The 2.5kW figure is for a server running 10 HC1 chips:
> The first generation HC1 chip is implemented in the 6 nanometer N6 process from TSMC. ... Each HC1 chip has 53 billion transistors on the package, most of it very likely for ROM and SRAM memory. The HC1 card burns about 200 watts, says Bajic, and a two-socket X86 server with ten HC1 cards in it runs 2,500 watts.
> Our second model, still based on Taalas’ first-generation silicon platform (HC1), will be a mid-sized reasoning LLM. It is expected in our labs this spring and will be integrated into our inference service shortly thereafter.
> Following this, a frontier LLM will be fabricated using our second-generation silicon platform (HC2). HC2 offers considerably higher density and even faster execution. Deployment is planned for winter.
Personally I think anything around the level of Sonnet 4.5 is worth burning to silicon because agentic workflows work. There are plenty of places where spending $50,000 for that makes sense (I have no idea of the pricing though)
I think maybe there are subsets of problems where you can have either a human or a smart LLM write a verifier (e.g. a property-based test?) and a performance measurement and let the dumb models generate candidates iterate on candidates?
Yeah, maybe, but then it would make much more sense to run a big model than hope one of the small ones randomly stumbles upon the solution, just because the possibility space is so much larger than the number of dumb LLMs you can run.
I don't work this way, so this is all a hypothetical to me, but the possibility space is larger than _any_ model can handle; models are effectively applying a really complex prior over a giant combinatorial space. I think the idea behind a swarm of small models (probably with higher temperature?) on a well-defined problem is akin to e.g. multi-chain MCMC.
I see; the chatbot demo in the Taalas page. If they could produce this cost effectively it would definitely be valuable. The only challenge would be getting models to market before their next revision.
I'd love to know what's going on with the Gemini Diffusion model - they had a preview last May and it was crazy fast but I've not heard anything since then.
A lot of this post-training recipe feels reminiscent of DINO training (teacher/student, use of stop gradients). I wonder if the more recent leJEPA SigREG regularization research might be relevant here for simpler post-training.
Seeing half of an AR LLM's output tokens go to generating a predefined json schema bothers me so much. I would love to have an option to use diffusion for infilling.
I do wonder why diffusion models aren't used alongside constraint decoding for programming - surely it makes better sense then using an auto-regressive model.
Diffusion models need to infer the causality of language from within a symmetric architecture (information can flow forward or backward). AR forces information to flow in a single direction and is substantially easier to control as a result. The 2nd sentence in a paragraph of English text often cannot come before the first or the statement wouldn't make sense. Sometimes this is not an issue (and I think these are cases where parallel generation makes sense), but the edge cases are where all the money lives.
But I do wonder if diffusion models will be used in more complex Software Architecture for their long-term coherence, no exposure bias, and their symmetric architecture could work well with interaction nets.
Scaling laws mean that there's not much need to actually scale things to the skies. Instead, you can run a bunch of experiments at small scale, fit the scaling law parameters, then extrapolate. If the predicted outcome is disappointing (e.g. it's unlikely to beat the previous scaled-to-the-sky model), you can save the really expensive experiment for a more promising approach.
It would certainly be nice though if this kind of negative result was published more often instead of leaving people to guess why a seemingly useful innovation wasn't adopted in the end.
This doesn't mention the drawback of diffusion language models, the main reason why nobody is using them: they have significantly lower performance on benchmarks than autoregressive models at similar size.
Can't wait for the day I can actually try a diffusion model on my own machine (128GB M4 Max) rather than as a hosted service. So far I haven't seen a single piece of software that supports it.
If this means there’s a 2x-7x speed up available to a scaled diffusion model like Inception Mercury, that’ll be a game changer. It feels 10x faster already…
One appeal of it is for RL. If it ends up being a lot faster for generation, you'll be able to do a lot more RL.
If people can make RL scalable-- make it so that RL isn't just a final phase, but something which is as big as the supervised stuff, then diffusion models are going to have an advantage.
If not, I think autoregressive models will still be preferred. Diffusion models become fixed very fast, they can't actually refine their outputs, so we're not talking about some kind of refinement along the lines of: initial idea -> better idea -> something actually sound.
