Why is this exchange reminding me of that You Can See Everything documentary?
The analogy made sense to me: Tokenmaxxing policies were the result of an irrational drive by execs who just craved an, ahem, explosion of metrics due to the excitement over new technology. Now that cost sensitivities are setting in, there’s a growing sense of unease that maybe they were in an altered state when they created the mess and now that the excitement has passed they have to clean it up.
So people had this drive to use tokens, and once they experienced the release from using the tokens they went into a refractory period where their reasoning is no longer being overridden by the need to use tokens? And later I guess they'll start to feel more and more of a need to use tokens and this will repeat?
Or is it just a throwaway phrase that makes no sense?
We’ll see if the cycle holds true for tokens or LLMs specifically, but for sure there is a larger “new sexy tech” cycle with a strongly emotional output===outcomes metrics irrationality phase to it.
It may well be that tokenmaxxing does not make sense. I don't think people were tokenmaxxing in some kind of compulsive urge to pursue a release of pent up tokens which were making them token horny. The analogy doesn't make sense.
I understand his assertion, and I can make up an argument that fits it too, probably a few arguments that aren't even compatible with each other, but what you said isn't what he said, you are just assuming all of this, and also nowhere here does it explain how this fits an analogy of "post nut clarity".
Their naming scheme seems a bit tough to follow. But it does seem like they finally learned their lesson and announced something at the same time it's actually available!
Maybe I'm not on the latest models (I do actually pay the min plan), but just this week I tried Gemini again. It still doesn't even get the current date/time of day correct. If it cant do that, I cant trust any output from it. I have no idea how Google fumbled so badly, but benchmark scores don't mean a thing if you cant even get your LLM to output the current date and time correctly.
You could vibe code a better LLM with Fable than what they have delivered. It is embarassing.
Is this a replacement for Flash-Lite? Because it’s being used to serve answers for Search.
The new naming scheme seems shockingly poor though, “Pro” and “Flash” made intuitive sense to me, and Anthropic and OpenAI kept a sense of scale. Nobody is going to intuitively understand “Oh yes, Hydrogen is smaller than Carbon”, even though it is.
It will be the same as these smaller models from OAI and Ant. Strong in benchmarks but pretty bad in real world use.
I've read this same comment every time a new Luna came out, and the difference between free GPT (Luna) and paid (Sol and Astra) is still outrageous for general use.
It's definitely not the case that small models only do well in benchmarks. We're working on a battle system on top of OpenCode, so have data for real people doign real work.
The naming of these Gemini 4 models is ridiculous.
In order of sophistication and cost: Hydrogen -> Argon -> Carbon
It's no longer ordered alphabetically. That would be Argon → Carbon → Hydrogen.
And it never was ordered by atomic mass, which is arguably the best ordering if you're going to name your model classes after chemical elements. That would be Hydrogen → Carbon → Argon.
It's also not ordered by prevalence in the universe. Hydrogen is, of course, the most abundant element in the universe, and Carbon is substantially more abundant than Argon. So that ordering would also be Hydrogen → Carbon → Argon.
It's not ordered by position on the periodic table, electronegativity, boiling point, melting point, or any other remotely defensible chemical property I can think of.
It's not even ordered by letter count, Scrabble score, or NFPA 704 Flammability Rating, all of which would be Argon → Carbon → Hydrogen.
The choice is really odd. Hydrogen and Carbon could have an organic chemistry theme going if they chose Oxygen, but then they go with Argon? Anthropic and Open AI have the names figured out in terms of obvious size: Luna->Terra->Sol->Astra and Haiku->Sonnet->Opus->Fable->Mythos makes intuitive sense. I have no idea how the Google model size class works.
Per Business Insider[0], Argon and Carbon were part of an alphabetical lineup:
> An internal document shows that Google has tested a series of Gemini 4 models, internally named Argon, Barium, and Carbon.
Then I suppose they tacked on Hydrogen since it's supposed to be a "light" model and hydrogen is the lightest element. I agree that it's super confusing though.
+1 to this the naming is totally opaque and ridiculous.
There is no hint of what each one does. I get that these were internal code names, fine, but to everyone outside Google it's not helpful to have these meaningless names.
ChatGPT and Claude aren't much better - there is still quite a lot of assumed knowledge there assuming you know what e.g. a sonnet is and what an opus is - but at least if you squint there is something there.
Google's names just seem to be randomly generated.
