There were so many memes about “Gemini Hydrogen” after they released Carbon. That they actually released a model named Hydrogen makes this very funny
But a new Flash model is very welcome. Argon and Carbon are very slow. In the time it takes for them to complete a single task, Opus already managed to complete multiple things. They’re so slow, I started using 3.8 Flash over them. (A lot of it is also because of how terrible agy is. Before doing any task, Gemini feels the need to hover up a bazillion files into its context when the task is something simple like “update the docs of this controller”)
Google’s segmentation here is keeping them from using their existing account dominance to eat away at OpenAI and Anthropic personal subscriptions, tbh. Adopt a lower usage limit for Carbon on the $20 plan and it’s suddenly insane value.
This is essentially what Anthropic does with the advisor mode and what GitHub does with their multi-model pipeline using the HydraFusion model, which is just a large set of different models with different qualities used in a pipeline to produce an answer.
What I don't understand is how they all make it work with caching in general. Looking at how the advisor mode in claude work you seem to only be billed the output tokens and cache hit on the whole session when it calls out to the advisor, making it super cheap compared to running the more expensive model for the entire session.
Pretty sure on advisor it needs to send the entire context to another model. The cost is mitigated a bit because there would be few output tokens from the advisor model normally.
Orchestrator seems to automatically delegate things to subagents as required is my read of it. This should be fine for caching.
When launching a swarm of smaller models, they can theoretically share cache amongst themselves, but not with the larger model (that already has it cached). They have different architectures.
Further, it wouldn't really make sense for them to share the same cache anyway! The swarm, that is, they'd be working on independent things...
My best guess is because models are segmenting more into BiS at specific tasks in a way that's orthogonal to small/good/better/best framing. Kind of the same framing as car models, where each one has its own pros and cons vs all other model, instead of being strictly better/worse or just smarter/faster/cheaper.
I'm really curious how folks feel about this. I regularly use subagents, but typically make my primary implementation agent use Carbon (or equivalent).
I set up my subagents with specific models based on how complex I believe the problem space is. I'm nervous to delegate that authority to a model (since models tend to be... optimistic about their capabilities.)
That said, not every coding task is a difficult coding task. Plenty are boilerplate, plumbing, or simple test authoring that could use a flash model, so maybe this would downgrade nicely.
On the other hand, it's trivial for me to set up some guidelines to create different subagents for different task domains, and just pin those models to those tasks -- implementor (hard) and implementor (easy) or whatever.
The only reason to do this is to save usage on the subscription. but because the value for money is so terrible in antigravity, you’re better off just using Claude Opus or GPT Sol for everything.
I’m glad Google is kinda back, but today there’s no reason to code with Gemini.
Im slowly migrating more of my work to Gemini too, Gemini is a workhorse now, 1M context is huge and their reasoning took a huge leap with Argon. I'm not in software, I'm in physical engineering, and where Claude is unpredictable w.r.t. engineering risk, Gemini is steady, measured. That's been my perception, anyway.
My qualitative take to support this is that Gemini seems to have the strongest spatial reasoning that translates nicely into CAD and engineering domain challenges. Less clear, but I also suspect a better world understanding of materials and their interactions in an intuitive way. I have this sense that Claude can tell me everything there is to know about how to spec a gasket, but it’s like Gemini actually knows what a gasket feels like and can intuit what it’s going to do.
My partner runs an internal computational mechanics benchmark that hasn’t tracked progress with model advancements (flat/no progress since early 2026) until suddenly saturating from argon and carbon. My take is that it’s due to spatial reasoning (most of the challenges are programming that require a lot of shape rotation type thinking), but we don’t know just yet why Gemini is doing so much better.
Carbon is the first "better than meh" model form Google in more than half a year. It's been out for less than a week. How can you consider it "steady"? Their subscription and even the API has been terribly unstable throughout the past few days, so I'm having really trouble believing any of this.
what source could their possibly be? Argon has been out for a while now and has been solid. Even flash 3.8 is great as long as you check it's work on more complex tasks, and doesn't need 12 hours to chew through problems like claude does.