> If not, I think autoregressive models will still be preferred. Diffusion models become fixed very fast, they can't actually refine their outputs, so we're not talking about some kind of refinement along the lines of: initial idea -> better idea -> something actually sound.
I'm really curious about this, I'm but a simple client developer, so I don't actually grok some of the differences.
For lack of a better word, there's a "normie" position that "omg diffusion means it can edit!!111! big unlock!" -- I think that's cute but I also don't see it as intuitively correct. And I guess I don't even know why I don't see it that way. But regardless, it sounds like I'm correct there.
> If not, I think autoregressive models will still be preferred.
But here I get lost, at least so far, diffusion models seem strictly significantly faster, and on par with models with the same parameter count.
If that is the case, why would autoregressive models still be preferred?
Asking this also makes me realize I am treating "diffusion models are better" as a premise, if I'm asserting they're always faster and ~same quality...
Feels like the sodium ion battery vs lithium ion battery thing, where there are theoretical benefits of one but the other has such a head start on commercialization that it'll take a long time to catch up.
Not really. Unlike with physical goods like batteries, the hardware for training a diffusion vs an autoregressive language model is more or less exactly the same.
Although the lab that did this research (Chris Re and Tri Dao are involved) is run by the world's experts in squeezing CUDA and Nvidia hardware for every last drop of performance.
At the API level, the primary differences will be the addition of text infill capabilities for language generation. I also somewhat expect certain types of generation to be more cohesive (e.g. comedy or stories where you need to think of the punchline or ending first!)
Didn't thinking tokens resolve the most problematic part of autoregressive models (the first few tokens set the constraints the model can't overcome later) and give it a massive advantage compared to diffusion models by showing the thinking trace? I can see diffusion models being used as a draft model to quickly predict a bunch of tokens and let the autoregressive model decide to use them or throw them away quickly, speeding it up considerably while keeping thinking traces available.
The reason I mentioned "purely autoregressive" is that realistically I expect hybrid diffusion + autoregressive models to be the first popular diffusion models. I could be wrong though. And diffusion models have other tricks like really easy integration with simple classifiers.
Diffusion model papers are always interesting to read but I always feel like they need some mechanism to insert or delete tokens. In the example in the figure in this post, once it has fixed "British munchkin cats _ _ and ..." you _can't_ get to "British munchkin cats are a new and controversial breed." because there's not the right number of tokens between "cats" and "and". In a coding context, if your model samples a paren or a comma or something which is entirely plausible at that position, it can still close off an expansion which would be syntactically correct.
OK, but then, in this regard, left to right generation is hardy better:
Once you get to "British cats <next-token-here>" you can't get to "British munchkin cats <next-token-here>"; the tokens to the left are done and dusted.
It's kind of a feature. Diffusion is used for images, right? It's like saying, once the image of a door has started to form right next to a kitchen counter, it cannot insert a refrigerator there any more. Well, maybe it doesn't "want to" because that layout is already settled by that time.
There is a new way to train diffusion models to insert tokens between existing tokens rather than unmasking <mask> tokens: https://openreview.net/forum?id=VbvXjs5f72
However, I believe this would "only" be able to insert tokens, not to delete tokens again it mistakenly produced before. (The deletion in the title refers to the reverse process during training, where tokens are progressively deleted rather than masked.)
But the "infilling" problem isn't exactly solved for AR LLMs, so it's a strange critique.
Further more, you're applying the logic of AR LLMs to diffusion models. AR LLMs are only seeking the probability of the next token (a chain of conditional probability), but diffusion LLMs are modeling the probability of the entire output at once. Because of this token structures that leads to invalid outputs should be extremely low probability if properly trained.
This blogpost references block diffusion which fixes this issue that you are describing.
The cat example is from the section on their block-causal attention mask. I really don't think this fixes the issue. So far as I can see, the block schedule dictates when they sample at each position. It does _not_ change that they basically have an array-of-token-vars representation, and once `t_i` is sampled, nothing can "move" that value left or right.