Maybe they were worried that the 3.1 Pro model being around for so lunch ruined the brand? But 3.8 Flash was pretty decent I thought so it's not like "flash" was burnt. I dunno.
This is a good blog post because of the graphs comparing directly to Terra and Sonnet for both wall-clock time and cost per task. This is the exact information I'm looking for.
My TL;DR is: seems like a better choice than Sonnet 5.5 in areas outside of coding, and at high reasoning it's smarter than Terra but somewhat more expensive.
I really hope hydrogen keeps the excellent conversational ability of 3.8 flash. For brainstorming or even decoding the dense output of the powerhouse models, flash was a real delight to use. Its the first non-SOTA model that I would actually go to often over the larger models, it was fast, plenty smart enough, and generally wrote really well. Here's to hoping hydrogen keeps that up
In the Gemini app of over the API? In the app Ive found it to write poorly, probably a system prompt thing. I've done some stuff with Argon over the API and it wrote well.
Pretty big Gemini fan, but my initial results with Hydrogen and Carbon are disappointing.
They are trading improvements on pass@1 pointless benchmarks for worse disasters at p90/99 that you're going to hit every week in agentic workflows (if not daily):
overreach (doing more than asked), overwork (when it should have failed fast), harm (doing the opposite of what was asked), orthogonal work (doing something completely unrelated to what was asked), etc.
Google doesn't seem to be the only lab making this trade...
The set of benchmarks they released are disappointing. 10 mostly random benchmarks. Really?
No one on HN seems to get that agentic workflows are not the only name in the game. Everyone doing agentic development is either on a plan or with a custom enterprise offering, either way using higher-performing models. These entry level models are (also) for building apps on top of while you are paying API prices.
Are we sure the guy that creates google naming schemes is not held hostage and those are his desperate attempts to communicate with the outside world in concealed way?
As this is a) being served in Search, b) supposed to be 2.5x faster than 3.8 Flash, c) the element with lowest atomic weight; I deduce that, despite being more capable than 3.8 Flash, now the “entry level” model set to replace 3.5 Flash Lite.
But it is priced the same as the non-discounted 3.8 Flash, considerably higher than what anyone is paying today for 3.8 Flash or 3.5 Flash Lite or the deprecated 3.0/2.x Gemini models. How does this work? Is Google abandoning the “fast and cheap” tier? That seems like a mistake since plenty of workflows are absolutely price inelastic, meaning when the 3.8 Flash discount ends at EOY, lots of current enterprise Gemini customers would have to take their business to another shop??
Same impression here. If this is Google's strategy we'll have to migrate to Luna, because costs are already high for our workload on 3.8 Flash (which in itself was already a 40% increase compared to 3.7 flash due to being more verbose and thinking more) and this now doubles cost once again.
According to Artificial Analysis, it's much cheaper per task than 3.8 flash, 3$ for H xhigh vs 6-7$ for 3.8 medium-high. I don't understand how though, since AA says it has more output tokens per task.
This is based on what we measure from the API - we see different endpoints depending on how they are served can have different cache hit rates (cache hit rate is defined as the % of cacheable tokens that actually come back as cache hits)
Are we going to talk about naming and confusing ways to access it, or are we going to discuss how capable the model is?
Every single time a Google model comes out, an army of Anthropic shills sidetrack the discussion.
Can we look at what the model does and discuss whether Google is still in the game or not? (naming models and whether you need to access it by clicking a drop-down menu or making a prayer ain't really relevant)
Argon, Carbon, now Hydrogen. If Google keeps climbing the periodic table, the roadmap writes itself: Californium finds gold in your codebase (enterprise only)
218 comments
[ 3.4 ms ] story [ 93.0 ms ] threadHow is tokenmaxxing like ejaculating exactly? I don't follow the analogy.
The analogy made sense to me: Tokenmaxxing policies were the result of an irrational drive by execs who just craved an, ahem, explosion of metrics due to the excitement over new technology. Now that cost sensitivities are setting in, there’s a growing sense of unease that maybe they were in an altered state when they created the mess and now that the excitement has passed they have to clean it up.
Or is it just a throwaway phrase that makes no sense?
1) There was a false narrative called tokenmaxxing whereby one is more productive by spending more tokens
2) This narrative was fueled by discount prices on tokens (which still happens to this day on subscriptions)
3) There is a point where we collectively realize that tokenmaxxing was a state of mind that is not rational
4) We will then pay API prices because LLMs are obviously amazing but no one will offer subscriptions
5) API costs of Opus 5.5 and Fable tier become uneconomical
6) That leaves us in Sonnet/Haiku tier
You could vibe code a better LLM with Fable than what they have delivered. It is embarassing.
