I'm scared to make this video

I need to be real with you guys. I'm legitimately scared to make this video. Not just because I think Google might cut me off from early access because I'm here to tell you how bad a lot of the things they're doing are, but because it might put my entire career at risk. The last time I came out to explain just how bad a Google product was, it was when Anti-gravity originally released. And my punishment for doing that was that they demonetized the video, stopped recommending it entirely, and they actually manually flagged it as enabling dishonest behavior. That flag killed distribution for a couple videos on my channel and risked my whole career. So, know that I'm doing this despite the fact that it could actually hurt me because I think it's important to share what's going on and why I'm so skeptical of Google's direction. I could easily do four videos of content about the things I'm about to talk about, but I don't want to waste your guys' time or energy. I want to really try to hone in on these topics and give you what you need as quickly as possible because we just experienced Google IO where they announced the new 3.5 flash model that I was lucky enough to have early access to as well as the new anti-gravity CLI which comes with a fun side effect of ending a lot of their open- source support. And then the stupidest thing by far, killing their own customers. railway service just went down entirely because Google Cloud as one of their layers banned their account that they're spending 2 million plus dollars a month on for no reason. You do not hate Google enough. This video is going to kill a lot of opportunities for me. It's going to potentially risk my channel entirely. So, I'd like to take a moment to thank today's sponsor for continuing to support us despite what's about to happen. I don't know about you guys, but I find myself waiting a lot more than I used to when I wrote software. Not literally me waiting. agents write code really fast. I'm talking about the actual things that I'm shipping. Since I have agents running all of these complex multi-step processes, it's kind of annoying to know what's going on when and I'm sitting there waiting for each test, not knowing if it's even working. At least that was the case before I started using trigger.dev. You will understand how powerful this is immediately. If you like TypeScript, and if you don't, your agents will. Trust me. The new real time stuff is so cool. You can take a job like some TypeScript you wrote to take an image that was generated, pass it through a filter, upscale it, and then send it. Or maybe you take a text prompt and then you refine it before doing that next step. Each of those individual pieces can be run as simple JavaScript functions that you write or trigger that are easy to deploy and manage from the back end and the front end. And the depth this goes to is insane. Ben implemented this on BTCA web. And I'm going to ask this agent to go explore the bun repo and figure out how much of the code is Zig and how much of the code is Rust. This is a complex process because it has to read the prompt to understand the intent, go run another job to clone the codebase, and then give a result after it does all of the traversing. And here we can see all of the work that is being done. You can click on the process as it's executing and see each step as it comes in. Do you know how hard it is to get a UI like this to explain these types of complex workflows? We have stuff like this internally, but it is so hard to get right. And the craziest part is that it even works on your local dev servers. As the work your code is doing gets more complex, you need tools like this to keep it simple. Save yourself a lot of pain at soyb.link/trigger. Now that my bills are paid, let's start with the Gemini 3.5 flash release. Speaking of which, flashbang warning. Gemini 3.5 Frontier Intelligence with action. To be very clear, this seems to be the best model Google has ever made. In the old days, this would have been one of those King Crown thumbnails that we used to do when a new model came out that was good. And I've actually historically really liked the Gemini Flash line. Gemini 20 Flash is probably the single model I've done the most tokens through because we use it for a ton of stuff in T3 Chat. Well, we did before they deprecated it and pushed us to models that were up to 30 times more expensive. We're not talking about costs yet. We're talking about numbers. So, let's take a look at how 35 Flash performs. In almost every single thing, in fact, I think actually in every single thing, it seems to outperform Gemini 31 Pro. There are exceptions to this. I ran this against Skatebench and in Skatebench it actually underperformed 31 Pro. In fact, it underperformed a lot of things which is crazy because 31 Pro was the best model by far in Skatebench. So, there's definitely less knowledge in this model, but they weren't really working on the knowledge. They were working on the capabilities. They wanted this model to be better at doing agentic work. They even told me as much when I got the early access test this for coding style tasks. And when you look at the numbers, you can see why. Their terminal bench score is the highest terminal bench