It's insane how Google has dropped the ball almost completely with AI.
I literally don't care about their models anymore because I can't even try them without paying for them?
I already pay for the plus plan but whatever their latest model was, is locked for me. Or otherwise it's just not accessible for me.
If I try to add balance to my Google payment account, it shows me errors and is asking me to verify my identity through government documents.
---
Not to mention the extremely confusing terminology. "Flash", "Experimental". Why are most of your models either of those two? I don't get the confidence that your models are any good because of such names. Give me a normal reliable model!
Even if Google releases AGI tier models in the future, I'll always look for other models because Google clearly thinks their models are too good for any of us.
AI Studio is legitimately the worst UX experience on the internet. For a company with unlimited cash and the ability to hire the best talent, I just don’t get it. Give nearly anyone else a few hours alone with that codebase and Opus-5.5 and it will be looking 1,000x better.
> "AI Studio is legitimately the worst UX experience on the internet."
I find it hard to believe there aren't worse ones out there. But off the top of my head I actually can't think of any...
Quite incredible that you can say a sentence like that, while talking about something within Google(/Alphabet) that's actually an important area for them (not just some random little side project), and yet it isn't hyperbolic, it's sincere!
Everyone there overseeing it (from direct management all the way up to C-suite) should truly feel embarrassed.
I wanted to pay them money to try their things. I just literally couldn't figure out how to do it. I think I then managed to get some trial credits but I never managed to activate anything with them and hit API limit after basically one prompt.
Our team figured out the only stable way to do this was to either use OpenRouter for API, or for subscriptions, to get Ultra Gemini accounts on personal Gmail accounts (no way to do this for individual members of an org without upgrading the entire org!) and hook those to the Antigravity CLI which Claude Code can then orchestrate for second opinions.
Which are amazing! Argon’s insights on code review and copywriting are phenomenal, and make these hoops well worth it! But geez, it’s horrifying that we had to figure this out!
Are you sure about Argon? My experience has been it seems to understand code but it completely destroyed a couple of projects I was working on. Others like Theo used it to work on T3 Code and said it benchmaxxed and was atrocious.
We use it primarily in the context of a code reviewer, and it occasionally finds subtle security/workflow/backwards-compat issues that Claude and GPT do not. Claude makes final judgments on whether the findings are material and how they should be fixed.
Argon also tends to write particularly warm copy and external messages - unsurprising with Google's particular access to email data and caches of writing from the early internet.
We haven't really used it to write code from scratch - generally, we're wary of these exact things, and feel like Claude continues to have good enough architectural instincts that we've doubled down into the Claude ecosystem, with mods and tooling (including a multi-agent reviewer) that accelerate our workflows.
A year ago I was trying their enterprise code offering and for the life of me I could not get it let me use a gemini 3+ model. I enabled everything in like 5 different interfaces and the GeminiCLI and UIs would only show me 2.5.
Eventually I gave up. Anthropic/OpenAI took < 10 minutes to get things going.
I paid for their AI models on the app because I wanted to try the Gemini experience on my device but its utterly incohesive, I just cant use my subscription anywhere else. Anthropic is close but at least they started offering API credits recently so my workflows dont break, but with google man.... its a horrible experience. Won't renew, thats for sure.
Me too! I wanted to just make an api call to get my current spend on google models.. and like 50 steps of creating this and that and linking it to here and then giving it this and that permissions, and being told this isn't acruall real time, you've just set up a query to be run against your data into this new data lake, and come back and check tomorrow.. I still couldn't get anything to work..
> AI Studio is legitimately the worst UX experience
Their backend looks very confused too. I consume their models through OpenRouter and requests are processed randomly by AI Studio or Vertex. API key management had several entry points for a while too, with the option to query the same model with a GCP, Firebase and/or AI Studio API key.
Yeah Tibo has a tweet where he talks about how he tried to do stuff at Google and they just wouldn't let him. Doesn't matter how much $ and talent you have if you are actively sandbagging them (and of course, for truly great stuff, you want the whole org from the very top driving you on, a la Musk or Jobs).
When I signed a few devs at work up for Gemini for Code I had to go into Google Cloud, jump through several different screens, assign special permissions to each dev to make it work and even when it did work...it was insanely confusing compared to every other option out there.
It's to the point where I'm better off asking Grok Bot set it up for me rather than ever deal with figuring it out again.
Is this true anymore? I feel like a lot of the best talent have better opportunities elsewhere where they won't get bogged down by entrenched bureaucracy. Not to mention the Google interview process and physical locations filter out a lot of talented people from even wanting to apply.
I love when my kid and and I are using AI Studio to make tower defense games, and it tries to squeeze in reasons to make Gemini API calls. No, I don't want a Gemini-populated random news ticker in my tower defense game.
If you have ever worked for a large company you can quickly imagine why they cant execute. Google is old now and entrenched. We will see how long it takes them to become IBM/GE etc. (if they are not already headed that direction).
Do they still have 5 messaging apps or did they finally trim it down.
No, it's exactly because of the copious resources.
Vertex AI existed as an ML platform before the GenAI era. DeepMind was doing it's thing in another silo. Then Gemini products came along. Then they acquired the Windsurf folks.
A smaller company would have no choice but to focus on one team and one product. Google has a Darwinian ecosystem and someone has to come along and smash teams together, clumsily refactoring the technology and user journeys.
Their employees read HN though and their VPs hear customers complain about onboarding. I hear there is an effort underway to create one unified agentic platform.
I've worked in some pretty big global orgs. Everything was decided by committee which meant that everything had to work for everybody, which meant that everything was shit. I spent some of the best years of my life building and rebuilding the same bullshit. I think I shipped a handful of truly novel experiences in that time (probably about one novel feature every 2 years) but the rest was beating a dead horse to make the experiences we already had "better."
Eng was neutered at that large org. Product people had the "vision" and the say. A lot of product people just used our team's product as a stepping stone onto their next role and in the end it showed. I would hope Google has a better culture...but theres a good chance it doesn't.
Google has an near monopoly on Internet ads (or rather a duo-poly combined with Facebook). Having such an monopoly makes them have hindsights on what they are doing and how they are failing, or even if they do, the incentives and management just aren't cleanly aligned with what ought to be the case.
Google has always been a bunker corporation, they exist only on shielded campuses and never interface with normal people or reality. One of the additional consequences of never concerning themselves with providing customer service for most of their products. They have no idea what's actually going on on the ground, until long after it has shaken the world (ChatGPT). They're deeply isolated from the front and from proper feedback iteration mechanisms.
There have been months where we've paid 1500€ for my AI models, but I can't bother with Google's models because I literally don't know how to get them. Every time I look it up it looks incredibly cumbersome and confusing. It's not clear which models I get, if I can get them on my personal accounts or need the work account. It's just too confusing.
I am a Google Workspace customer by the way (have multiple accounts, actually!).
Google must really not care about me as a customer, and instead only care about the enterprise.
I'm quite happy using antigravity fwiw. I use Claude code for work. Only tried agy because I was already paying for Google one (youtube music, youtube without ads/premium and a few TB of storage).
Antigravity come with Google's models and anthropic - so Gemini 4 Argon, Gemini 3.8 Flash, Claude Opus 5.5
Claude Sonnet 5.5 and
GPT-OSS 120B.
Agy is better than Claude code in some ways and vice versa.
Google models are great.
