I wonder how big the Pro model is that Google is using behind the scenes to train these smaller ones.
Going on baseless speculation, the lack of accompanying pro models with these flash releases either means: 1) the model is too big to be economical, 2) google doesn't have the compute to serve the big model, 3) their big model has too many alignment issues to serve to the public.
edit: looks like benchmarks are up on https://artificialanalysis.ai/models/gemini-3-6-flash. It's solidly middle-of-pack. However, if you want to be most fair to flash, look at the intelligence vs time per task and intelligence vs outputspeed benchmarks. This is a very fast model.
edit 2: I use antigravity from time to time and in my experience, 3.5 flash is an underrated model, so long as you know what it's good for. It's very good at frontend (much better than gpt 5.5) and it's fast, so it's a great tool for iteration. I expect 3.6 to be no different.
It's also very possible that they know their big model underperforms chatgpt 5.6 and fable by too much, so they are focusing on what they can get wins in like speed instead.
That and/or the business case isn’t as clear when serving enormous models? You’re constantly stuck in a red queen’s race where your profitability window is increasingly measured in weeks because the Chinese are right behind you.
For small models (which are probably distilled from their big ones) you can serve them economically all the time and not hemorrhage money.
> For small models (which are probably distilled from their big ones) you can serve them economically all the time and not hemorrhage money.
For smaller models, you're competing with DeepSeek V4 Flash. (Which I think is a 284B A13B?) Subjectively, this feels about as smart as Sonnet 4.5, give or take. And it costs $0.09/$0.18 on Open Router, compared to $1/$5 for the latest Claude Haiku. See https://openrouter.ai/deepseek/deepseek-v4-flash#providers The developer antirez of Redis fame uses this as a local coding model.
DeepSeek did some extremely clever research on hybrid attention to get the prices that low, reducing per-user context cache sizes dramatically.
So, no, when it comes to low-price models, the US models probably can't sustain their current margins there, either.
There are also US based companies like Fireworks serving up the best open weight models with the compliances we need in US enterprise. Depending on the company, they may offer more/different jurisdictions, EU probably needs a Fireworks like company (haven't heard about one, maybe it already exists?)
The AI labs have been subsidising. When they try turn a profit, people will move to the fast followers. The only people that won’t are those that compete on leveraging the very latest models and even then, once spend and scale goes to the cheaper providers, we’ll see deeper research from those providers too. Think “PC compatibles beat IBM, Sun, SGI eventually”.
without any hard data one way or another your comment is worthless. "pile up subscriptions" - based on what? neither company is public. "piling up subscriber counts", "piling up API usage"? cool. how much money are they making? oh you don't know because they're not public.
the reality is one way or another that as long as there exists an alternative that a USA company could serve with the same compute rented from hyperscalers, this represents a threat, even if the extent to which is unknown
Why was this flagged? This is absolutely true, Americans seem to not actually want to use Chinese models at least for coding, maybe for other inference use cases but I haven't seen it. No one I know uses anything but OpenAI and Anthropic even if Chinese models are better or cheaper in many use cases.
First 6 of the most used models on openrouter currently are Chinese; that's true for code generation and other coding-relevant tasks too when ranked by share of tokens.
Of course openrouter is not representative because most users directly go to the model provider but it still proves your claim is very far from "absolutely true".
Even if not majority, it doesn't invalidate the point that just because OpenRouter has them as its top models that they are in general use among the population, rather than localized.
Honestly go use Deepseek v4 Flash, the data hosting in PRC is it's only downfall. It truly is an excellent model at the moment. It is opensource thankfully but there is friction there and this won't be as popular.
Why wouldn't the hyperscalers run these open models since they're much better than OpenAI and Anthropic at operating compute at scale?
I think the reason OpenAI and Anthropic stay ahead in revenues right now is because the models are improving too quickly to reliably compete with them on cost.
But, once model performance reaches a plateau -- they have to at some point, though perhaps years away -- that's when ability to operate compute infrastructure at scale becomes the secret sauce.
The big AI labs are likely safe until models stop improving fast enough to protect them from competition on cost.
This similar pattern has repeated in most technical booms prior to this.
When hard drive technology was improving fast enough that old hard drives were quickly obsolete, IBM could maintain good margins making hard drives. But once hard drives got good enough and advances were slow enough that innovation was not the only factor considered by drive purchasers, commodity hard drives started to take over and IBM had to exit those businesses.