score other than GPT55. And it's very close to that. It's absolutely crushing Opus and Three Flash and even beating out 31 Pro. SWB Pro, similarly did very well. It's underperforming Opus 47 and 55, but it's still outperforming 31 Pro. MCP Atlas, it got a best-in-class score in Toolathon. It got state-of-the-art for Binance Agent. It got state-of-the-art for Chars of Reasoning. It did state-of-the-art. MMU Pro, it did state-of-the-art. It's good numbers. This model seems to be quite good. One of the crazier things they shared here is the artificial analysis intelligence index. You're familiar. I use these numbers a lot. I love artificial analysis. Their intelligence bench might not be the best overall, but there's a lot of good insights to be had from them, especially when you're comparing the speed, the price, the performance, and all of these other aspects. I find them to be a good general reference point. And you'll see in this chart they're showing 35 Flash has by far the craziest ratio of speed to performance. It is performing comparably to 31 Pro, Opus 47, and 55 at a significantly faster speed, hitting easily the near 300 token per second range. It's clear artificial analysis is trying to form a better relationship with Google, so they have not been able to talk too much [ __ ] here. But that's what I'm here for because this model [ __ ] sucks. I want to show you guys just how bad it is to use. But first, I have to show you guys just how misleading this all is. First, we have to talk about the price. Noticing that this page doesn't have a dollar sign anywhere on it. So, they're trying to hide the cost here. The reason they're trying to hide it is because they tripled the price again. It is now $1.50 per mill tokens in and $9 per million tokens out. $9 per mill out on a flash model. For reference, three flash was $3 per mill out and 50 cents per mill in, a third the price. And to go back to my beloved 20 flash, where this all started, one of my old favorite models, sadly deprecated soon, 10 cents per mill in, 40 cents per mill out. That is a greater than 20x increase purely based on the input and output token prices. But even just measuring this would be dishonest because there's a bigger issue which is that this model is a reasoning model and it certainly does a lot of reasoning. Here are the output tokens used for artificial analysis to run their benchmarks across various models. The worst offender for too many tokens by far is Deepc V4 Flash which did 240 million tokens. GBD54 Mini didn't do much better, roughly around the same. Sonnet did 200 mil which is absurd. But I want to go further to the right because Gemini does perform okay in that regard, but not when you compare it to GBT 555 Medium which did 22 million tokens. A tenth as many as V4 Flash did and roughly a tenth as many as Sonnet did. Token efficiency is a very underrated aspect of all of the things that we're doing nowadays. And OpenAI seems to be the only company taking it very seriously. Everyone else is just inflating the amount of tokens the models are willing to use to solve a problem which does mean that they behave smarter. They have more tokens. But how many tokens does it take to get a certain score? That's a whole separate measurement and Google's not doing great there. For what it's worth, the efficiency is only slightly worse than Gemini 3 flash was. 72 mil versus 73 mil. When you look at the intelligence versus output tokens chart, you'll see GPT55 medium and X high are absolutely slaughtering Google. And none of the models from Claude or Deepseek or anything else are even measuring in this section at all. But ready for where things get really messy, the costs. 35 Flash is now the fourth most expensive model that has been through these benchmarks in modern times. It is almost twice as expensive as 31 Pro was because it generated so many more tokens. So sure, the per token cost might be half as much as Pro, but the cost in actually using it is massive because it wastes tokens like I've never [ __ ] seen. So, I don't care how fast it is because it's not going to complete tasks faster because it's going to generate such an absurd amount of tokens. If it's 2x faster than other models, but it generates four times the amount of tokens it needs, it still ends up being 2x slower. But if it's good, it's worth it, right? Well, let me show you something. One of my recent favorite agentic benches is taking my old game I was working on, Fish Slop, and asking the model to rewrite it with access to the original source, but the ability to rebuild it in a better, more stable way so that I can finish the game with a less sloppy codebase. So, how did 35 Flash do? Not very good. It is the only model I have run this test on that didn't manage to make it work. It wrote code that was broken. It didn't check. It didn't run anything. It didn't do anything. So, I told it to fix it, and it did. And somehow the actual version is worse. It has this awful glow. The fish it put in the game are too big to actually do anything with. The feeding mechanic doesn't work. The aging mechanic doesn't work. This guy doesn't do anything at all. It's just entirely broken. Like, it didn't even kind of get it right. Oh, I guess that his thing works. It generated new images. They