I think that having an early draft of the output is part of the appeal of this type of models.
Early draft yes. But when you write an early draft of prose or code, you leave yourself the ability to insert or remove material in a way that _changes the indexes of the tokens you already put in your draft_. If you write a letter, you may know that it ends with "Yours Truly, <your name>", but not know the absolute number of tokens the letter will use. In this framework, once you say that "Yours Truly, John Hancock" are tokens 501 to 506, infilling the preceding sentences requires that you exactly preserve the number of tokens before that point ... which to me seems silly. I'm sure it's computationally messy to be able to slide stuff around, but if it meaningfully changes the topology of the search process, it may be worth it.
IIRC, some researchers are working on mixed AR+diffusion models for this sort of thing.
I think the gap is, if they're building hybrids with _forward_ AR and diffusion, they risk giving up the cool part of diffusion which is reasoning back. I may be imposing unreasonable human biases on to this, but I really think it would be interesting to have the model engage with the structure of the text, rather than just being either a sequence or an array of tokens. E.g. "I'm going to _ tomorrow." If the _ is not just a token but an expansion in context, which might be a noun phrase, a verb phrase etc, it could be filled in with "the mall", "practice guitar". In code "if (_1) { return _2; }", _1 could be an expression whose type is bool, and which makes sense as a check to confirm that some process is finished. I don't care specifically how many tokens either of those is, but I do care that it makes sense in context.
I was thining of something like LLaDa that uses a Transformer to predict forward masked tokens:
https://arxiv.org/abs/2502.09992
I wish there would be more of this research to speed things up rather than building ever larger models
Notice how all the major AI companies (at least the ones that don't do open releases) stopped telling us how many parameters their models have. Parameter count was used as a measure for how great the proprietary models were until GPT3, then it suddenly stopped.
And how inference prices have come down a lot, despite increasing pressure to make money. Opus 4.6 is $25/MTok, Opus 4.1 was $75/MTok, the same as Opus 4 and Opus 3. OpenAI's o1 was $60/MTok, o1 pro $600/MTok, gpt-5.2 is $14/MTok and 5.2-pro is $168/MTok.
Also note how GPT-4 was rumored to be in the 1.8T realm, and now Chinese models in the 1T realm can match or surpass it. And I doubt the Chinese have a monopoly on those efficiency improvements
I doubt frontier models have actually substantially grown in size in the last 1.5 years, and potentially have a lot fewer parameters than the frontier models of old
You're hitting on something really important that barely gets discussed. For instance, notice how opus 4.5's speed essentially doubled, bringing it right in line with the speed of sonnet 4.5? (sonnet 4.6 got a speed bump too, though closer to 25%).
It was the very first thing I noticed: it looks suspiciously like they just rebranded sonnet as opus and raised the price.
I don't know why more people aren't talking about this. Even on X, where the owner directly competes in this market, it's rarely brought up. I strongly suspect there is a sort of tacit collusion between competitors in this space. They all share a strong motivation to kill any deep discussion of token economics, even about each other because transparency only arms the customers. By keeping the underlying mechanics nebulous, they can all justify higher prices. Just look at the subscription tiers: every single major player has settled on the exact same pricing model, a $20 floor and a $200 cap, no exceptions.
These AI companies are all in the same boat. At current operating costs and profit margins they can't hope to pay back the investment, so they have to pull tricks like rebranding models and downgrading offerings silently. There's no oversight of this industry. The consumer protection dept in the US was literally shut down by the administration, and even if they had not been, this technology is too opaque for anyone to really be able to tell if today they're giving you a lower model than what you paid for yesterday.
I'm convinced they're all doing everything they can in the background to cut costs and increase profits.
I can't prove that Gemini 3 is dumber than when it came out because of the non deterministic nature of this technology, but it sure feels like it.
> a $20 floor and a $200 cap, no exceptions
Google caps at $250
opus 4.6 was going to be sonet 5 up until week of release. The price bump is even bigger than you realize because they don't let you run opus 4.6 at full speed unless you pay them an extra 10x for the new "fast mode"
If that's true, it would be surprising; the current Sonnet 4.6 is not in the same league as either Opus 4.5 or 4.6, either anecdotally or on benchmarks.
why is that surprising?