Ironically google music is so terrible I've uninstalled it and use Metrolist exclusively.
I think google models do a decent job helping me put together a blog, for example:
In the past, whenever benchmarks used to show google as being on par with the rest, I tried them and found them extremely lacking.
It's been maybe 6 months since I last tried them in any proper fashion (other than my usual benchmark which is non-coding related), but I was so let down previously, and found it so exhausting to just figure out how to get a subscription working with them, that I really feel like I've given up.
I'm sure I'll get around to trying them. Heck, it could happen today if the mood hits, but....it does need a specific mood. The mood to get over bullshit UX and a profound disdain for the company.
I've not been impressed with their models. Argon seems good but the usage limits on Pro feel worse than Sol/Astra/Opus on competing plans. But where Luna and Haiku have gotten to point of being 'good enough' for grunt work, I don't feel like 3.8 Flash is there.
That view somewhat biased by Flash getting stuck looping in Antigravity's CLI when I asked if it could add plan usage % to the status line. Instead of "no", it first pretended agy was Claude Code and gave me a wrong answer, then resorted to trying to analyse the binary.
The only sane way is to do it through openrouter. I spent hours trying to figure out how to use one of their two competing APIs that expose the models natively, it ended up being a waste of time.
Yes, I do that to do quick tests of the model, but I am looking for the subsidized monthly subscription, which everyone else has. I won't use it otherwise.
For unknown reasons workspace accounts get ridiculously slow roll outs of new models.
The only way to get the latest models with decent limits on a subsidized plan is to pay the full cost (monthly or yearly) for the pro plan or higher on a personal account.
Promo "pro" accounts and discounted "pro" accounts get much worse limits and model selection.
If that last is accurate, that seems like a terrible choice. I signed up for a 3 month discounted Pro trial to see if Gemini was worth trying against Claude and ChatGPT, then ran into an hour long lockout after my first serious prompt to Argon. Not a great first impression.
Would want assurances they won’t permaban a Gmail for a false positive for something on their LLM offering. Like the risk with Google Photos (thinking about the dad sending a doctor a photo of his son’s rash).
Thanks, I didn't know about this page, and I'm a Google Workspaces subscriber.
By all rights, people should be debating among Codex, Claude, and Gemini plans. They obv don't care about that mindshare but they're a big enough organization that they could carve off a marketing and dev outreach team.
On X the only Google person I follow is Richard Seroter, and he hypes Antigravity sometimes, but not the plans you need for it. No one ever talks about it.
They should be the 3rd or 4th, certainly jockeying for position with Grok/Cursor. I would never sign up for an Elon AI plan under any circumstances, and there's enough distaste for Anthropic and OpenAI to fuel a few DCs, and Google isn't even playing.
Love that they have "Personal", "Business" and ... "Developer"
Perfect Venn Diagram.
Thanks for the links. I'm pretty sure last time I tried this I got mysterious errors and a couple of redirect loops. Maybe this time will be the one that works.
It's confusing for enterprises, too. Believe it or not, the cloud sales team recommends hiring partners to set up basic auth and "landing zones" for the APIs.
These are the (approximately) equivalent model classes between the labs
Experimental means not production, so liable to change in the near future if you decide to build on it.
Omni is a video editing model, and flash implies it's a lighter weight one than what they have/could have. The other labs don't have competing models here.
What's Carbon then? I haven't been paying much attention to Google but their naming has been confusing me.
Luna/Terra/Sol/Astra was a bit pretentious but at least the sizing makes sense. (I imagined the meeting... "Ok guys, we're getting mogged on naming... what's more epic than a Myth?" -"The SUN!")
Anthropic makes sense with the literary "scale" naming, mostly.
Is a Carbon bigger than an Argon though? Maybe it has to do with the atomic number?
This would be by (suspected) parameter size, not benchmarked ability.
It's possible Argon/Carbon are pro sized, but given the internal rumblings and that it is based on a fresh pretrain, it makes most sense for it to be a "big" model.
Google did at one point have an "Ultra" sized model, but nothing more came of it after 1.5 Ultra.
People really seem to struggle with this and I can't figure out why... Like it kind of just reads like astroturfing in every thread. I downloaded agy, purchased a subscription that the splash page said "includes Argon" and it... works?
3.8 flash and argon have been great FYI. Can't tell a difference between any of the claude models I use at work, minus actually being able to read what it writes.
> People really seem to struggle with this and I can't figure out why... Like it kind of just reads like astroturfing in every thread. I downloaded agy, purchased a subscription that the splash page said "includes Argon" and it... works?
Google does tend to have terrible messaging/branding/etc that results in part of these problems.
When Argon first came out, it wasn't included in Antigravity/agy and I personally never saw a Google stated plan on when they would include it, and because they removed Gemini Pro when Argon was released there was a brief window where Antigravity/agy didn't support a "Pro" model at all even for people with a "Pro" subscription.
That said, I'm with you on the idea that Google is roughly on par with everyone else despite the rhetoric here on HN that makes it look like they are way behind. They do have typical Google problems with messaging/branding for their AI products, just like all their other products, but when I use Google models for personal use I don't feel like I'm missing out on anything I get when using Anthropic for work stuff (paid for by the company).
And for context I say this all as someone who is a bit of a Google hater these days (not happy with some recent Android decisions, moved off of gmail to avoid too much Google power concentration, etc).
I can tell you that I have a Gemini plus subscription I still don't have anything more recent than 3.6 Flash. In consequence, I never use Gemini anymore, and instead have a PAYG plan on Deep Infra.
The search model is not the same model, and it's quite obviously worse when you use it, so that's not quite a reason I would personally find the hostility excessive.
This sentiment is the top comment I'm seeing on every recent Google release. That it's impossible to actually access the tech, or even pay for it.
I heard someone say on the last such thread that you can use their models thru OpenRouter, and that this is much more pleasant than trying to use them thru Google directly. Looks like it's up there already:
That, and gemini (app or web) has been bugging for months for me. I'd get "something wrong happened" every 3-4 queries, sometimes a few retries made it work, sometimes not. That this happens for a day is understandable, that this is unsolved after weeks or months is incomprehensible to me and literally forced me to a competitor (which I immediately found to be better anyway)
I agree and I'm guessing the reason is that Google's real incentive is using their own AI in search and their internal products. While they should care about other developers building on top of them (via Google Cloud), they have an even bigger customer which is their internal teams.
I paid, ran antigravity cli, getting 500 on the image tool because the model the cli wants to use is not available?
There are no other models to choose. Im so confused
I just use their models via openrouter. I have an account with OpenAI, Ollama , even deepseek platform. But for all google ai stuff, I just use openrouter.
Not only is it easy to pay but I also don’t risk getting my account banned.
I can't use it. The webpage reloads continuously after logging in. It's probably some ad blocker add-on, but it's the only web AI that doesn't work, so I simply don't use it at all.
The real problem is you can't even try them with paying for them! It's been more than a week since their previous model was announced, "Gemini 4 Carbon", if there is such a thing. Even paying customers still can't use it!
On top of that, the inference is painfully slow. I’m now convinced Gemini is no better than ChatGPT or Claude. They haven’t even managed fast responses on the Flash model, and I’m paying for Pro. Gemini, you suck.
Alphabet is a company with too much money and resources, which means they don't end up trying very hard to acquire and retain customers for any of these products since it's just a blip in their overall revenue. It's time to split them up.