I use argon and opus daily and I can tell you I have to use opus to review argon because argon goes and build things and also creates bunch of bugs along with it. I wish it weren't but that's the reality. It's great with deep cross inspection better than claude but it's slow and still need review by another frontier model.
I can say the exact opposite was also true? No model was able to one-shot perfect solutions in my reasonably complex codebase and always required a through red-team by at least one other model. Doesn't mean one model is bad and another is better, just means no model is perfect.
Whatever they thought of when they said "measured"? Is that too much to ask?
Ok, you're either a troll or you objectively don't know what you're talking about. Google should be ashamed they even released 3.8 Flash. They published models with such high frequency and always made a huge jump on benchmarks. Then they were caught, when a new version of AA came out and their numbers dropped like a stone.
Glad you found what works for you but before Carbon, Gemini models have been utter garbage for more than 6 months now. I'm happy to debate anyone on this. Even DeepMind engineers agreed with this.
Gemini 3.8 Flash is the most benchmaxxed model to date and even Argon hallucinated so much, it was impossible to use it for any professional work.
I disagree with this. Flash 3.7 and Flash 3.8 are excellent models and put (for me) Gemini back on the radar. I’ve been using them a lot since release, and still do for simple implementations and grunt work.
They’re also excellent (together with Flash-Lite) to use for consumer facing features where speed and cost matters
I've been using 3.8 Flash (and Argon and Carbon) to clean up a lot of Claude mess. But I fully agree on 3.1 Pro, it was only good for conversational stuff and some deep dives, but not for actual agent work.
I'm honestly surprised they only see 25% savings with co-trained models and special harness support. With Mjolnir bridging Opus to DeepSeek Flash over ACP we see savings of 60%.
I.e. in this context Opus is the big brains orchestrator and it delegates to DeepSeek Flash subagents?
Intriguing, I might have to try this out - I have an OpenCode Go sub for the deepseeks of the world, and a Claude sub too. Sounds like I could have the best of both worlds here.
Isn't that because Chrome is forced to use the same rendering engine on iOS? Whereas on macOS, Safari is total garbage across all of Google's sites and products compared to Chrome, for very obvious reasons.
> Orchestrator takes a dynamic, bidirectional approach between Carbon and Gemini 4 Hydrogen within the evolution of a single trajectory. For example, Carbon can start in a read-only phase to explore dependencies, resolve ambiguities, and build a concrete plan with explicit re-planning triggers. Hydrogen could then take over to write edits, update packages, and run tests. If Hydrogen hits friction, it would hand back to Carbon to inspect the state, repair the plan, and resume execution. This allows Orchestrator to reach near-Carbon task success while being faster and more quota-efficient. On Terminal-Bench 4.1, Orchestrator reached near-Carbon task success rates while, on average, consuming 24% less quota and completing 19% faster.
Is this now how e.g. Opus works now? I use Opus 5.5 on high for just about everything now, and my bill's not gone up (gone down actually). I had assumed under the hood it was doing this (e.g. I'm sure its not literally Opus doing the log checking / file searches). Towards Google's choice, I'm skeptical introducing a new concept (Orchestrator) for this is really necessary, when they could have just done it by default, unless its plainly lower quality than pinning Carbon.
I found when fable first came out when I switched from opus my usage actually went down. Personal theory was that since it was smarter it took fewer tokens to accomplish the same tasks.
Indeed, all the 4/Hydrogen submissions were from different sources, but *sigh* the exact same conversation is happening here as other threads.. almost as if it's the same topic.
I don't understand which model is supposed to be better. I think Carbon is the big model and Hydrogen the Flash equivalent? Only because of a Twitter leak though. Like I have no earthly idea based on the naming.
Seems weird to pick a name that is so much a noble gas, and then not have the other names follow it. Is that because Nitrogen doesn't have parameters enough? Or Neon being a weird model name if it came to that? And why not Helium, instead of Hydrogen? Is that supposed to indicate rock-bottom pricing? What about the cheap cheap 3.1-flash-lite for pipelines then? Will they phase it out, and replace it with Hydrogen as the new lower price range? None of it makes sense!