Because Opus 4.6 is better than 4.5. So if it's true that Sonnet 5 was so good they gave it the Opus name, does that mean there was an Opus upgrade that didn't pan out? And what is Sonnet 4.6? An upgraded Haiku? Just trying to follow the red yarn in the conspiracy board here.
I don't know whether there was an opus that ran into trouble or if they just looked at the model they had and decided that they could charge more than originally intended. sonet 4.6 presumably is either a version of sonet 4.5 with optimizations for cost instead of perf (or a haiku that also got upscaled). Anthropic is preparing for IPO this year, so it's not exactly a stretch to suggest that they might be trying to decrease their losses and increase inference margin.
It's quite plausible to me that the difference is inference configuration. This could be done through configurable depth, Moe experts, layers etc. Even beam decoding changes can make substantial performance changes.
Train one large model, then down configure it for different pricing tiers.
I dont think thats plausible because they also just launched a high-speed variant which presumably has the inference optimization and smaller batching and costs about 10x
also, if you have inference optimizations why not apply them to all models?
Isn’t Gemini $300 a month for their most expensive plan? That includes stuff like Genie and all that though.
Doubling speed can likely come from MoE optimizations such as reducing the amount of active parameters.
It kind of makes sense, at least a year or so ago, I know $20.00 unlimited plans were costing these companies ~$250.00 averaged out, they're still lighting money on fire with $200.00 but probably not nearly as bad, however, I'm not sure if costs have gone up with changes in models, seems like the agentic tooling is more expensive for them (hence why they're pushing anyone they can to pay per token).
Cite a source. Your concrete claim is that, on average, for every $1 of subscription revenue on a monthly subscription, OpenAI and Anthropic were losing $11.50?
It seems completely implausible.
I could believe that if a $20 sub used every possible token granted, it would cost $250. But certainly almost no one was completely milking their subscription. In the same way that no one is streaming netflix literally 24/7.
From what I've gathered, they've been mostly training limited. Better training methods and cleaner training data allows smaller models to rival or outperform larger models training with older methods and lower-quality training data.
For example, the Qwen3 technical report[1] says that the Qwen3 models are architecturally very similar to Qwen2.5, with the main change being a tweak in the attention layers to stabilize training. And if you compare table 1 in Qwen3 paper with table 1 in Qwen 2.5 technical report[2], the layer count, attention configuration and such is very similar. Yet Qwen3 was widely regarded as a significant upgrade to Qwen2.5.
However, for training, they doubled the pre-training token count, and tripled the number of languages. It's been shown that training on more languages can actually help LLMs generalize better. They used Qwen2.5 VL and Qwen 2.5 to generate additional training data by parsing a large number PDFs and turning them into high quality training tokens. They improved their annotation so they could more effectively provide diverse training tokens to the model, improving training efficiency.
They continued this trend with Qwen3.5, where even more and better training data[3] made their Qwen3.5-397B-A17B model match the 1T-parameter Qwen3-Max-Base.
That said there's also been a lot of work on model architecture[4], getting more speed and quality per parameter. In the case of Qwen3-Next architecture which 3.5 is based on, that means such things as hybrid attention for faster long-context operation, and sparse MoE and multi-token prediction for less compute per output token.
I used Qwen as an example here, from what I gather they're just an example of the general trend.
[1]: https://arxiv.org/abs/2505.09388
[2]: https://arxiv.org/abs/2412.15115
[3]: https://qwen.ai/blog?id=qwen3.5
[4]: https://qwen.ai/blog?id=4074cca80393150c248e508aa62983f9cb7d...
Similar trend in open text-to-image models: Flux.1 was 12B but now we have 6B models with much better quality. Qwen Image goes from 20B to 7B while merging the edit line and improving quality. Now that the cost of spot H200s at 140GB came down to A100 levels, you can finally try larger scale finetuning/distillation/rl with these models. Very promising direction for open tools and models if the trend continues.
> Parameter count was used as a measure for how great the proprietary models were until GPT3, then it suddenly stopped.