They may have dropped the ball on AI-for-programmers, but the Google slopbot is the one the normies use. As much as their AI summaries used to be the biggest joke in the industry, today they're the one I think normies who aren't "AI people" are using the most day-to-day. I see a lot of people doing it:
Search with Google, see the AI summary and the search results, realise they want more detailed info, and put that request for clarification into the AI reply and now they're off to the races on a Google AI conversation.
Now, it's questionable how "sticky" that is and latest financials revealed it was extremely unprofitable, but in terms of raw "who's using AI" it seems to be working for them.
Bingo, this is exactly how I use AI personally. There’s just too much SEO spam to wade through on search results anymore, and Google’s AI Mode RAG is pretty good at citing trustworthy sources when I want to dive in.
DuckDuckGo’s equivalent feature is unfortunately too prone to hallucination, especially whatever cheap model they are using for AI summaries. Once you click in it switches to GPT Luna, which will sometimes hilariously correct the previous response.
Yeah, I try to use DDG as my "default" search but it's just all around mediocre compared to Google. Even simple dumb things like "image search for a meme I remember that I want to post on bsky": Google has it near top result, DDG doesn't even have it on the first SRPs for the same query.
how i have been using it 'at home'. its ok conversationally. It honestly is a very 'mid' search engine. Sometimes it can zoom right onto what I want. Other times it gets very fixated on the 'wrong thing' and you have to club it into expanding its view. It just randomly gets 'lost' and you have to nudge it along to actually dig in. even then even if you explicitly tell it to review its sources it will just kinda ignore the request and keep going.
It truly is. I acutallu signed up for an annual plan because I used the image models for design ideas on my house. Few weeks ago they announced changes, no more pro models. It’s laughable how much they keep screwing up.
This is a new video model. It is not a replacement for Gemini 4, it is doing something altogether different.
It has been available, since yesterday, within Google Flow. I know, because I used it yesterday. Creatives normally are using Flow or similar third party tools, not APIs or AI Studio. The claim here is that there is more character consistency, which is helpful if true. But that's not something I can see from a little work. Despite the "More Affordable" claim, it costs more, and the outputs doesn't, from my limited experience so far, seem to be significantly better. So I could only, personally, justify pulling this up when character consistency is punching me in the face. Which, to be fair, is a lot of the time, both in my experience, and the experience of other AI film makers I talk to.
This isn't for (most) coders, it hardly matters if it shows up on Flow a little earlier that Google Vids, and, perhaps, it is a little silly to act like it is a nightmare that a rollout occurs over a few days, especially the rollout prioritizes the paying creatives on the tool that are most likely to use.
The rollout is so confusing. Going straight to the Gemini app and asking when I can use this model gets me vague and contradicting answers. Ultimately it sounds like the Vids rollout will take 1-3 days, and you can go to vids.new on desktop to see if you’ve won the lottery and have it available now, apparently on any free Google account. Business accounts can use it too. Meanwhile the Gemini app is getting it on paid plans, but with different video length limits and no clear rollout timeline. If you pay for the API, it’s available now apparently.
I think they are doing a terrible job with their product packaging. Like others, I also gave up with Google models long ago due to difficulty in accessing and low limits.
My gut is that this is reflective of their fragmented internal product organization around AI.
It's okay to say that Google's engineering is also poor. If they had good engineering, they'd study their users and lead to product improvements. It's really not either/or.
It has been a while, but I used to read AOSP source code quite regularly, and always wondered where do all those engineers that pass the famous Google interviews land, because they certainly weren't writing the Java and C++ code in Android, given how it looked like, or the whole mess with NDK headers, and other stuff in the tooling.
Ah, and the Eclipse/Android Studio stable releases, that usually have bugs as if they were still preview, release candidate ones.
Still can't handle the prompts I throw at it, though the 15 seconds gives more breathing room. Not being able to use first and last frames for the longer videos doesn't render it useless, but it gets pretty close (the model just cannot reproduce unusual or idiosyncratic character designs without first and last frames; fat, muscular, ugly, and non-human characters regress to tv commercial blandness).
I just don't know how video is going to make it, long-term. You have to go through so many iterations to get something with even a semblance of what you were going for. I've been making AI videos since early 2024, and while image quality and coherence has gotten better, the ability of the models to actually follow your directions precisely hasn't. I would rather have to wait 5-10 minutes for a clip that actually does what I tell it to than get something useless in 30 seconds.
When playback ends on one of the videos, the page jumps back to that video for some reason. I saw this too when I had multiple videos playing at once and the earlier clip on the page finished while watching another one lower on the page.
Completely random, but I was inspecting the source code for their Omni 1.2 Flash web page, and there are insane amount of newlines. There are seven newlines between the first line and the DOCTYPE. There are huge amount of newlines between each section. What did they use to generate this website?
Enterprise websites tend to have very complicated content generation systems with little care about performance (including html/css/js size). What's hypocritical is that Google penalizes in SERP sites which have poor performance score yet Google's own sites are far from exemplary.
they might have a composer layer which loads certain components (transparent to you) with and without authentication and mixing different backend sources.
Then they might have templating engines and frameworks.
Sometimes they just don't care but lets be honest: a few new lines are a few bytes.
But the most critical pages go through optimizers for traffic
> Today, we’re rolling out Gemini Omni 1.2 Flash, and making it generally available across the Gemini API and Gemini for business. Gemini Omni 1.2 Flash introduces native 15-second video creation, multi-turn conversational editing capabilities, and enhanced reference following. We have dropped the price 30% from Gemini Omni 1.1 making it available at $0.07 per second for 720p video.
I desperately wish Google would focus on improving the Gemini capabilities that are built-in to their office suite (docs, slides, etc.). The current built-in Gemini is terrible, and if I could use e.g. Claude natively with those products my workday would be an order of magnitude more productive.
I'm shocked that they don't seem to be investing in agentic capabilities there at all.
I built https://smalldocs.org / https://github.com/espressoplease/smalldocs basically for this. It lets any agent you want create rich documents - word, slides, spreadsheets, mermaid diagrams, rich code rendering, etc. About to launch a cloud first version of it, so what's on the site now will evolve soon.
What's this name and versioning now, NOBODY understands Google's Ai products, including subscribers to them.
Why their models go from 3 to 4 but then whatever this is starts at 1 I honestly don't care anymore
My friend on Google ai plus is still stuck on gemini 3.6 flash and now they're taking away 3.1 pro from them too even though it's long obsolete, like zero regard for user interactions here
honestly the naming on all of them from all companies is weird unless you 'know where to look'. I get having a cool code name for each one. But give me a clear way to look at it and see why should I use one model over another? like 'low/medium/high' something?! please!
That is sad... I tried to guess and ignore AI videos (created to be deliberately misleading as "realistic") by length, but I guess that time is over now.
It's amazing to me how poor adoption has been for video models, at least in my corner of the internet. I see considerably more Code to Video content from libraries like HyperFrames than any native AI video. It seems that users are much more open to AI coded "Explained style" videos than anything else.
It's still pretty inconsistent, slow, and expensive. I have claude write up massive prompts with reference images and it'll still come out poorly, but since I'm not generating much revenue with the product I use these models for, I just accept a lot of the sloppiness. I have tried a lot of different prompts, I think video just has a lot more free space to differ from our mental models than something like code that is a lot more well-defined.