all suck. And some of them don't even have their transparency set. Right. This is embarrassing. This is like legitimately embarrassing for a allegedly state-of-the-art model. I gave the same task to 55 and it did it so well I asked it to make the game 3D instead and it did that too. Insane that they were shipping something this [ __ ] and claiming it's good at this type of thing. Like actually absurd. And that's separate from all of the terrible bugs I ran into with the anti-gravity CLI. And here is where the next stage of this crash out happens. First off, scroll is broken. I'm trying to scroll and it's putting my inputs from before in the input box instead of scrolling. I'm not running T-Ux. I'm not SSH. This is a boring ghosty terminal instance and the CLI just doesn't work. Oh, also email leak uh face do your thing. C doesn't work to exit it. You have to manually slashexit and every time you open it has to resign in. Okay, that sucks. But how is it to actually use? Let's start with some slash commands. I'll make sure we're using flash high. Did that right. Didn't leave UI in the bottom like it did before. Let's hope it continues to do things right. Thankful I'm filming this off stream because it doesn't seem like there's any way to hide my email from inside of the CLI. Love that. I don't know if you saw that the generating getting frozen and locked up. Tip. To disable tips, toggle show tips in settings. Cool. Notice how the input box is moving all around. It consistently does that. Even better is it often will ask you for feedback. Like it'll ask questions. It'll ask you to literally submit feedback to the team. Those types of things. And when it does that, see if I can force it. Crl S to submit for some reason. Okay, that time it cleared properly. Earlier when I did that, it left all the other UI in the bottom and I couldn't get rid of it and it would just change where I was typing. It's the buggiest CLI I've used in a while. It was really, really bad. I ran into so many broken buggy states when I was trying it earlier. I actually couldn't fathom it. And the fact that I have to close with their special/exit command is just the icing on the [ __ ] cake. Is that Gemini CLI wasn't trash. If you're not familiar, Gemini CLI is an open source project that Google made for using their models. This came out a little after the codec cli and the cloud code stuff blew up originally like early last year. I believe Gemini CLI was not my favorite thing, but you might have noticed I don't talk too much trash on it. Well, we'll get to why in a bit. I actually had a lot of faith in this team though and it was cool having a good open-source option in particular because a lot of the other labs, specifically the Chinese ones, were using this as a starting point to make their own CLIs. And it seemed like a good starting point there. They had some cool patterns in it for making and sharing skills and workflows. Like this was a fine project that was going places. Flashbang warning. An important update. Transitioning Gemini CLI to the anti-gravity CLI. When we shipped Gemini CLI last year, our goal was to bring the magic of Gemini directly into your terminal. Along the way, we've learned a lot from our community of billions of users with over 100K GitHub stars, 6,000 merged PRs, and hundreds of contributors, including you love a good terminal UI you appreciate. I think this is a bug in the formatting of this post. Uh yeah, it is. It is what it is. They had over 100k GitHub stars, 6,000 PRs merged, and according to them, you love a good terminal UI. You appreciate that we ship weekly releases, and your workflows have simply outgrown those early days of 2025. Gemini CLI proved the terminal could be an incredible interface for agentic tasks, but your needs shifted. You require multiple agents communicating with each other to split up the work and solve complex problems. This means your terminal tools need to share a unified backend with the rest of your workflow. They're folding it into anti-gravity and they put out a new anti-gravity CLI that was written from scratch and Golang. Anti-gravity CLI focuses on all of the core functionalities of invoking, monitoring, and interacting with anti-gravity agents. It does not try to be a gooey, which is better suited for anti-gravity 2. This way, it remains perfect for users who prefer a true fast and lightweight CLI experience. Obviously, not fun, but we can keep using Gemini CLI, right? Well, on June 18th, Gemini CLI and Gemini Code Assist IDE extensions will stop serving requests for Google AI Pro and Ultra. So, those who are in the subscription plans like I am to use Gemini models in our coding tools will only be allowed to use it with anti-gravity going forward. But at least we can fix the new CLI, right? Nope, cuz it's closed source. So, Google killed a promising open-source CLI that was building faith with the community in favor of a closed source CLI that doesn't [ __ ] function. Are you joking? So, so far we have a model that sucks at coding that doesn't actually write code that works. We have a CLI that's replacing a decent open-source one with an absolutely broken closed source one. At least it's written in Go, right? Cool. Instead of TypeScript, so terrible. But that's not even the