AFAICT that's mostly because what you're getting when you select a "model" from most of these cloud chat model providers today, isn't a specific concrete model, but rather is a model family, where your inference request is being routed to varying models within the family during the request. There's thus no one number of weights for "the model", since several entirely-independent models can be involved in generating each response.
And to be clear, I'm not just talking about how selecting e.g. "ChatGPT 5.2" sometimes gets you a thinking model and sometimes doesn't, etc.
I'm rather saying that, even when specifically requesting the strongest / most intelligent "thinking" models, there are architectural reasons that the workload could be (and probably is) routed to several component "sub-models", that handle inference during different parts of the high-level response "lifecycle"; with the inference framework detecting transition points in the response stream, and "handing off" the context + response stream from one of these "sub-models" to another.
(Why? Well, imagine how much "smarter" a model could be if it had a lot more of its layers available for deliberation, because it didn't have to spend so many layers on full-fat NLP parsing of input or full-fat NLP generation of output. Split a model into a pipeline of three sub-models, where the first one is trained to "just understand" — i.e. deliberate by rephrasing whatever you say to it into simpler terms; the second one is trained to "just think" — i.e. assuming pre-"understood" input and doing deep scratch work in some arbitrary grammar to eventually write out a plan for a response; and the third one is trained to "just speak" — i.e. attend almost purely to the response plan and whatever context-tokens that plan attends to, to NLP-generate styled prose, in a given language, with whatever constraints the prompt required. Each of these sub-models can be far smaller and hotter in VRAM than a naive monolithic thinking model. And these sub-models can make a fixed assumption about which phase they're operating in, rather than having to spend precious layers just to make that determination, over and over again, on every single token generation step.)
And, presuming they're doing this, the cloud provider can then choose to route each response lifecycle phase to a different weight-complexity-variant for that lifecycle phase's sub-model. (Probably using a very cheap initial classifier model before each phase: context => scalar nextPhaseComplexityDemand.) Why? Because even if you choose the highest-intelligence model from the selector, and you give it a prompt that really depends on that intelligence for a response... your response will only require a complex understanding-phase sub-model if your input prose contained the high-NLP-complexity tokens that would confuse a lesser understanding-phase sub-model; and your response will only require a complex responding-phase sub-model if the thinking-phase model's emitted response plan specifies complex NLP or prompt-instruction-following requirements that only a more-complex responding-phase sub-model knows how to manage.
Which is great, because it means that now even when using the "thinking" model, most people with most requests are only holding a reservation on a GPU holding a copy of the (probably still hundreds-of-billions-of-weights) high-complexity-variant thinking-phase sub-model weights, for the limited part of that response generation lifecycle where the thinking phase is actually occurring. During the "understanding" and "responding" phases, that reservation can be released for someone else to use! And for the vast majority of requests, the "thinking" phase is the shortest phase. So users end up sitting around waiting for the "understanding" and "responding" phases to complete before triggering another inference request. Which brings the per-user duty cycle of thinking-phase sub-model use way down.
[flagged]
I'd suggest that a measure like 'density[1]/parameter' as you put it will asymptotically rise to a hard theoretical limit (that probably isn't much higher than what we have already). So quite unlike Moore's Law.
Obviously, there’s a limit to how much you can squeeze into a single parameter. I guess the low-hanging fruit will be picked up soon, and scaling will continue with algorithmic improvements in training, like [1], to keep the training compute feasible.
I take "you can't have human-level intelligence without roughly the same number of parameters (hundreds of trillions)" as a null hypothesis: true until proven otherwise.
[1] https://arxiv.org/html/2602.15322v1
Why don't we need them? If I need to run a hundred small models to get a given level of quality, what's the difference to me between that and running one large model?
You can run smaller models on smaller compute hardware and split the compute. For large models you need to be able to fit the whole model in memory to get any decent throughput.
Ah interesting, I didn't realize MoE doesn't need to all run in the same place.
[flagged]
> There is not one person on the planet, who wouldn't prefer a doctor who is deeply considerate of the complexities and feedback-loops of the human body, over a doctor who is simply not smart enough to do so and, thus, can't. He can learn texts all he wants, but the memorization of text does not require deeper understanding.