Okay, I just wanted to vent: I'm on the Pro plan and I've not been able to use Argon for the past 4 days - every time I send a prompt I get downgraded to a Flash model.
Maybe I’m using antigravity wrong, so I’d love to be corrected here:
* For scientific computing it constantly cuts corners and writes awful buggy code. Bugs that will flip signs, or clamp values. Then when the model (or harness..?) discovers it, rather than fixing the underlying hacks it just adds more code with ifs, cases, and deceptive tests.
* using argon on high by itself is fast enough but seems dumb. /teamwork will catch a dozen errors from High’s laziness and sloppy tendencies
* /boost and /teamwork are ridiculously slow. Much slower than Codex Ultra. And it’s not clear when to use it.
I basically lost all trust in Argon output after running verification through “/teamwork Argon High” or “Sol 6.1 Ultra”.
It would be incredibly helpful to me if AI model names followed cars.
I'm not scratching my head about a "2025 Corolla." Windows 95 and 2000 had this built in, but then they lost it.
I'd prefer it was called Omni 2026.2 instead. It embeds a date and model and would package it as a product for on-premise or local hosting. It'd be easier to sell to the bean counters in the future: "We're running Omni 2026.2, we want to upgrade to Omni 2028."
They aren't far from cars. Cars have generations and each generation has distinct models, for market (region), body style, maybe generational steps with facelift, etc. So your "2005 Corolla" is Corolla Gen9 (2000-2007), designated E120/E130 (no idea how they differ), and can be 3-door/5-door hatchbacks, 4-door sedan, station wagon, varying which market you look at.
But it still starts with “Gemini”. It would be better if they called it Gemini Omni 4, 4.1 etc. Otherwise call it just Omni 1.2 and drop the Gemini prefix.
Not going to even read this post. While I have been really bullish on Google I am now tired of them constantly baiting us with vaporware announcements unfortunately.
TFA is literally a marketing piece for why someone might want to use it. If you personally don't need it, that's ok, you're just not their target audience.
Hot take, Google should focus on their open weight models. I use Qwen 3.8 and it’s been amazing. I would pay for Max but there’s no provider or server here in the states. If they had an amazing Gemma 5, I would seriously consider trying their closed source models. Model comfort and familiarity is a real thing.
The average length of a shot in full length feature film today is like three seconds. Even back in the old days, the average shot was like 12 seconds. Most movies are made up sequences and scenes made up of lots and lots of shots.
Increasing the available range of shot lengths from 10 seconds to 15 seconds gives film makers a wider pallet of options, if the longer shots have continuity, which is the other claim here.
I'm not aware of any "one-shot", one and done, use cases that are much more compelling the clip is 15s vs 10s. But that's not relevant to many people actually using AI video models like this on a daily basis.
That's cool, but I don't see why one would want to watch a movie like that.
For B-Roll it's different. I can somewhat see the why behind that.
It's just to give your brain something to look at so that you don't have a voice-only podcast.
But movies OTOH are these creative endeavors where the shots _are_ the content instead of just augmenting the content.
This won't stop the industry from still trying to use this tech like that, of course.
Imagine a hybrid movie, where film maker spends $5000 filming a couple of actors, but uses $50 of AI shots for special effect or other footage that would have been completely out of budget before.
I mean, if you just want to claim that AI lacks soul, and its an abomination if every frame in film isn't make by a chemical process involving silver nitrate, like the good Lord intended, that's fine. But there are certainly film makers who don't share that sentiment.
I can imagine that, but I don't think that this is the right AI tool for the job.
I would rather expect people to use AI to drive more conventional VFX tools, as that is reproducible, can be versioned, and, most importantly, will contain true intent in aspects.
it’s a shame that Google/Gemini decided to restrict and fuck other adventurous yet extremely gracious paying customers; who were excited at the prospect of not being locked in to two obnoxious companies.
oh well, I guess we’ll just have to take that enthusiasm and roll it over to all the open weights on providers like fireworks.ai. or others.
i’m so happy I got in early and paid for $200 a month for shit ass Opus 4.6 and sonnet 4.6 and being locked into the absolutely limited Gemini 3.8 flash.
It's insane how Google has dropped the ball almost completely with AI.
I literally don't care about their models anymore because I can't even try them without paying for them?
I already pay for the plus plan but whatever their latest model was, is locked for me. Or otherwise it's just not accessible for me.
If I try to add balance to my Google payment account, it shows me errors and is asking me to verify my identity through government documents.
---
Not to mention the extremely confusing terminology. "Flash", "Experimental". Why are most of your models either of those two? I don't get the confidence that your models are any good because of such names. Give me a normal reliable model!
Even if Google releases AGI tier models in the future, I'll always look for other models because Google clearly thinks their models are too good for any of us.
AI Studio is legitimately the worst UX experience on the internet. For a company with unlimited cash and the ability to hire the best talent, I just don’t get it. Give nearly anyone else a few hours alone with that codebase and Opus-5.5 and it will be looking 1,000x better.
> "AI Studio is legitimately the worst UX experience on the internet."
I find it hard to believe there aren't worse ones out there. But off the top of my head I actually can't think of any...
Quite incredible that you can say a sentence like that, while talking about something within Google(/Alphabet) that's actually an important area for them (not just some random little side project), and yet it isn't hyperbolic, it's sincere!
Everyone there overseeing it (from direct management all the way up to C-suite) should truly feel embarrassed.
The average SAP internal app is probably worse.
Off the top of my head: workaday
I wanted to pay them money to try their things. I just literally couldn't figure out how to do it. I think I then managed to get some trial credits but I never managed to activate anything with them and hit API limit after basically one prompt.
Our team figured out the only stable way to do this was to either use OpenRouter for API, or for subscriptions, to get Ultra Gemini accounts on personal Gmail accounts (no way to do this for individual members of an org without upgrading the entire org!) and hook those to the Antigravity CLI which Claude Code can then orchestrate for second opinions.
Which are amazing! Argon’s insights on code review and copywriting are phenomenal, and make these hoops well worth it! But geez, it’s horrifying that we had to figure this out!
Are you sure about Argon? My experience has been it seems to understand code but it completely destroyed a couple of projects I was working on. Others like Theo used it to work on T3 Code and said it benchmaxxed and was atrocious.
We use it primarily in the context of a code reviewer, and it occasionally finds subtle security/workflow/backwards-compat issues that Claude and GPT do not. Claude makes final judgments on whether the findings are material and how they should be fixed.
Argon also tends to write particularly warm copy and external messages - unsurprising with Google's particular access to email data and caches of writing from the early internet.
We haven't really used it to write code from scratch - generally, we're wary of these exact things, and feel like Claude continues to have good enough architectural instincts that we've doubled down into the Claude ecosystem, with mods and tooling (including a multi-agent reviewer) that accelerate our workflows.
A year ago I was trying their enterprise code offering and for the life of me I could not get it let me use a gemini 3+ model. I enabled everything in like 5 different interfaces and the GeminiCLI and UIs would only show me 2.5.
Eventually I gave up. Anthropic/OpenAI took < 10 minutes to get things going.
I paid for their AI models on the app because I wanted to try the Gemini experience on my device but its utterly incohesive, I just cant use my subscription anywhere else. Anthropic is close but at least they started offering API credits recently so my workflows dont break, but with google man.... its a horrible experience. Won't renew, thats for sure.