The same is likely to happen once model improvement slows.
Kimi K3 is ahead of Fable 5 on several benchmarks.
So basically the angle went from "China cannot ever compete" to "China is six months behind" to "China is six weeks behind" to "China is six days behind but that's because they're distilling" and now you're saying "Yup sure, Kimi K3 is ahead on several benchmarks but you cannot host it yourself so this thing will go absolutely nowhere".
I mean: is it not a bit early to draw conclusions? It's been days since a chinese model is ahead of the very best / frontier US model on several benchmarks and you compare it to models who were clearly behind on everything.
You’re absolutely right and it’s heartening to see. I maintain a client with ~every provider you can think of and llama.cpp and it was really tiring the last few days to see people laundering other stuff through Kimi and Qwen. They’re not even open yet, the hype was based on their own blog posts, no one’s actually running these locally, the Qwen Max’s have never been open, Kimi’s API was 1/2 the speed the benchmarks was based on, when it was up, and had 60% downtime before they had to stop accepting new accounts, and their EULAs are “your inputs and outputs are ours.” May being clear-eyed benefit us both in the long run.
"You’re absolutely right and it’s heartening to see"
Damnit, I usually don't jump to LLM speech patterns, but this opening had me thinking you were a bot. But after checking your profile, I think you pass as human. I wonder when will be the time, this does not work anymore for me. (Creation date is a strong hint, but abandoned accounts can be hijacked)
Hehe, cheers, it really is funny & odd habit I have (usually when I'm in "everyone is wrong!" mode, haven't bothered to argue that, and see someone else arguing it :p)
That's the only explanation that makes sense. If it was frontier but cost or compute were limiting factors, they'd release it at an obscene price for the bragging rights. Google doesn't care that much about alignment, and I don't think it's likely to be significantly different than 3.5 anyway. The only reason it would need to be soft-canceled is if it's terrible, and has to end up in a ditch like Llama 4 to avoid shareholder panic.
3.5 pro was clearly a miss. It should have been in prod mid may, not MIA in late July. The brain drain at deep mind is a clear indicator that the people who know the most think that they can’t stay at the frontier.
Antigravity NEEDED to be game-changing. Without the stream of data that Claude, Codex, and Cursor enjoy there is little chance of getting an effective reinforcement learning loop. For the first time in its history, GOOG is at a meaningful data disadvantage, and apparently a cultural one as well.
"Without the stream of data that Claude, Codex, and Cursor enjoy there is little chance of getting an effective reinforcement learning loop"
Google literally giving everyone + student 18 month free subscription, those are source of cheap gemini + sonet,opus model that people selling/use with rotator proxy with thousands of account
I don’t think this is necessarily true, did we all forget how much Google cared about alignment that their AI wasn’t able to render a white polar bear?
My view then was they are optimising the models for inference ability on their own hardware AND use cases, which is often speed and time to first token.
They've somehow seemed to end up with terrible compute shortages, which again is surprising given how good Google is at infra deployments AND have their own hardware. From rumors out there they are turning down enterprise deals for Gemini because they don't have the compute.
The problem is they're falling further and further behind on frontier class on coding especially, and since I wrote that article it's got even worse with open weights models undercutting them on price AND intelligence.
They include an LLM response with every single Google search, whether it is warranted or not. That scale is, my guess, many orders of magnitude higher than what OpenAI and Anthropic serve. And for Google none of these are paid interactions since their LLMs do not (YET) insert ads into the responses.
So my guess is that Google will continue having compute shortages until the Gemini enshittification starts.
I don't think so. According to some very basic research there are around 8bn searches a day, or 250bn a month.
Let's assume Google serves AI overviews on every SERP (they don't) and don't cache them (they do, afiak).
And let's assume that each AI overview is 2000 tokens (blended input/output), that's 500T tokens a month.
It's rumoured that anthropic is serving somewhere close to 10Q tokens a month.
Now it may be that AI overviews uses vastly more tokens than that per search, but I doubt it based on speed to render the overview.
My very rough napkin math on this is that maybe AI overviews is consuming 100T tokens/month max (after adjusting for caching and SERPs that don't have them), which would be 1% of Anthropic token volume.
Well I asked the google AI mode thing what it thinks about your comment and it told me this (edited obviously):
"10 Quadrillion tokens a month means: 333 Trillion tokens per day and 3.85 Billion tokens generated/processed every single second, 24/7."