But a smart person who hasn’t read all the texts won’t be a good doctor, either.
Chess players spend enormous amounts of time studying openings for a reason.
> Multiple small models, specifically trained for high reasoning/cognitive capabilities, given access to relevant texts
So, even assuming that one can train a model on reasoning/cognitive abilities, how does one pick the relevant texts for a desired outcome?
Doesn’t the widely accepted Bitter Lesson say the exact opposite about specialized models vs generalized?
Bitter Lesson is about exploration and learning from experience. So RL (Sutton's own field) and meta learning. Specialized models are fine from Bitter Lesson standpoint if the specialization mixture is meta learned / searched / dynamically learned&routed.
The corollary to the bitter lesson is that in any market meaningful time scale a human crafted solution will outperform one which relies on compute and data. It's only on time scales over 5 years that your bespoke solution will be over taken. By which point you can hand craft a new system which uses the brute force model as part of it.
Repeat ad-nauseam.
I wish the people who quote the blog post actually read it.
Why not both?
Scaling laws are real! But they don't preclude faster processing.
It's the same thing. Quantize your parameters? "Bigger" model runs faster. MOE base model distillation? "Bigger" model runs as smaller model.
There is no gain for anyone anywhere by reducing parameter count overall if that's what you mean. That sounds more like you don't like transformer models than a real performance desire
Is anyone doing any form of diffusion language models that are actually practical to run today on the actual machine under my desk? There's loads of more "traditional" .gguf options (well, quants) that are practical even on shockingly weak hardware, and I've been seeing things that give me hope that diffusion is the next step forward, but so far it's all been early research prototypes.
I worked on it for a more specialized task (query rewriting). It’s blazing fast.
A lot of inference code is set up for autoregressive decoding now. Diffusion is less mature. Not sure if Ollama or llama cpp support it.
Did you publish anything you could link wrt. query rewriting?
How was the quality?
Quality was about the same. I will say it was a pain to train since it isn’t as popular and there isn’t out of the box support.
Interesting, thanks! That's pretty cool though!
Because diffusion models have a substantially different refining process, most current software isn't built to support it. So I've also been struggling to find a way to play with these models on my machine. I might see if I can cook something up myself before someone else does...
Based on my experience running diffusion image models I really hope this isn't going to take over anytime soon. Parallel decoding may be great if you have a nice parallel gpu or npu but is dog slow for cpus
Releasing this on the same day as Taalas's 16,000 token-per-second acceleration for the roughly comparable Llama 8B model must hurt!
I wonder how far down they can scale a diffusion LM? I've been playing with in-browser models, and the speed is painful.
https://taalas.com/products/
Nothing to do with each other. This is a general optimization. Taalas' is an ASIC that runs a tiny 8B model on SRAM.
But I wonder how Taalas' product can scale. Making a custom chip for one single tiny model is different than running any model trillions in size for a billion users.
Roughly, 53B transistors for every 8B params. For a 2T param model, you'd need 13 trillion transistor assuming scale is linear. One chip uses 2.5 kW of power? That's 4x H100 GPUs. How does it draw so much power?
If you assume that the frontier model is 1.5 trillion models, you'd need an entire N5 wafer chip to run it. And then if you need to change something in the model, you can't since it's physically printed on the chip. So this is something you do if you know you're going to use this exact model without changing anything for years.
Very interesting tech for edge inference though. Robots and self driving can make use of these in the distant future if power draw comes down drastically. 2.4kW chip running inside a robot is not realistic. Maybe a 150w chip.
The 2.5kW figure is for a server running 10 HC1 chips:
> The first generation HC1 chip is implemented in the 6 nanometer N6 process from TSMC. ... Each HC1 chip has 53 billion transistors on the package, most of it very likely for ROM and SRAM memory. The HC1 card burns about 200 watts, says Bajic, and a two-socket X86 server with ten HC1 cards in it runs 2,500 watts.
https://www.nextplatform.com/2026/02/19/taalas-etches-ai-mod...