Me too! I wanted to just make an api call to get my current spend on google models.. and like 50 steps of creating this and that and linking it to here and then giving it this and that permissions, and being told this isn't acruall real time, you've just set up a query to be run against your data into this new data lake, and come back and check tomorrow.. I still couldn't get anything to work..
> AI Studio is legitimately the worst UX experience
Their backend looks very confused too. I consume their models through OpenRouter and requests are processed randomly by AI Studio or Vertex. API key management had several entry points for a while too, with the option to query the same model with a GCP, Firebase and/or AI Studio API key.
isn't the random choice of AI Studio or Vertex a feature of OpenRouter?
you could pin the provider to always get one or the other
Yeah Tibo has a tweet where he talks about how he tried to do stuff at Google and they just wouldn't let him. Doesn't matter how much $ and talent you have if you are actively sandbagging them (and of course, for truly great stuff, you want the whole org from the very top driving you on, a la Musk or Jobs).
https://x.com/thsottiaux/status/2083596911060324570?a=iif1W7...
Agreed.
When I signed a few devs at work up for Gemini for Code I had to go into Google Cloud, jump through several different screens, assign special permissions to each dev to make it work and even when it did work...it was insanely confusing compared to every other option out there.
It's to the point where I'm better off asking Grok Bot set it up for me rather than ever deal with figuring it out again.
> the ability to hire the best talent
Is this true anymore? I feel like a lot of the best talent have better opportunities elsewhere where they won't get bogged down by entrenched bureaucracy. Not to mention the Google interview process and physical locations filter out a lot of talented people from even wanting to apply.
What's wrong with it? I haven't used it, just curious to hear its faults
I love when my kid and and I are using AI Studio to make tower defense games, and it tries to squeeze in reasons to make Gemini API calls. No, I don't want a Gemini-populated random news ticker in my tower defense game.
Ironically, Gemini is excellent with UI/UX, and somehow they can't use it to improve themselves.
They're banned from using Opus-5.5 or other competitor models so they can't fix it.
Opus 5.5 and Fable 5.1 are available now
What about Workday? That's the standard in terms of horrible experience.
If you have ever worked for a large company you can quickly imagine why they cant execute. Google is old now and entrenched. We will see how long it takes them to become IBM/GE etc. (if they are not already headed that direction).
Do they still have 5 messaging apps or did they finally trim it down.
>For a company with unlimited cash and the ability to hire the best talent
Right. They lack no resources. Therefore, what is happening must be what they want to happen.
No, it's exactly because of the copious resources.
Vertex AI existed as an ML platform before the GenAI era. DeepMind was doing it's thing in another silo. Then Gemini products came along. Then they acquired the Windsurf folks.
A smaller company would have no choice but to focus on one team and one product. Google has a Darwinian ecosystem and someone has to come along and smash teams together, clumsily refactoring the technology and user journeys.
Their employees read HN though and their VPs hear customers complain about onboarding. I hear there is an effort underway to create one unified agentic platform.
I've worked in some pretty big global orgs. Everything was decided by committee which meant that everything had to work for everybody, which meant that everything was shit. I spent some of the best years of my life building and rebuilding the same bullshit. I think I shipped a handful of truly novel experiences in that time (probably about one novel feature every 2 years) but the rest was beating a dead horse to make the experiences we already had "better."
Eng was neutered at that large org. Product people had the "vision" and the say. A lot of product people just used our team's product as a stepping stone onto their next role and in the end it showed. I would hope Google has a better culture...but theres a good chance it doesn't.
Google has an near monopoly on Internet ads (or rather a duo-poly combined with Facebook). Having such an monopoly makes them have hindsights on what they are doing and how they are failing, or even if they do, the incentives and management just aren't cleanly aligned with what ought to be the case.
Google has always been a bunker corporation, they exist only on shielded campuses and never interface with normal people or reality. One of the additional consequences of never concerning themselves with providing customer service for most of their products. They have no idea what's actually going on on the ground, until long after it has shaken the world (ChatGPT). They're deeply isolated from the front and from proper feedback iteration mechanisms.
wow that is a really good way to describe it.
There have been months where we've paid 1500€ for my AI models, but I can't bother with Google's models because I literally don't know how to get them. Every time I look it up it looks incredibly cumbersome and confusing. It's not clear which models I get, if I can get them on my personal accounts or need the work account. It's just too confusing.
I am a Google Workspace customer by the way (have multiple accounts, actually!).
Google must really not care about me as a customer, and instead only care about the enterprise.
I'm quite happy using antigravity fwiw. I use Claude code for work. Only tried agy because I was already paying for Google one (youtube music, youtube without ads/premium and a few TB of storage).
Antigravity come with Google's models and anthropic - so Gemini 4 Argon, Gemini 3.8 Flash, Claude Opus 5.5 Claude Sonnet 5.5 and GPT-OSS 120B.
Agy is better than Claude code in some ways and vice versa.
Google models are great.
Ironically google music is so terrible I've uninstalled it and use Metrolist exclusively.
I think google models do a decent job helping me put together a blog, for example:
https://hypertekst.net
(Content is typical 90s homepage stuff, not great. But model voice isn't as grating as Claude - and while it has a "look" I think it's s bit better).
In the past, whenever benchmarks used to show google as being on par with the rest, I tried them and found them extremely lacking.
It's been maybe 6 months since I last tried them in any proper fashion (other than my usual benchmark which is non-coding related), but I was so let down previously, and found it so exhausting to just figure out how to get a subscription working with them, that I really feel like I've given up.
I'm sure I'll get around to trying them. Heck, it could happen today if the mood hits, but....it does need a specific mood. The mood to get over bullshit UX and a profound disdain for the company.
Fwiw the homepage serves as a current demo - for better or worse. Practically vibe coded (I haven't really looked at the code), but not one-shotted.
YMMV
I've not been impressed with their models. Argon seems good but the usage limits on Pro feel worse than Sol/Astra/Opus on competing plans. But where Luna and Haiku have gotten to point of being 'good enough' for grunt work, I don't feel like 3.8 Flash is there.
That view somewhat biased by Flash getting stuck looping in Antigravity's CLI when I asked if it could add plan usage % to the status line. Instead of "no", it first pretended agy was Claude Code and gave me a wrong answer, then resorted to trying to analyse the binary.
Their harness situation is just awful.
I paid and tried it today and the imagen tool just returns a 500, not sure if a fluke but not a great start
The only sane way is to do it through openrouter. I spent hours trying to figure out how to use one of their two competing APIs that expose the models natively, it ended up being a waste of time.
Yes, I do that to do quick tests of the model, but I am looking for the subsidized monthly subscription, which everyone else has. I won't use it otherwise.
For unknown reasons workspace accounts get ridiculously slow roll outs of new models.
The only way to get the latest models with decent limits on a subsidized plan is to pay the full cost (monthly or yearly) for the pro plan or higher on a personal account.
Promo "pro" accounts and discounted "pro" accounts get much worse limits and model selection.
If that last is accurate, that seems like a terrible choice. I signed up for a 3 month discounted Pro trial to see if Gemini was worth trying against Claude and ChatGPT, then ran into an hour long lockout after my first serious prompt to Argon. Not a great first impression.
> get them on my personal accounts
Would want assurances they won’t permaban a Gmail for a false positive for something on their LLM offering. Like the risk with Google Photos (thinking about the dad sending a doctor a photo of his son’s rash).