"At an incredibly cheap, subsidized infrastructure cost of $1 per million tokens, serving 10 Quadrillion tokens would cost Anthropic $10 Billion per month ($120 Billion a year) just in inference compute."
It also had this to say about how google's AI overview works: "Google doesn't just feed the LLM your 5-word search query. The system scrapes the top 10–20 web results, feeds thousands of words (tens of thousands of tokens of context) into the model, processes it, and then outputs the result."
Oh, and it does all of that in less than two seconds. Honestly, whatever Google is doing with its infrastructure is so far ahead of everyone else, I can't believe you fell for such an obvious lie.
There are also extremely obvious holes in your comment:
>Let's assume Google serves AI overviews on every SERP (they don't) and don't cache them (they do, afiak).
Try it out for yourself. Add a few random letters or punctuation. They cache nothing.
The tokens served number might include cache tokens which are a huge chunk of agentic token spend - and even with that the estimated burn rate for anthropic doesn't seem wildly off? They spend 1.25 bn per month on their deal with SpaceX alone.
They definitely cache results - I've searched and re-searched an identical query back to back a few times and seen identical results from overview. They are definitely throwing a stupid amount of compute towards these ai results nobody is paying for - changing punctuation and stuff does get you a different response - but they're not doing no caching.
Certainly what they're doing with their infrastructure is impressive but it's not super meaningful at the end of the day for a for profit company to be really impressively good at burning tens of billion dollars on a service nobody pays for while the same tech from their competitors is quickly becoming one of the largest spend categories for many software engineering teams
> focusing on what they can get wins in like speed instead
Speed as a differentiator has always been Google's thing. They (used to?) show the microseconds it took to query & rank web-scale search results. Chrome, notoriously, focused on speed at the expense of resource use. The very many efforts to efficiently speed up Android & its runtime since its inception, and so on...
> their big model underperforms chatgpt 5.6
Possible but TFA claims:
We have started our most ambitious pre-training run yet, for Gemini 4 ...
There was some recent reporting that a July release of the Pro model got pushed back for exactly that reason. Its performance was not good compared to the OpenAI/Anthropic big models. They are having a lot of problems with posttrain.
It would be a shame if they cannot beat Kimi K3 or Qwen3.8 Max, both of which are claimed to be Fable-like. If that is true, it will be [or would be] the first time a major American lab falls behind a Chinese competitor.
Google can't compete with China, neither can Meta. Only two labs in the US can keep chucking billions at the frontier race. Everyone else has a real business to run.
China can keep up because it's cheaper to run a frontier lab there. They also have more researchers and a stronger cultural inclination for this sort of thing. And I guess the business case in China doesn't have to work as well as it does in the US.
> China can keep up because it's cheaper to run a frontier lab there
Not sure if this is what you meant, but their training runs are significantly cheaper. This was one of the big shockers from the Deepseek R1 paper. US foreign policy has helped to ensure that the Chinese are compute constrained, so they literally cannot buy the most expensive and powerful training rigs.
This has led to a steady drumbeat of innovations which are not revolutionary on their own but stack together to make things much more efficient.
China will accelerate on ML research and optimization regardless of US export control policy. They want to win or at least not lose just as much as the US and will pull every reasonable lever at their disposal to do so. NVIDIA and the US government have no say in this.
I’ve often thought it was the cost of electricity and crypto mining being banned a few years ago. China has a lot of “stranded electricity” which fits this usecase well.
I think they've been behind for a while - flash isn't even that competitive with glm 5.2 and from their hype around 3.5 flash at launch - that was certainly not intended to be the case
When it is cheaper, and the "lower quality" model is adequate for the task at hand.
Plenty of problems have a low(er) skill/intelligence floor, anyone who uses the dual-mode agent paradigm (plan, then act) figures out the second phase can be completed by a less capable model. Even when disregarding costs - speed is important here because the agent can rapidly iterate without human supervision, based on compiler errors, lint and test failures
A coding agent driven by a large LLM can delegate smaller tasks to a faster model. For example searching through the codebase for references, examples, or established patterns. They are treated as tools and don't pollute the main agent's context.
I wonder if the broad use of AI overviews on Google search results is having an impact. Maybe the numbers make it more profitable to use their compute on several billion searches a day rather than selling API access.
