Where the f*** is Gemini 3? 🔗 **Source:** [Where the f*** is Gemini 3?](https://www.youtube.com/watch?v=eUG6r8ixngs) > [!note]+ Document Notes > tags: [ai-models, google-gemini, model-release] > author: Theo - t3․gg > source: YouTube > reference: https://www.youtube.com/watch?v=eUG6r8ixngs > date created: 2024-12-19 10:30 > > # 🏷️ Title: Where the f*** is Gemini 3? > <!-- deep-read:sections start --> ## 🧰 Action Inbox - [H1] Track Gemini API deprecation schedule changes — November 18th deadline #monitoring - [H2] Test CLI access for unreleased models — authentication workarounds #testing - [H3] Monitor competitive positioning against Claude/GPT releases — market timing #strategy ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - [Gemini API Documentation](https://ai.google.dev/) - [Google AI Studio](https://aistudio.google.com/) ## 📝 Summary - Gemini 3.0 Pro has been rumored for months but remains unreleased despite multiple leaks and CLI access attempts - Google's aggressive API deprecation schedule suggests imminent release but keeps getting delayed - Internal organizational fragmentation between DeepMind, AI Studio, and Vertex teams may be causing coordination issues - Google faces pressure to release state-of-the-art model after competitors like Claude 4.5 Sonnet raised the bar - Evidence suggests model exists internally but various factors (benchmarks, legal issues, technical problems) are blocking release ## 📖 Main Content Google's highly anticipated Gemini 3.0 Pro model remains mysteriously absent despite months of speculation and leaked evidence of its existence. The delay appears connected to Google's unique position in the AI landscape as the only company competing across applications, foundation models, inference APIs, and custom hardware simultaneously. The organizational complexity created by separate teams for DeepMind, AI Studio, and Vertex AI in different locations may be hampering coordinated releases. Meanwhile, competitive pressure from recent releases like Claude 4.5 Sonnet could be forcing Google to delay until they can ensure state-of-the-art performance rather than risk releasing a second-place model. ## 🗒️ Notes #### ⏱️ [00:00](https://www.youtube.com/watch?v=eUG6r8ixngs?t=0) — [H1] Missing Gemini 3 despite competitor releases - Key idea: Gemini 2.5 Pro released March, many competitor models since, but no Gemini 3 - Evidence: "Llama 4, Quen 3, Claude 4, Claude 4.1, Cloud 4.5, GPT5, 03, 04 Mini, GPT5 Codeex" #### ⏱️ [00:30](https://www.youtube.com/watch?v=eUG6r8ixngs?t=30) — [H2] Persistent rumors without delivery - Key idea: Weekly release rumors for 3 months with no actual release - Evidence: "I've been hearing rumors that this model's coming out this week for about 3 months now" #### ⏱️ [02:15](https://www.youtube.com/watch?v=eUG6r8ixngs?t=135) — [H1] Aggressive API deprecation schedule suggests release timing - Key idea: Multiple Gemini snapshots being deprecated November 18th indicates planned replacement - Evidence: "All of these are being deprecated on November 18th" #### ⏱️ [03:45](https://www.youtube.com/watch?v=eUG6r8ixngs?t=225) — [H2] Deprecation timeline changes hint at delays - Key idea: Original November 18th deprecation date spread out, suggesting release delays - Evidence: "Previously all of these were put on this week of November 18th because they planned to release it that week" #### ⏱️ [04:30](https://www.youtube.com/watch?v=eUG6r8ixngs?t=270) — [H1] Competitive pressure driving delays - Key idea: Google delaying to avoid releasing second-place model after Claude 4.5 Sonnet - Evidence: "They might have had a model that was state-of-the-art for code and then 4.5 sonnet dropped which scared them" #### ⏱️ [05:15](https://www.youtube.com/watch?v=eUG6r8ixngs?t=315) — [H2] Stock price concerns with benchmark performance - Key idea: Google's equity at risk if Gemini 3 benchmarks below competitors - Evidence: "If they are behind open models, like if Kimmy K2 is benching higher than Gemini 3 when it drops, Google's equity in stock is going to plummet" #### ⏱️ [06:00](https://www.youtube.com/watch?v=eUG6r8ixngs?t=360) — [H1] Google's unique multi-platform position - Key idea: Google competes on infrastructure hosting and their own models simultaneously - Evidence: "They're the only company I know of that's competing on both sides here" #### ⏱️ [07:30](https://www.youtube.com/watch?v=eUG6r8ixngs?t=450) — [H1] Organizational fragmentation at Google - Key idea: Separate teams for AI dev, DeepMind, AI Studio, Vertex in different locations - Evidence: "These are all different teams with different metrics with different OKRs with different goals in different buildings in different states" #### ⏱️ [09:45](https://www.youtube.com/watch?v=eUG6r8ixngs?t=585) — [H2] CLI access evidence for Gemini 3 - Key idea: CLI allows prompting with Gemini 3 model name but returns rate limiting - Evidence: "If you do Gemini-model Gemini 3 Pro preview 112025, it does let you prompt" #### ⏱️ [11:00](https://www.youtube.com/watch?v=eUG6r8ixngs?t=660) — [H3] UI generation quality differences - Key idea: Alleged Gemini 3 outputs show different styling from 2.5 Pro - Evidence: "These little pills like that, not something I've ever seen Gemini 3 do" #### ⏱️ [12:00](https://www.youtube.com/watch?v=eUG6r8ixngs?t=720) — [H2] Access blocked during video production - Key idea: CLI access to Gemini 3 was blocked while video was being made - Evidence: "In the time it took me to post this, it seems that Google has blocked it" #### ⏱️ [12:30](https://www.youtube.com/watch?v=eUG6r8ixngs?t=750) — [H3] JSON test files reference Gemini 3.0 Ultra - Key idea: Recent CLI updates include references to unreleased Ultra model - Evidence: "Recent update to some JSON test files for the Gemini CLI that includes Gemini 3.0 Ultra" ## 📚 References - [Google AI