I’m confused then. They need 10 of these to run an 8B model?
No. 250 watts to run an 8B model.
> Our second model, still based on Taalas’ first-generation silicon platform (HC1), will be a mid-sized reasoning LLM. It is expected in our labs this spring and will be integrated into our inference service shortly thereafter.
> Following this, a frontier LLM will be fabricated using our second-generation silicon platform (HC2). HC2 offers considerably higher density and even faster execution. Deployment is planned for winter.
From https://taalas.com/the-path-to-ubiquitous-ai/
Personally I think anything around the level of Sonnet 4.5 is worth burning to silicon because agentic workflows work. There are plenty of places where spending $50,000 for that makes sense (I have no idea of the pricing though)
Just tried this. Holy fuck.
I'd take an army of high-school graduate LLMs to build my agentic applications over a couple of genius LLMs any day.
This is a whole new paradigm of AI.
A billion stupid LLMs don't make a smart one, they just make one stupid LLM that's really fast at stupidity.
I think maybe there are subsets of problems where you can have either a human or a smart LLM write a verifier (e.g. a property-based test?) and a performance measurement and let the dumb models generate candidates iterate on candidates?
Yeah, maybe, but then it would make much more sense to run a big model than hope one of the small ones randomly stumbles upon the solution, just because the possibility space is so much larger than the number of dumb LLMs you can run.
I don't work this way, so this is all a hypothetical to me, but the possibility space is larger than _any_ model can handle; models are effectively applying a really complex prior over a giant combinatorial space. I think the idea behind a swarm of small models (probably with higher temperature?) on a well-defined problem is akin to e.g. multi-chain MCMC.
Man, I'm in the exact opposite camp. 1 smart model beats 1000 chaos monkeys any day of the week.
What did you try and how?
https://chatjimmy.ai
I see; the chatbot demo in the Taalas page. If they could produce this cost effectively it would definitely be valuable. The only challenge would be getting models to market before their next revision.
When that genrates 10k of output slop in less latency than my web server doing some crud shit....amazing!
This is exceptionally fast (almost instant) whats the catch? Answer was there before I lifted return key!
This is crazy!
I'd love to know what's going on with the Gemini Diffusion model - they had a preview last May and it was crazy fast but I've not heard anything since then.
A lot of this post-training recipe feels reminiscent of DINO training (teacher/student, use of stop gradients). I wonder if the more recent leJEPA SigREG regularization research might be relevant here for simpler post-training.
Seeing half of an AR LLM's output tokens go to generating a predefined json schema bothers me so much. I would love to have an option to use diffusion for infilling.
One trick I learned for this was to use csv for LLM I/I and translate json <-> csv at the boundary layer
Oh neat. So have the llm output csv instead of JSON and then convert it? How would handle nested structures?
Depending on how it's nested, you could denormalize, think of how you could denormalize a one-to-many SQL relationship
So if you have a user that has many automobiles, maybe instead of Autos: [...] you could do Auto1Make Auto2Make etc.
I do wonder why diffusion models aren't used alongside constraint decoding for programming - surely it makes better sense then using an auto-regressive model.
Diffusion models need to infer the causality of language from within a symmetric architecture (information can flow forward or backward). AR forces information to flow in a single direction and is substantially easier to control as a result. The 2nd sentence in a paragraph of English text often cannot come before the first or the statement wouldn't make sense. Sometimes this is not an issue (and I think these are cases where parallel generation makes sense), but the edge cases are where all the money lives.
But I do wonder if diffusion models will be used in more complex Software Architecture for their long-term coherence, no exposure bias, and their symmetric architecture could work well with interaction nets.
Google is working on a similar line of research. Wonder why they haven't rolled out a GPT40 scaled version of this yet
Probably because it's expensive.
But I wish there were more "let's scale this thing to the skies" experiments from those who actually can afford to scale things to the skies.
Scaling laws mean that there's not much need to actually scale things to the skies. Instead, you can run a bunch of experiments at small scale, fit the scaling law parameters, then extrapolate. If the predicted outcome is disappointing (e.g. it's unlikely to beat the previous scaled-to-the-sky model), you can save the really expensive experiment for a more promising approach.