worst thing Google did today. We need to talk about Railway for a second. I know they've been a sponsor in the past. They will likely sponsor future videos. I love their service. It's actually one of my favorite ways to deploy persistent servers and databases for my quick one-off projects. I built a lot of random things on Railway. It's been awesome to work with. Railway made a bad choice early on to build on top of Google Cloud because Google Cloud's primitives best matched what Jake had in mind for Railway initially. Also, fun fact, I was interviewing at Railway in 2022. I would have been the fourth employee at the time, but Jake encouraged me to go do my own thing, even though he gave me an incredible offer. I chose to go build what became paying in T3 tools and also make this channel. That's why I'm here now. [ __ ] I misremembered. It was 2021, not 2022. This was way, way back. I'm very thankful to Jake for being a mentor and helping me start my business and make it what it is today. I owe him a ton. I've never been hesitant to push back on his thoughts. And when he told me they were building on Google Cloud, I was very, very skeptical. But he insisted they had the right pieces, as frustrating as they were. That went so poorly that Railway ended up building their own metal and building their own cloud and moving as much to it as they could. Their servers run on railway hardware, but their web layer, their CDN, and all of that is still built largely on Google Cloud. I don't think it will be much longer because Google cloud blocked the railway account entirely which took any webf facing service at railway offline. On one hand, you own the service. Jake's responsible. Railway's responsible for any downtime and outages. This is a really brutal one. I think it's still out right now. I wish I could say this is the first time Google has done something like this, but I know way too many examples. This is one from two years ago where Google accidentally deleted an Australian customer's account. This isn't some random Australian. This is a $135 billion company, well 90 bill US in pension fund management. Google Cloud CEOs confirmed that the disruption arose from an unprecedented sequence of events whereby an inadvertent misconfiguration during provisioning of UniSup's private cloud services ultimately resulted in the deletion of Unisuper's private cloud subscription. The funniest part about this one is they had a backup in place with a different provider. Thank god they were using something other than Google Cloud for their backups or they would have lost everything. and they were able to minimize the data loss by having that other backup. It's almost unprecedented to think that an organization of that size could just have their entire cloud account completely deleted. Your entire search, all the applications you're running, all the data those apps store, all views, yada yada is actually insane. People originally thought this was a cyber attack because of the severity of the error. I have heard stories like this far too many times. things as absurd as rate limits that Google didn't even know were in place being hit by railway causing their stuff to go down and nobody internally could respond to it. As someone who shits a lot on Azure and deals a lot with AWS, these problems are Google Cloud specific. Azure is sometimes weird and slow, but if you raise enough alarms, they'll do something about it, and they certainly won't accidentally delete your account if you're spending $2 million a [ __ ] month on them. Azure might be [ __ ] but they know how to make money. AWS still has their stuff together. There's a reason they're number one is because they actually work. Google Cloud is a joke. It is actually insane how bad and unreliable and terrible it is. But I haven't talked much [ __ ] on it for a while. There's three core reasons why. Reason one is because the last time I tried to talk [ __ ] on Google Cloud, nobody cared. It was one of the worst performing videos I ever did because Google Cloud is so irrelevant. Most people don't give a [ __ ] Reason two is I'm scared for my channel. I mentioned this at the start, but Google has had a bad track record here of demonetizing, limiting, and marking my videos as dishonest because I talked [ __ ] on Google product. This did get reverted and I did get apologized to for it. But I am still scared because this is my livelihood and I do love YouTube. I build on top of this platform for a lot of reasons. It's not just the money. I watch a lot of YouTube. I love YouTube. A lot of my friends are people I met through this platform. I need YouTube to succeed. I I need it to do well, but that doesn't mean I'm not scared when these things happen. The third reason is one you might not expect. It's actually three reasons. They're named Dmitri, Jack, and Gal. These three people have been awesome to work with. They have never been shy to DM me, to correct things I'm wrong about, but also thank me for the things I was right about, to give me the insights and information I need to cover these things responsibly, and to give me faith that the Gemini CLI was going in a good direction. They are why I didn't talk [ __ ] because believe it or not, I only tend to go public with my complaints when the