Why are there 7 different ways to call this model from Google? Which one should I use?
Is no one at the wheel calling the shots at Google and it's a PIP every man for himself? I want to try them out, don't know how.
The new naming scheme seems shockingly poor though, “Pro” and “Flash” made intuitive sense to me, and Anthropic and OpenAI kept a sense of scale. Nobody is going to intuitively understand “Oh yes, Hydrogen is smaller than Carbon”, even though it is.
I've read this same comment every time a new Luna came out, and the difference between free GPT (Luna) and paid (Sol and Astra) is still outrageous for general use.
Right now DeepSeek Flash 4.1 Flash has a 64% win rate against GPT 6.1 Sol. Small data still, but it has 33 wins vs. 18 losses, which is confidence interval of 52% - 78% win rate. https://agentduel.com/leaderboard/models/deepseek/deepseek-v...
In order of sophistication and cost: Hydrogen -> Argon -> Carbon
It's no longer ordered alphabetically. That would be Argon → Carbon → Hydrogen.
And it never was ordered by atomic mass, which is arguably the best ordering if you're going to name your model classes after chemical elements. That would be Hydrogen → Carbon → Argon.
It's also not ordered by prevalence in the universe. Hydrogen is, of course, the most abundant element in the universe, and Carbon is substantially more abundant than Argon. So that ordering would also be Hydrogen → Carbon → Argon.
It's not ordered by position on the periodic table, electronegativity, boiling point, melting point, or any other remotely defensible chemical property I can think of.
It's not even ordered by letter count, Scrabble score, or NFPA 704 Flammability Rating, all of which would be Argon → Carbon → Hydrogen.
At this point even using Starbucks sizes would be better.
> An internal document shows that Google has tested a series of Gemini 4 models, internally named Argon, Barium, and Carbon.
Then I suppose they tacked on Hydrogen since it's supposed to be a "light" model and hydrogen is the lightest element. I agree that it's super confusing though.
[0] https://www.businessinsider.com/google-employees-test-new-ge...
alphabetical skipping the weird names
next would be iron krypton and lithium
There is no hint of what each one does. I get that these were internal code names, fine, but to everyone outside Google it's not helpful to have these meaningless names.
ChatGPT and Claude aren't much better - there is still quite a lot of assumed knowledge there assuming you know what e.g. a sonnet is and what an opus is - but at least if you squint there is something there.
Google's names just seem to be randomly generated.
Maybe they were worried that the 3.1 Pro model being around for so lunch ruined the brand? But 3.8 Flash was pretty decent I thought so it's not like "flash" was burnt. I dunno.
The naming scheme alone has no rhyme or reason. Model names should not be a source of cognitive load.
Argon, Carbon, Hydrogen?
Luna, Terra, Sol, Astra, Betelgeuse?
My TL;DR is: seems like a better choice than Sonnet 5.5 in areas outside of coding, and at high reasoning it's smarter than Terra but somewhat more expensive.
They are trading improvements on pass@1 pointless benchmarks for worse disasters at p90/99 that you're going to hit every week in agentic workflows (if not daily):
overreach (doing more than asked), overwork (when it should have failed fast), harm (doing the opposite of what was asked), orthogonal work (doing something completely unrelated to what was asked), etc.
Google doesn't seem to be the only lab making this trade...
The set of benchmarks they released are disappointing. 10 mostly random benchmarks. Really?
But it is priced the same as the non-discounted 3.8 Flash, considerably higher than what anyone is paying today for 3.8 Flash or 3.5 Flash Lite or the deprecated 3.0/2.x Gemini models. How does this work? Is Google abandoning the “fast and cheap” tier? That seems like a mistake since plenty of workflows are absolutely price inelastic, meaning when the 3.8 Flash discount ends at EOY, lots of current enterprise Gemini customers would have to take their business to another shop??
It's cheaper than the last Pro models.
Every single time a Google model comes out, an army of Anthropic shills sidetrack the discussion.
Can we look at what the model does and discuss whether Google is still in the game or not? (naming models and whether you need to access it by clicking a drop-down menu or making a prayer ain't really relevant)
Feel like this pricing only makes sense for a product where you don't even select the model. Maybe Antigravity is like that.