I doubt it's more profitable - ai overview is free/runs even when signed out in incognito, while the frontier models from openai and anthropic are nauseatingly expensive, especially for enterprise, and have a ton of users who are willing to cough up that money. Even if they aren't profitable because their cogs is even higher, they certainly have revenue with ai overview doesn't really have
AI overview is just a summarization of the top 2-3 results. Of course at Google scale that will still need a ton of compute, but the requirement for generating an overview is many orders of magnitude lower than asking the same question in Gemini.
It’s a small multilingual embedding model designed for things like search, RAG, and semantic similarity. It supports a fairly large context window and is designed to run efficiently on a CPU in a GPU starved world.
The interesting part is that it builds on BitNet, using ternary weights of -1, 0, and 1 instead of the usual floating-point weights. That should make indexing and searching large amounts of text much cheaper without giving up too much accuracy.
Considering how cheap the subscriptions are, it looks like agentic coding is a low margin business. If they can sell you a subscription for the chat, it is profitable, but if you try to use the subscription to its limits, you're probably making them lose money.
absolutely, knowing OpenAI try to break into ads market tell the whole direction that pure AI is not that profitable tbh (especially with how cheap chinnese model are)
It seems like there are some credible rumors that Google is actually winning in terms of actually building models that work and don't lose money- between how they're able to price them, the TPU advantage and their capex advantage (being able to raise debt + just having a lot of cash - well I said not lose money... more like not go bankrupt).
From the outside they look like they're behind in terms of frontier models, but I think they might be the best positioned to not go out of business when the bubble pops.
Also look at the fact that they've been able to deploy AI-assisted search at google scale. It must be another order of magnitude larger (at least) than the model deployments for OpenAI and Anthropic.
Of course unless you're inside Google it's impossible to know for sure.
In terms of open models, Gemma 4 beats the pants off everything else to the point that paying for APIs becomes hard to justify. Qwen has the meme-share for coding, but it feels much less well rounded. I have no doubt that Google have both the infrastructure and the expertise to curb stomp everyone else, should they resolve in earnest to do so.
Lest we forget, "Attention is All You Need" came from Google.
It also came directly from the university of Toronto, and the university of Toronto seeded all American frontier labs (including Grok (why do you think they could start so fast))
Interesting, glad to hear. We have gemma4 at work, and I was considering localhosting qwen, but gemma4 is so far behind the Opus and Fable I have at home that I've decided to hold off for another model release.
"We have a company provided Toyota at work but it is so far behind the Ferrari I rent at home that I've decided to hold off for another model release."
At 20x the parameter count I should hope GLM beats Gemma! But is it 20x better? Expertise is demonstrated, not by making big models, but by making small ones. Bigger isn't better if you can't run it at all.
It's rumored that Gemini 3.5 flash has a >50% margin, and I'd imagine 3.6 flash is even higher.
I do not think OpenAI or Anthropic are actively chasing margins - though, Anthropic is supposed to be profitable on some form of non-GAAP accounting...
I suspect Google isn't really interested in seeing how far it can get dragged into a race of selling dollars for $0.25, and is more interested to see if it can stay in the race selling $0.50 for a dollar - when everyone else is losing or barely breaking even.
It kind of doesn't make sense though, because typically a large org like Google can afford to crush competitors on pricing. They could probably even go toe to toe with chinese model pricing for years without feeling it.
Maybe they don't want to price war with the other labs so they can comfortably maintain healthy margins on selling them compute?
google has to make money. flash is awesome. you can run it free on their infra and the performance and latency is excellent for what you wait and pay for right now, with great perf per watt. every person in the world going to google.com runs it. every query. its far larger than free gpt, localhost qween and what not.
it's their pro that isn't awesome at all. in fact, their pro kinda suck now that everyone else woke up.
But maybe the TPU advantage is in inference? That's what I assume because the number of compute cycles are going to be all in inference vs training. So they could train on GPUs if they want.
"As of October [2025], OpenAI's compute margins reached 70%, up from 52% at the end of 2024 and double the rate in January 2024, [The Information] said, citing a person familiar with the figures."
As for Anthropic, the rumors I remember seeing for their API margins were more like 85-90%, but I don't have a reference at hand for those. But once you know the API is wildly profitable and the subscriptions are roughly break-even and not even a big slice of their income, all of the investment makes a lot more sense.
It says the original report was in the Information, which I can't see, but I'm skeptical that they includes the training cost? And how much that changes the figure?