It's right here: https://gemini.google/subscriptions/
Have used them both for work with google workspace accts, and personally, with antigravity in linux.
Thanks, I didn't know about this page, and I'm a Google Workspaces subscriber.
By all rights, people should be debating among Codex, Claude, and Gemini plans. They obv don't care about that mindshare but they're a big enough organization that they could carve off a marketing and dev outreach team.
On X the only Google person I follow is Richard Seroter, and he hypes Antigravity sometimes, but not the plans you need for it. No one ever talks about it.
They should be the 3rd or 4th, certainly jockeying for position with Grok/Cursor. I would never sign up for an Elon AI plan under any circumstances, and there's enough distaste for Anthropic and OpenAI to fuel a few DCs, and Google isn't even playing.
Love that they have "Personal", "Business" and ... "Developer"
Perfect Venn Diagram.
Thanks for the links. I'm pretty sure last time I tried this I got mysterious errors and a couple of redirect loops. Maybe this time will be the one that works.
It's confusing for enterprises, too. Believe it or not, the cloud sales team recommends hiring partners to set up basic auth and "landing zones" for the APIs.
Flash-lite haiku luna
Flash sonnet tera
Pro Opus Sol
Argon Fable Astra
These are the (approximately) equivalent model classes between the labs
Experimental means not production, so liable to change in the near future if you decide to build on it.
Omni is a video editing model, and flash implies it's a lighter weight one than what they have/could have. The other labs don't have competing models here.
What's Carbon then? I haven't been paying much attention to Google but their naming has been confusing me.
Luna/Terra/Sol/Astra was a bit pretentious but at least the sizing makes sense. (I imagined the meeting... "Ok guys, we're getting mogged on naming... what's more epic than a Myth?" -"The SUN!")
Anthropic makes sense with the literary "scale" naming, mostly.
Is a Carbon bigger than an Argon though? Maybe it has to do with the atomic number?
Carbon is the next iteration (via RL training) of the Argon sized model.
A B (not released) C
Carbon is to Argon as Mythos is to Fable.
Carbon is to Aragon what Fable is to Mythos.
I thought carbon was smarter?
Yes, Carbon and Mythos are the secret super-smart models you can't use without an invitation and a million dollars.
Argon is not same tier as Astra and Fable. the upcoming, similarly unavailable Carbon maybe is.
This would be by (suspected) parameter size, not benchmarked ability.
It's possible Argon/Carbon are pro sized, but given the internal rumblings and that it is based on a fresh pretrain, it makes most sense for it to be a "big" model.
Google did at one point have an "Ultra" sized model, but nothing more came of it after 1.5 Ultra.
Argon is totally in those tiers in my experience. And you actually get to use it for a few days before you run out of tokens...
People really seem to struggle with this and I can't figure out why... Like it kind of just reads like astroturfing in every thread. I downloaded agy, purchased a subscription that the splash page said "includes Argon" and it... works?
3.8 flash and argon have been great FYI. Can't tell a difference between any of the claude models I use at work, minus actually being able to read what it writes.
> People really seem to struggle with this and I can't figure out why... Like it kind of just reads like astroturfing in every thread. I downloaded agy, purchased a subscription that the splash page said "includes Argon" and it... works?
Google does tend to have terrible messaging/branding/etc that results in part of these problems.
When Argon first came out, it wasn't included in Antigravity/agy and I personally never saw a Google stated plan on when they would include it, and because they removed Gemini Pro when Argon was released there was a brief window where Antigravity/agy didn't support a "Pro" model at all even for people with a "Pro" subscription.
That said, I'm with you on the idea that Google is roughly on par with everyone else despite the rhetoric here on HN that makes it look like they are way behind. They do have typical Google problems with messaging/branding for their AI products, just like all their other products, but when I use Google models for personal use I don't feel like I'm missing out on anything I get when using Anthropic for work stuff (paid for by the company).
And for context I say this all as someone who is a bit of a Google hater these days (not happy with some recent Android decisions, moved off of gmail to avoid too much Google power concentration, etc).
I can tell you that I have a Gemini plus subscription I still don't have anything more recent than 3.6 Flash. In consequence, I never use Gemini anymore, and instead have a PAYG plan on Deep Infra.
So, no, it is not astroturfing.
> I literally don't care about their models anymore because I can't even try them without paying for them?
Where are you finding ways to try e.g. Opus without paying for it?
well, search is free right? the hostility here is a little excessive
The search model is not the same model, and it's quite obviously worse when you use it, so that's not quite a reason I would personally find the hostility excessive.
This sentiment is the top comment I'm seeing on every recent Google release. That it's impossible to actually access the tech, or even pay for it.
I heard someone say on the last such thread that you can use their models thru OpenRouter, and that this is much more pleasant than trying to use them thru Google directly. Looks like it's up there already:
https://openrouter.ai/google/gemini-omni-1.2-flash
Also Vercel:
https://vercel.com/ai-gateway/models/gemini-omni-1.2-flash
If you're using any sort of router middle-man new models are usually easy to try.
That, and gemini (app or web) has been bugging for months for me. I'd get "something wrong happened" every 3-4 queries, sometimes a few retries made it work, sometimes not. That this happens for a day is understandable, that this is unsolved after weeks or months is incomprehensible to me and literally forced me to a competitor (which I immediately found to be better anyway)
I agree and I'm guessing the reason is that Google's real incentive is using their own AI in search and their internal products. While they should care about other developers building on top of them (via Google Cloud), they have an even bigger customer which is their internal teams.
I paid, ran antigravity cli, getting 500 on the image tool because the model the cli wants to use is not available? There are no other models to choose. Im so confused
I just use their models via openrouter. I have an account with OpenAI, Ollama , even deepseek platform. But for all google ai stuff, I just use openrouter.
Not only is it easy to pay but I also don’t risk getting my account banned.
I can't use it. The webpage reloads continuously after logging in. It's probably some ad blocker add-on, but it's the only web AI that doesn't work, so I simply don't use it at all.
> I can't even try them without paying for them?
The real problem is you can't even try them with paying for them! It's been more than a week since their previous model was announced, "Gemini 4 Carbon", if there is such a thing. Even paying customers still can't use it!
On top of that, the inference is painfully slow. I’m now convinced Gemini is no better than ChatGPT or Claude. They haven’t even managed fast responses on the Flash model, and I’m paying for Pro. Gemini, you suck.
Alphabet is a company with too much money and resources, which means they don't end up trying very hard to acquire and retain customers for any of these products since it's just a blip in their overall revenue. It's time to split them up.
> It's insane how Google has dropped the ball almost completely with AI.
Aside from winning the Nobel prize in chemistry.
They may have dropped the ball on AI-for-programmers, but the Google slopbot is the one the normies use. As much as their AI summaries used to be the biggest joke in the industry, today they're the one I think normies who aren't "AI people" are using the most day-to-day. I see a lot of people doing it:
Search with Google, see the AI summary and the search results, realise they want more detailed info, and put that request for clarification into the AI reply and now they're off to the races on a Google AI conversation.
Now, it's questionable how "sticky" that is and latest financials revealed it was extremely unprofitable, but in terms of raw "who's using AI" it seems to be working for them.