There is nothing confusing about it if you're at least a little bit educated. If anything, they are the least confusing, Ant drops the ball at opus->fable.
For OpenAI, it's:
Luna, means "moon", smallest.
Terra, means "earth", bigger than the moon.
Sol, means "sun", the biggest object in the solar system.
Astra, plural of the word "astrum" which means "star".
All of those are deducible, sure. The problem is that the pattern works only when they are all listed together. Person sees the list and can logically infer where they belong. But when I see the single model name detached from it's siblings, like "OAI releases new Terra 7.7" headline, for quite a long time I had zero idea where it belongs just from my memory. Was that a big model, a small model, a medium model - no idea. Now after a few HN topics I do remember the lineup and can place the new announce where it belongs.
To be fair, I knew all of those words already, and it never even occured to me that these names are supposed to be compared to eachother to indicate size or something. So, at least to me, it's not that it confuses me, it just imparted no information, because a collection of astronomical names and terms doesn't automatically indicate an ordering to me, it just indicates that nerds like naming their stuff after scientific terms and concepts, sci-fi stuff, etc.
I mean, this is what numbers are for. And not only do numbers indicate ordering(without having to do a mapping through severely atrophied school science class knowledge from decades ago), they can also convey meaningful size information using the wonderful technology of units of measurement.
I refuse to walk among you rabble. You rabid mouth-breathers. I refuse I tell you.
The only models worth _my_ time are Σεληνη, Γη, Ἡλιος and Αστρα.
I don’t know what to tell you illiterate heretics other than you should know you are bad people. You should be ashamed of yourselves.
The truly αμεταπειστοι among us, the people _I_ hang out with, know these things by their proper name. Which is to say, naturally, 𒌗 and 𒆠 or, if you were even somewhat educated: 𓇹, 𓇾 and 𓇳 (praise be to Ra, the Great Unborn).
> Today, we’re rolling out Gemini 4 Hydrogen inside Google Antigravity alongside a new model dropdown option called Orchestrator, which pairs Hydrogen with our strongest models for long-horizon tasks, including Gemini 4 Carbon, in an optimized balance of quality, efficiency, and latency.
Before I go and burn through my Pro tokens in four minutes again (looking at you, Argon): is this something usable with a mere Pro account? Or will this require Ultimate for agy?
"Gemini 4 Hydrogen" shows up as something I can select in antigravity CLI, while "Gemini 4 Carbon" is "Unavailable · Gemini 4 Carbon is currently available to Ultra plan subscribers."
Also 3.8 flash is still visible for me (contrary to what a previous poster experienced.)
hear me out, finance: two hydrogens, but we call one carbon and allow it to do five more things than hydrogen can. and you can't do the pair-bonded mode for less than $100. by golly it just might work
Amen, what the heck—after double-tapping to zoom and center the text on this page in Safari, scrolling down with a 2-finger touch gesture...also scrolls to the left? I struggle to understand how or why one would ever mess with such behavior.
Terrible naming, unless the atomic weight is some indication of size? When is Uranium-235 coming out? Not sure what this model is, is Hydrogen like Sonnet where Carbon is like Opus?
Is google playing chemistry? First they launch Carbon, then hydrogen now next will be oxygen. Are they making models around periodic table of elements?
But a new Flash model is very welcome. Argon and Carbon are very slow. In the time it takes for them to complete a single task, Opus already managed to complete multiple things. They’re so slow, I started using 3.8 Flash over them. (A lot of it is also because of how terrible agy is. Before doing any task, Gemini feels the need to hover up a bazillion files into its context when the task is something simple like “update the docs of this controller”)
Also this sounds like Argon is dead now
Is Carbon actually released yet?
https://blog.google/innovation-and-ai/models-and-research/ge...
EDIT: /s
Separately, I'm surprised not to see any HN link for this (that made it to the front page) in the past few days.
What I don't understand is how they all make it work with caching in general. Looking at how the advisor mode in claude work you seem to only be billed the output tokens and cache hit on the whole session when it calls out to the advisor, making it super cheap compared to running the more expensive model for the entire session.