private ones aren't being addressed. That's why I've been so hard on anthropic. That's why I'm generally relatively quiet about OpenAI because when I have an issue with OpenAI, I tell them and they usually fix it. Dimmitri and crew have been awesome about this. They were very upfront when I first complained about the Gemini CLI saying, "My DMs are open. Give me all of the feedback. We want to address it." And they have. All three of these people have been incredible to interact with and work with, and they built a ton of good faith from me. Enough that despite not using the Gemini CLI and other Google tools that much, I knew they were going in the right direction. All of their roles have been replaced by the anti-gravity team. You might notice the new anti-gravity UI looks a lot like the Codeex app. To their credit, they were early here. I even said as much in my anti-gravity crashout video, and I would defend them on this saying, well, they were first. But something really funny happens a few seconds later. This is the official Google account announcement video for the new anti-gravity app. He's adding a folder from documents. And one of the folders is named Codeex. They can't even hide it in their demos. They are biting so hard and at the same time ignoring the awesome people internally who are actually trying to make a good product. This all comes down to internal politics. Google spent a bunch of money buying the founders off of Windsurf to have them come in and build this anti-gravity thing. And as a result, all of the wonderful people who have been trying their hardest to make a good open solution inside of Google for the world to use have had their work just retired. And it sucks because they've all done a phenomenal job of building community, building good relations between people like me and their team, and giving me faith they were actually going somewhere with it. And what they've been replaced with is somebody who can't stop biting codecs for long enough to hide it from their announcement videos. Insulting is putting it lightly. This is absurd. This is big company politics [ __ ] getting in the way of not just the best ideas or some things getting away with the best people. It is driving good people to madness and resulting in absolute [ __ ] slop coming from the Google machine. And that's not even talking about how horrible the models are to actually use, which is its own whole separate thing. It seems like Google still hasn't really cracked the RL thing. And the models just don't know how to check their own work and steer themselves in the right direction. They just burn tokens for no good reason. I'm done. I I'd love to be wrong. Enthropic and OpenAI need competition, but I have more faith in Cursor than Google at this point. Like I think cursor has a much bigger chance of making useful things for devs than Google ever could at this rate is insane. And if you're looking for the types of people that can keep me from crashing out about your broken [ __ ] that can build something good in the open, take feedback well, and run communities, I don't know what the employment status is for these people, but I am sure that Dimmitri, Jack, and Gal would at the very least accept a DM right now. These people are good. I will stand behind them even if I have to stand against a trillion dollar company in the process because all three of them did such a good job that they prevented this video from happening for over a year. And now that they've been screwed over, I'm not waiting anymore. I'll risk it all because what the [ __ ] are you doing, Google? How do you throw away the work of these awesome individuals in favor of absolute [ __ ] slop the same day you mislead everybody with numbers about how fast and cheap this model is? It's just so bad faith to put this chart in this announcement and not include the price anywhere and also take down railway entirely the same day. Google is not just a company that doesn't care. It's a company that's incapable of caring in its current structure. And I have honestly spent more time helping my friends at Google leave than I have spent building or using anything Google makes right now. They can still turn it around, but I don't have any faith they will. Google has all the things they need to make something incredible. They have the infrastructure. They have the talent. They have the ecosystem. They have the TPUs. They have the research. But the result is that no one's ever allowed to work together long enough to make anything good. And half the time when talking to somebody there who might actually make something cool happen, they've left the company before the thing is done. It's a slot fest and I can't support it anymore. If you're building things on Google stuff, get it off. You can't trust any of this. And I'm scared as hell because my whole career is currently built on top of YouTube. I have nothing else to say on this one. Hopefully this video still makes it into your feeds even though Google's going to do everything in their power to hide it. I hope I'm wrong on this. I really do, but I don't think I am. I am curious how y'all feel. Am I overreacting here or is this as insane as I think it is? Give those people a follow. They all deserve it. And until next time, peace nerds.