Bingo, this is exactly how I use AI personally. There’s just too much SEO spam to wade through on search results anymore, and Google’s AI Mode RAG is pretty good at citing trustworthy sources when I want to dive in.
DuckDuckGo’s equivalent feature is unfortunately too prone to hallucination, especially whatever cheap model they are using for AI summaries. Once you click in it switches to GPT Luna, which will sometimes hilariously correct the previous response.
Yeah, I try to use DDG as my "default" search but it's just all around mediocre compared to Google. Even simple dumb things like "image search for a meme I remember that I want to post on bsky": Google has it near top result, DDG doesn't even have it on the first SRPs for the same query.
how i have been using it 'at home'. its ok conversationally. It honestly is a very 'mid' search engine. Sometimes it can zoom right onto what I want. Other times it gets very fixated on the 'wrong thing' and you have to club it into expanding its view. It just randomly gets 'lost' and you have to nudge it along to actually dig in. even then even if you explicitly tell it to review its sources it will just kinda ignore the request and keep going.
I think the gap here is you're thinking that you are their customer.
We literally paid them for a Gemini subscription. How does that not make us a customer?
Google is a dying company. Abandon the ship! Everything they do in terms of AI is really worthless.
It truly is. I acutallu signed up for an annual plan because I used the image models for design ideas on my house. Few weeks ago they announced changes, no more pro models. It’s laughable how much they keep screwing up.
Gemini is completely integrated in my company's ecosystem. They don't care that you don't use it.
skill issue?
This is a new video model. It is not a replacement for Gemini 4, it is doing something altogether different.
It has been available, since yesterday, within Google Flow. I know, because I used it yesterday. Creatives normally are using Flow or similar third party tools, not APIs or AI Studio. The claim here is that there is more character consistency, which is helpful if true. But that's not something I can see from a little work. Despite the "More Affordable" claim, it costs more, and the outputs doesn't, from my limited experience so far, seem to be significantly better. So I could only, personally, justify pulling this up when character consistency is punching me in the face. Which, to be fair, is a lot of the time, both in my experience, and the experience of other AI film makers I talk to.
This isn't for (most) coders, it hardly matters if it shows up on Flow a little earlier that Google Vids, and, perhaps, it is a little silly to act like it is a nightmare that a rollout occurs over a few days, especially the rollout prioritizes the paying creatives on the tool that are most likely to use.
It's double the cost in Flow, but 30% cheaper via the API. And Flow has always been much cheaper than the API (by 5-6x).
The "omni" naming for a video-only model is so confusing. I was under the impression omni referred to models that can output text and audio too.
The rollout is so confusing. Going straight to the Gemini app and asking when I can use this model gets me vague and contradicting answers. Ultimately it sounds like the Vids rollout will take 1-3 days, and you can go to vids.new on desktop to see if you’ve won the lottery and have it available now, apparently on any free Google account. Business accounts can use it too. Meanwhile the Gemini app is getting it on paid plans, but with different video length limits and no clear rollout timeline. If you pay for the API, it’s available now apparently.
Why is this so difficult to figure out?
https://flow.google.com/ is where most people would use the model. You get credits with your monthly sub that you can spend on flow.
Never even heard of this until today.
I think they are doing a terrible job with their product packaging. Like others, I also gave up with Google models long ago due to difficulty in accessing and low limits.
My gut is that this is reflective of their fragmented internal product organization around AI.
Google has been like this forever.
Excellent engineering and awful product.
It's okay to say that Google's engineering is also poor. If they had good engineering, they'd study their users and lead to product improvements. It's really not either/or.
It has been a while, but I used to read AOSP source code quite regularly, and always wondered where do all those engineers that pass the famous Google interviews land, because they certainly weren't writing the Java and C++ code in Android, given how it looked like, or the whole mess with NDK headers, and other stuff in the tooling.
Ah, and the Eclipse/Android Studio stable releases, that usually have bugs as if they were still preview, release candidate ones.
What's clear to me: leetcode and quality code are not the same skillset.
Uhg that’s my issue. I keep getting it mixed up with https://wave.google.com/
Go with the Flow, don’t ride the Wave.
I opened ai studio and i see it popping up, on a normal account: https://aistudio.google.com/prompts/new_chat?model=gemini-om...
>Going straight to the Gemini app and asking when I can use this model gets me vague and contradicting answers.
In other words, business as usual.
https://x.com/signulll/status/2067446889956430273
Still can't handle the prompts I throw at it, though the 15 seconds gives more breathing room. Not being able to use first and last frames for the longer videos doesn't render it useless, but it gets pretty close (the model just cannot reproduce unusual or idiosyncratic character designs without first and last frames; fat, muscular, ugly, and non-human characters regress to tv commercial blandness).
I just don't know how video is going to make it, long-term. You have to go through so many iterations to get something with even a semblance of what you were going for. I've been making AI videos since early 2024, and while image quality and coherence has gotten better, the ability of the models to actually follow your directions precisely hasn't. I would rather have to wait 5-10 minutes for a clip that actually does what I tell it to than get something useless in 30 seconds.
Somewhere in the middle of the video it would suddenly scroll one screen back, happens with all videos on the page, eh
When playback ends on one of the videos, the page jumps back to that video for some reason. I saw this too when I had multiple videos playing at once and the earlier clip on the page finished while watching another one lower on the page.
noticed that too
Completely random, but I was inspecting the source code for their Omni 1.2 Flash web page, and there are insane amount of newlines. There are seven newlines between the first line and the DOCTYPE. There are huge amount of newlines between each section. What did they use to generate this website?
Enterprise websites tend to have very complicated content generation systems with little care about performance (including html/css/js size). What's hypocritical is that Google penalizes in SERP sites which have poor performance score yet Google's own sites are far from exemplary.
Enterprise websites have often enough this issue.
they might have a composer layer which loads certain components (transparent to you) with and without authentication and mixing different backend sources.
Then they might have templating engines and frameworks.
Sometimes they just don't care but lets be honest: a few new lines are a few bytes.
But the most critical pages go through optimizers for traffic
Perhaps they used:
https://en.wikipedia.org/wiki/Whitespace_(programming_langua...
It's the watermarking of the AI they used :)
A simple observation I have
> https://storage.googleapis.com/gweb-uniblog-publish-prod/med...
Text from the first paragraph on the page says
> Today, we’re rolling out Gemini Omni 1.2 Flash, and making it generally available across the Gemini API and Gemini for business. Gemini Omni 1.2 Flash introduces native 15-second video creation, multi-turn conversational editing capabilities, and enhanced reference following. We have dropped the price 30% from Gemini Omni 1.1 making it available at $0.07 per second for 720p video.
Focus on this part
> We have dropped the price 30%
The text to speech audio says
> We have dropped the price 3%
They built a youtube-immolation machine - but to what end?
Parity with their admitted Search-immolation machine maybe
I desperately wish Google would focus on improving the Gemini capabilities that are built-in to their office suite (docs, slides, etc.). The current built-in Gemini is terrible, and if I could use e.g. Claude natively with those products my workday would be an order of magnitude more productive.
I'm shocked that they don't seem to be investing in agentic capabilities there at all.
I built https://smalldocs.org / https://github.com/espressoplease/smalldocs basically for this. It lets any agent you want create rich documents - word, slides, spreadsheets, mermaid diagrams, rich code rendering, etc. About to launch a cloud first version of it, so what's on the site now will evolve soon.