It would certainly be nice though if this kind of negative result was published more often instead of leaving people to guess why a seemingly useful innovation wasn't adopted in the end.
This doesn't mention the drawback of diffusion language models, the main reason why nobody is using them: they have significantly lower performance on benchmarks than autoregressive models at similar size.
Can't wait for the day I can actually try a diffusion model on my own machine (128GB M4 Max) rather than as a hosted service. So far I haven't seen a single piece of software that supports it.
You can try it today. You can get them from huggingface. Here is an example:
https://huggingface.co/tencent/WeDLM-8B-Instruct
Diffusion isn’t natively supported in the transformers library yet so you have to use their custom inference code.
I don't subscribe to the Python craze, but this could be interesting. Thanks!
I think diffusion makes much more sense than auto-regressive (AR) specifically in code generation comparing to chatbot.
Is this available as open source anywhere to try?
If this means there’s a 2x-7x speed up available to a scaled diffusion model like Inception Mercury, that’ll be a game changer. It feels 10x faster already…
Diffusion language models seem poised to smash purely autoregressive models. I'm giving it 1-2 years.
One appeal of it is for RL. If it ends up being a lot faster for generation, you'll be able to do a lot more RL.
If people can make RL scalable-- make it so that RL isn't just a final phase, but something which is as big as the supervised stuff, then diffusion models are going to have an advantage.
If not, I think autoregressive models will still be preferred. Diffusion models become fixed very fast, they can't actually refine their outputs, so we're not talking about some kind of refinement along the lines of: initial idea -> better idea -> something actually sound.
> If not, I think autoregressive models will still be preferred. Diffusion models become fixed very fast, they can't actually refine their outputs, so we're not talking about some kind of refinement along the lines of: initial idea -> better idea -> something actually sound.
I'm really curious about this, I'm but a simple client developer, so I don't actually grok some of the differences.
For lack of a better word, there's a "normie" position that "omg diffusion means it can edit!!111! big unlock!" -- I think that's cute but I also don't see it as intuitively correct. And I guess I don't even know why I don't see it that way. But regardless, it sounds like I'm correct there.
> If not, I think autoregressive models will still be preferred.
But here I get lost, at least so far, diffusion models seem strictly significantly faster, and on par with models with the same parameter count.
If that is the case, why would autoregressive models still be preferred?
Asking this also makes me realize I am treating "diffusion models are better" as a premise, if I'm asserting they're always faster and ~same quality...
RL suffers both from sparsity in a huge amount of dimensions as well as convergence brittleness when it's extremely difficult to get it to converge.
Feels like the sodium ion battery vs lithium ion battery thing, where there are theoretical benefits of one but the other has such a head start on commercialization that it'll take a long time to catch up.
Not really. Unlike with physical goods like batteries, the hardware for training a diffusion vs an autoregressive language model is more or less exactly the same.
Although the lab that did this research (Chris Re and Tri Dao are involved) is run by the world's experts in squeezing CUDA and Nvidia hardware for every last drop of performance.
At the API level, the primary differences will be the addition of text infill capabilities for language generation. I also somewhat expect certain types of generation to be more cohesive (e.g. comedy or stories where you need to think of the punchline or ending first!)
Same with digital vs analog
Digital came later but beat analog at almost everything?
Didn't thinking tokens resolve the most problematic part of autoregressive models (the first few tokens set the constraints the model can't overcome later) and give it a massive advantage compared to diffusion models by showing the thinking trace? I can see diffusion models being used as a draft model to quickly predict a bunch of tokens and let the autoregressive model decide to use them or throw them away quickly, speeding it up considerably while keeping thinking traces available.
The reason I mentioned "purely autoregressive" is that realistically I expect hybrid diffusion + autoregressive models to be the first popular diffusion models. I could be wrong though. And diffusion models have other tricks like really easy integration with simple classifiers.
Check out this paper where they use diffusion during inference on the autoencoded prediction of an autoregressive model: https://openreview.net/forum?id=c05qIG1Z2B