Gemini 4 Hydrogen and Orchestrator
antigravity.google
antigravity.google
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...
This is a lot of purchasing complexity for an old retired guy like myself!
I wish Google had finer grained $30/month and $40/month subs: then I would probably just make my life easier by only using local models and Google (their sub also provides some Opus, but the Anthropic subscription is better if all you care about is Opus or Fable.)
My only gripe is that it feels like they get slower and slower with time.
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.
One other thing I wondered about: is all that additional and possibly irrelevant information harmful? I believe that it is. I think it causes the model to make bad decisions. In long workflows, once a bad decision is made, I find it frequently compromises the rest of the implementation.
So I try to keep it tightly scoped. Otherwise I get hilariously bad results sometimes. Ex: yesterday I let Argon medium run for 40 minutes. Welp, that was a big revert.
Like they had really fast hardware or something.
E.g. something like Haiku is "obviously" faster than Opus. Yet if you test Opus 5.5 on medium vs Haiku on max (because Haiku needs to be cranked pretty high on effort to be reasonably intelligent), you'll observe that Opus is ~2x faster to complete.
In fairness the plan came with my phone, which is the only reason I have it, but the limitation is a nice reminder of why it'll be getting canceled the second it'd actually charge me a penny I haven't already spent.
At least, if you know your periodic table, there is a link to the size of the elements where hydrogen is small and argon is big. Just wondering if the new flash-lite will become helium lol.
Luna, Terra, and Sol are in order of celestial size, and Haiku, Sonnet, and Opus are in order of length. But Hydrogen, Argon, and Carbon? Maybe they intended the order to be by density in pure form.
Yeah, I noticed this with Argon also. It did a good job on my tasks, but the absurd amount of reading it needed to do first seemed really weird, relative to other models.
It also felt more robotic than Claude or any of the other models (GLM-5.3 is basically an excited teenager, while claude is a teacher's pet (sorry Dave, I can't let you do that)).
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...
I have absolutely no clue how these things works, but could it be that all the newer models are actually just the same model, with looped transformers, so it's just about how many passes it does internally that decides if it's Haiku / Sonnet / Opus / Fable?
I might have misread the article a little bit, as you say, this is just workflow orchestration with subagents. Anthropic seems to think that this isn't the best approach anymore with their recent posts around subagent use and with the experimental advisor mode. So who knows.
I've accepted Google is just bad at names.
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.
The problem is when 3.8 or Argon were reviewing things with context poisoned by what they were working with, as gemini models generally feel over optimistic about things.
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.
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
Using 3.x flash series heavily over last 6 months, and along with experiences on Argon and Carbon (haven't yet played much with Hydrogen, but initial experience looks great too)
I have not experienced the hallucinations or benchmaxxing people are complaining about. People keep banging on about how the hallucinations are so terrible that it cannot be used for anything etc, yet benchmarks like https://artificialanalysis.ai/evaluations/omniscience#aa-omn... show the top two best models for not hallucinating are Gemini models.
Something does not align with these complaints from people who are trashing the Gemini models - one has to wonder if they just tried it for 5 mins then give up because it doesn't behave exactly like Claude does.
As mentioned I used 3.8 Flash for a lot of things because I don't like working with Claude all day (its generally a sanctimonious prick to work with) and I way prefer working with Gemini's personality. 3.8 Flash seemed to me to be a really decent model for every day run of the mill get-shit-done work. True I have not tried to cure cancer or solve navier-stokes recently, but for your standard day to day coding works its been really solid.
I am excited to try 4 Hydrogen now.
First 3 were GPT-6 Astra, GPT-6.1 Sol, and Fable 5.2. And it did so at about 1/3 the cost of Astra and Fable (but still more expensive than Sol)
It's also an extraordinary improvement over Gemma 4 from just 6 months ago.
It looks like Gemma per token is like 1/20th of the cost anyway
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
Safeguards around any binary decompiling, disassembly, device level verification, etc. Static work only. No way to get cyber access.
Sticking with Codex and Daybreak Blue for now + Muse and Devin for lighter work.
Sounds painful.
The Bitcoin Red Team had great success in cinching up a lot of issues relying more than anything on Kimi K3. Claude may technically have an edge but when everything you ask gets you dropped to a lower model and still refused, it doesn't really help much.
Many pointed out that even if you've just built something with Claude, asking it to review for security holes got you looked at like you were a North Korean superhacker and treated accordingly.
I'll pass on that sort of disrespect from companies I patronize.
This when antigravity already has a pretty solid "auto" mode.
I suspect we will see people bitten by this.
They could have eaten Anthropic and OpenAI's lunch with the models they have and with the right strategy!
Not sure how brain dead they have to be when they don't even promote or offer any of their models at discount/freemium with schools, universities, community colleges, public libraries or other educational institutions.
Execs are too scared to take risks to remain relevant. Since its easier to lose taking big risks when a nice yearly bonus will roll in with a safer but less relevant options.
Issue is that they will obsolete themselves in no time if they are this scared to dominate a landscape that they ruled for 20 years.
I have not seen Argon successfully complete a medium, well-planned refactor. The behavior is very unpredictable. Yes, sometimes it one shots amazing things but most of the time does not do a good job on anything but planning and/or simple changes.
That said, I think the underlying models are great. I can't believe that an org with Google scale cash leaves Antigravity to rot. Whoever is leading that project needs to be fired immediately.
Also what's with the comically large 'Antigravity' in the footer. It's humongous.
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.
My own implementation workflow is orchestrator based, but I generally dont use the small models because they're fucking morons
Ignore all other models.
And argon? Are we going on atomic number or is it the noblest and thus cleverest, or is it the most chilled out model that doesn’t over react
From the company that brought us Android ice cream sandwich though, I guess these are their clearest names yet!
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.
I wonder how many billions in revenue they lost at this point from just offering a piss-poor developer experience.
ps. it's freaking not