That's the profit margin on inference, not overall. Each model does end up being profitable over its lifetime, but the money they're making is being immediately churned into buying more data centers & the training for the next giant model up, so they're not profitable overall at the moment. It's a bit like how Amazon kept churning their profits into more growth instead of taking the profit early.
That said, Anthropic has supposedly crossed over into profitability and made $1 Billion in profit so far this year, in the lead up to their IPO. Being profitable sounds good for launching on the stock market! But as a customer, that's noticeable in the downtime due to lack of compute, and only getting 50% access to Fable.
OpenAI might not be profitable, but they've got so much compute access that they've been able to give their customers full access to Sol, and as a result they've almost doubled their Codex subscriber base in the last two weeks (6 million on July 12, 10 million on July 21 - that would be an extra $1-$10 Billion in Annual Recurring Revenue that they've gained in just these 2 weeks). Doing the unprofitable thing in the short term can result in outsized rewards in the long term.
On the other hand, the consumer side of the market seems to be less competitive right now.
OpenAI's new Mac app doesn't even have a normal "Chat" option now. OpenAI might be chasing coding and b2b sales more now that they realise very few regular consumers pay for subscriptions.
I was already impressed by how fast 3.5 Flash was. But I've never compared it to other models in its class for coding.
Why? Coz models in that class are not very useful to me. Time saved waiting for responses usually just turns into time wasted replying to low quality responses.
Google need to release a Pro model ASAP. I am skeptical of the "maybe they don't have the compute to run it" thing. Anthropic were (probably) in that situation with Mythos and they announced it anyway - that's the obvious play for investor relations as well as hype for your product.
Pre-trains take a huge chunk of your compute offline, incurring both an raw expense (24/7 max power for all training clusters) and an opportunity cost (could have sold excess compute during that time). They also don't come with any great guarantees, as lots of techniques look good on small scale and crumble or plateau once scaled.
I wish someone would convince Google to may be leave the Google search be without AI responses and use all their resources for a Gemini subscription/API..
I've been using 3.5 flash in Android Studio on a Dart/Flutter project, some of it pretty complex and using brand new API's and native code for agentic tool calling in the app.
It used to be that you could find the edges of the training set pretty easily. No longer.
I think it's 2. I frequently get told there's no capacity for Pro and the query is answered by Flash with extended thinking. And tbh it's hard to tell the difference between the two, especially if you're not coding with it.
Artificial analysis always seemed sketchy as hell. If you read some of there methodology you’ll see a lot of <=3 repetitions on a particular pass for a given model. So low for calling a frontier model over the public internet ????
Flash versions were often ultra competitive and their best in the range along with openai mini models. Always been gemini most useable and best model with nano b. Frontier is much more competitive. Anthropic haiku is like 2025 flash…
I suspect the entire 3.x family is fundamentally problematic and we'll need to see an architecture change before they're half decent like back in the 2.5 days again.
What I think is really going on is an attempt to segment the market in favor of Google's strengths. It's a bet that models are "good enough" for many use cases even before they reach human-level intelligence, and Google is trying to capture workflows where quantity beats quality.
They are likely deliberately avoiding the SoTA race for a few reasons:
1. Their best models are marginally better than current SoTA releases.
2. They'd like to let Ant/OAI make mistakes with safeguards / let them get the regulatory heat. The unknown unknowns are huge with SoTA models (eg OAI accidentally hacking huggingface) and they are protecting their reputation.
3. They want to encourage companies to become cost conscious because they can likely win on price in the long run. Getting market share in "quantity beats quality" workflows forces companies to establish processes to choose the "cheapest acceptable model", which is a good environment for Google.
Going on baseless speculation, the lack of accompanying pro models with these flash releases either means: 1) the model is too big to be economical, 2) google doesn't have the compute to serve the big model, 3) their big model has too many alignment issues to serve to the public.
edit: looks like benchmarks are up on https://artificialanalysis.ai/models/gemini-3-6-flash. It's solidly middle-of-pack. However, if you want to be most fair to flash, look at the intelligence vs time per task and intelligence vs outputspeed benchmarks. This is a very fast model.
edit 2: I use antigravity from time to time and in my experience, 3.5 flash is an underrated model, so long as you know what it's good for. It's very good at frontend (much better than gpt 5.5) and it's fast, so it's a great tool for iteration. I expect 3.6 to be no different.