Price is $1.50 in Input and $9.00/$12.00 (text/video) in Output per 1M tokens
$0.07 per second for 720p video.
Is the previous Gemini 4 Carbon model live yet? I still don't see it on OpenRouter... but this one seems to be there
Not sure why this is so mysterious to everyone.
Google's main source of revenue is advertising (yes, still).
This is one of paths that serve that.
Still confused? Look at what Meta is doing with end-to-end ad creation.
I'm not saying Omni makes good ads. It's all part of the toolbox.
What's this name and versioning now, NOBODY understands Google's Ai products, including subscribers to them.
Why their models go from 3 to 4 but then whatever this is starts at 1 I honestly don't care anymore
My friend on Google ai plus is still stuck on gemini 3.6 flash and now they're taking away 3.1 pro from them too even though it's long obsolete, like zero regard for user interactions here
honestly the naming on all of them from all companies is weird unless you 'know where to look'. I get having a cool code name for each one. But give me a clear way to look at it and see why should I use one model over another? like 'low/medium/high' something?! please!
Extending an exiting video by 15 seconds is pretty nuts.
That is sad... I tried to guess and ignore AI videos (created to be deliberately misleading as "realistic") by length, but I guess that time is over now.
It's amazing to me how poor adoption has been for video models, at least in my corner of the internet. I see considerably more Code to Video content from libraries like HyperFrames than any native AI video. It seems that users are much more open to AI coded "Explained style" videos than anything else.
It's still pretty inconsistent, slow, and expensive. I have claude write up massive prompts with reference images and it'll still come out poorly, but since I'm not generating much revenue with the product I use these models for, I just accept a lot of the sloppiness. I have tried a lot of different prompts, I think video just has a lot more free space to differ from our mental models than something like code that is a lot more well-defined.
Okay, I just wanted to vent: I'm on the Pro plan and I've not been able to use Argon for the past 4 days - every time I send a prompt I get downgraded to a Flash model.
Makes me consider cancelling the plan entirely.
Maybe I’m using antigravity wrong, so I’d love to be corrected here:
* For scientific computing it constantly cuts corners and writes awful buggy code. Bugs that will flip signs, or clamp values. Then when the model (or harness..?) discovers it, rather than fixing the underlying hacks it just adds more code with ifs, cases, and deceptive tests.
* using argon on high by itself is fast enough but seems dumb. /teamwork will catch a dozen errors from High’s laziness and sloppy tendencies
* /boost and /teamwork are ridiculously slow. Much slower than Codex Ultra. And it’s not clear when to use it.
I basically lost all trust in Argon output after running verification through “/teamwork Argon High” or “Sol 6.1 Ultra”.
I pay and I still can't test it, how is Google this incompetent?
Why can't they just name it 'Gemini Video'?
Wait, I thought we were on Gemini 4?
It would be incredibly helpful to me if AI model names followed cars.
I'm not scratching my head about a "2025 Corolla." Windows 95 and 2000 had this built in, but then they lost it.
I'd prefer it was called Omni 2026.2 instead. It embeds a date and model and would package it as a product for on-premise or local hosting. It'd be easier to sell to the bean counters in the future: "We're running Omni 2026.2, we want to upgrade to Omni 2028."
They aren't far from cars. Cars have generations and each generation has distinct models, for market (region), body style, maybe generational steps with facelift, etc. So your "2005 Corolla" is Corolla Gen9 (2000-2007), designated E120/E130 (no idea how they differ), and can be 3-door/5-door hatchbacks, 4-door sedan, station wagon, varying which market you look at.
Omni is a video model.
Which, from a naming standpoint, doesn't make any sense
But it still starts with “Gemini”. It would be better if they called it Gemini Omni 4, 4.1 etc. Otherwise call it just Omni 1.2 and drop the Gemini prefix.
It’s still bad versioning and will confuse people.
Not going to even read this post. While I have been really bullish on Google I am now tired of them constantly baiting us with vaporware announcements unfortunately.
It seems to me that LLMs are using simulation for cloth generations.
They move a lot like 3d models.
It’s not the LLMs themselves, it’s the training data that they are trained on being supplemented by synthetic video from simulations / 3D rendering.
Trying to "test" the model on OpenRouter, and keep failing with Gemini uptime being "68%".
What a great way to start.
please integrate it to antigravity cli
Why do we need this?
TFA is literally a marketing piece for why someone might want to use it. If you personally don't need it, that's ok, you're just not their target audience.
its pretty telling that the only compelling argument put forth by their own article is 'advertising'
Hot take, Google should focus on their open weight models. I use Qwen 3.8 and it’s been amazing. I would pay for Max but there’s no provider or server here in the states. If they had an amazing Gemma 5, I would seriously consider trying their closed source models. Model comfort and familiarity is a real thing.
I'd also love to see Gemini 4 Flash
Google, just go back to making products instead of throwing announcements.
At least this one is available immediately.
I don't trust Google on AI. They should do an spin off if they want to be relevant in AI.
DeepMind would be a good name for it!
Waiting for someone to test it with "~~A pelican~~Will Smith eating pasta while riding a bicycle" or something...
If I understand correctly, the use-case for this is video B-roll? Because what else am I supposed to do with 15s of videos.
So, to badly impersonate N-Gate, "Business Model: Uber for Getty Images"
The average length of a shot in full length feature film today is like three seconds. Even back in the old days, the average shot was like 12 seconds. Most movies are made up sequences and scenes made up of lots and lots of shots.
Increasing the available range of shot lengths from 10 seconds to 15 seconds gives film makers a wider pallet of options, if the longer shots have continuity, which is the other claim here.
I'm not aware of any "one-shot", one and done, use cases that are much more compelling the clip is 15s vs 10s. But that's not relevant to many people actually using AI video models like this on a daily basis.
That's cool, but I don't see why one would want to watch a movie like that.
For B-Roll it's different. I can somewhat see the why behind that. It's just to give your brain something to look at so that you don't have a voice-only podcast.
But movies OTOH are these creative endeavors where the shots _are_ the content instead of just augmenting the content.
This won't stop the industry from still trying to use this tech like that, of course.
Imagine a hybrid movie, where film maker spends $5000 filming a couple of actors, but uses $50 of AI shots for special effect or other footage that would have been completely out of budget before.
I mean, if you just want to claim that AI lacks soul, and its an abomination if every frame in film isn't make by a chemical process involving silver nitrate, like the good Lord intended, that's fine. But there are certainly film makers who don't share that sentiment.
I can imagine that, but I don't think that this is the right AI tool for the job.
I would rather expect people to use AI to drive more conventional VFX tools, as that is reproducible, can be versioned, and, most importantly, will contain true intent in aspects.
Something this simply cannot do.
Of all the uses of AI, producing convincing video is just about the worst one in my opinion.
it’s a shame that Google/Gemini decided to restrict and fuck other adventurous yet extremely gracious paying customers; who were excited at the prospect of not being locked in to two obnoxious companies.
oh well, I guess we’ll just have to take that enthusiasm and roll it over to all the open weights on providers like fireworks.ai. or others.
i’m so happy I got in early and paid for $200 a month for shit ass Opus 4.6 and sonnet 4.6 and being locked into the absolutely limited Gemini 3.8 flash.
> who were excited at the prospect of not being locked in to two obnoxious companies.
Is being locked in to three obnoxious companies that much better?