Orchestrator seems to automatically delegate things to subagents as required is my read of it. This should be fine for caching.
When launching a swarm of smaller models, they can theoretically share cache amongst themselves, but not with the larger model (that already has it cached). They have different architectures.
Further, it wouldn't really make sense for them to share the same cache anyway! The swarm, that is, they'd be working on independent things...
Anyways when a new model is released they want to have some interest and hopefully some tyre kickers will stick and hoping some social medial coverage
Really
I set up my subagents with specific models based on how complex I believe the problem space is. I'm nervous to delegate that authority to a model (since models tend to be... optimistic about their capabilities.)
That said, not every coding task is a difficult coding task. Plenty are boilerplate, plumbing, or simple test authoring that could use a flash model, so maybe this would downgrade nicely.
On the other hand, it's trivial for me to set up some guidelines to create different subagents for different task domains, and just pin those models to those tasks -- implementor (hard) and implementor (easy) or whatever.
I’m glad Google is kinda back, but today there’s no reason to code with Gemini.
My partner runs an internal computational mechanics benchmark that hasn’t tracked progress with model advancements (flat/no progress since early 2026) until suddenly saturating from argon and carbon. My take is that it’s due to spatial reasoning (most of the challenges are programming that require a lot of shape rotation type thinking), but we don’t know just yet why Gemini is doing so much better.
Carbon is the first "better than meh" model form Google in more than half a year. It's been out for less than a week. How can you consider it "steady"? Their subscription and even the API has been terribly unstable throughout the past few days, so I'm having really trouble believing any of this.
Ok, you're either a troll or you objectively don't know what you're talking about. Google should be ashamed they even released 3.8 Flash. They published models with such high frequency and always made a huge jump on benchmarks. Then they were caught, when a new version of AA came out and their numbers dropped like a stone.
https://maxxbench.unlap.ai/
Gemini 3.8 Flash is the most benchmaxxed model to date and even Argon hallucinated so much, it was impossible to use it for any professional work.
They’re also excellent (together with Flash-Lite) to use for consumer facing features where speed and cost matters
https://blog.slopcop.com/a-cheapskates-guide-to-coding-with-...
Intriguing, I might have to try this out - I have an OpenCode Go sub for the deepseeks of the world, and a Claude sub too. Sounds like I could have the best of both worlds here.
scenario 2 DSF drives, but gets to ask Opus for advice
Is this now how e.g. Opus works now? I use Opus 5.5 on high for just about everything now, and my bill's not gone up (gone down actually). I had assumed under the hood it was doing this (e.g. I'm sure its not literally Opus doing the log checking / file searches). Towards Google's choice, I'm skeptical introducing a new concept (Orchestrator) for this is really necessary, when they could have just done it by default, unless its plainly lower quality than pinning Carbon.
Ignore all other models.
they are same base model
Luna, means "moon", smallest.
Terra, means "earth", bigger than the moon.
Sol, means "sun", the biggest object in the solar system.
Astra, plural of the word "astrum" which means "star".
Anyone who gets confused should be embarrassed.
LLM's really destroyed the average person's ability to read and comprehend text.
I mean, this is what numbers are for. And not only do numbers indicate ordering(without having to do a mapping through severely atrophied school science class knowledge from decades ago), they can also convey meaningful size information using the wonderful technology of units of measurement.
The only models worth _my_ time are Σεληνη, Γη, Ἡλιος and Αστρα.
I don’t know what to tell you illiterate heretics other than you should know you are bad people. You should be ashamed of yourselves.
The truly αμεταπειστοι among us, the people _I_ hang out with, know these things by their proper name. Which is to say, naturally, 𒌗 and 𒆠 or, if you were even somewhat educated: 𓇹, 𓇾 and 𓇳 (praise be to Ra, the Great Unborn).
It saddens me to see HN has fallen to this level.
Wow, that's a lot of new words
Also 3.8 flash is still visible for me (contrary to what a previous poster experienced.)
I have the "Google AI Pro" plan.
ps. it's freaking not