Studio](https://aistudio.google.com/) - [Gemini API Documentation](https://ai.google.dev/) - [Google Cloud Vertex AI](https://cloud.google.com/vertex-ai) ## 🗣️ Paraphrase - Google's Gemini 3.0 Pro has been anticipated for months with constant rumors but no actual release, despite evidence the model exists internally - The company's complex organizational structure with separate teams for different AI products may be creating coordination challenges for releases - Competitive pressure from recent high-performing models like Claude 4.5 Sonnet appears to be causing Google to delay release until they can ensure market-leading benchmarks ## 📰 Newsletter **The Mystery of Google's Missing Gemini 3.0 Pro Model** After months of rumors and leaked evidence, Google's highly anticipated Gemini 3.0 Pro remains conspicuously absent from public release. Recent API deprecation schedules and CLI access attempts suggest the model exists internally but faces coordination challenges across Google's fragmented AI teams. The delay highlights the pressure facing major AI companies to maintain competitive benchmarks. With Claude 4.5 Sonnet raising performance bars, Google may be holding back until they can ensure state-of-the-art results rather than risk a second-place launch. For developers waiting to integrate advanced capabilities, this uncertainty around release timing makes planning challenging. Consider preparing fallback strategies using currently available models while monitoring for official announcements. Resources: [Google AI Studio](https://aistudio.google.com/), [Gemini API Docs](https://ai.google.dev/), [Model Comparison Tools](https://artificialanalysis.ai/) ## ✅ Action Items #### DevRel - Developer advocacy: Create comparison guides between available models while waiting for Gemini 3 - Education: Workshop on handling model deprecations and API migrations - Developer experience: Document CLI access patterns and authentication flows - Developer marketing: Content strategy around model evaluation and selection criteria - Community management: Monitor developer frustration with release delays and provide alternatives #### OSPO - License Compliance: Review implications of using leaked or preview model access - Security Management: Assess risks of CLI workarounds and unofficial access methods - Software Inventory: Track deprecated model dependencies across projects - Developer Support: Training on graceful model migration strategies - Strategic Alignment: KPIs for model adoption and competitive positioning #### Community - Maintainers: Guidance on handling breaking API changes from deprecations - Contributors: Documentation of model comparison methodologies - Ecosystem: Coordination between projects using different model providers ## 🧪 Hypotheses - Google's organizational fragmentation between AI teams is causing release coordination delays - Competitive benchmark pressure is driving repeated postponements until performance meets thresholds - Internal model snapshots exist but face legal or compliance review bottlenecks - Stock price sensitivity is making Google more conservative about releasing potentially second-place models ### 📊 Methodology - Study design to test hypothesis 1: Survey coordination challenges across Google AI teams - Data sources and analysis for hypothesis 2: Benchmark timing correlation with competitor releases - Experiments to validate hypothesis 3: API access pattern analysis and deprecation schedule changes ### ❓ Questions - How do organizational silos at large tech companies impact AI model release cadence? - What benchmark thresholds trigger release delays versus launches in competitive markets? - How should developers plan integration timelines when model releases are unpredictable? ### 🔮 Future Research - Analysis of model release patterns across major AI companies - Investigation of internal coordination challenges at scale - Study of market timing strategies in AI model launches ### 💡 Implications - Practice: Developers need contingency plans for model deprecations and delayed releases - Policy: API versioning and deprecation policies need clearer communication timelines - Security: Unofficial access methods create compliance risks for enterprise users - Education: Training needed on model migration strategies and competitive analysis - Funding: Budget planning complicated by unpredictable model release schedules ### 🧠 Learning Styles Assessment Reading level estimate: 8th Grade (Flesch-Kincaid) #### 👀 Visual learners - Present: Screenshots of CLI commands, API documentation interfaces - Improve: Timeline diagrams of competitor releases, organizational charts of Google AI teams #### 🔊 Aural learners - Present: Video format with narrator explanation - Improve: Audio summary of key timeline events and technical concepts #### 🗣️ Verbal learners - Present: Clear definitions of technical terms, analogies about organizational challenges - Improve: Glossary of AI model terminology, comparison frameworks #### 🧩 Kinesthetic learners - Present: Step-by-step CLI testing instructions, hands-on API exploration - Improve: Interactive model comparison tools, migration checklists #### 🧠 Logical learners - Present: Timeline analysis, cause-effect relationships between releases and delays - Improve: Decision trees for model selection, benchmark comparison matrices #### 🧑‍🤝‍🧑 Social learners - Present: Community discussion prompts, collaborative speculation - Improve: Developer forum integration, peer review of findings #### 🧘‍♀️ Solitary learners - Present: Individual research methodology, self-guided investigation - Improve: Reflection questions on competitive strategy, personal note-taking templates #### 🌿 Naturalistic learners - Present: Ecosystem analysis of AI companies and market dynamics - Improve: Industry lifecycle thinking, competitive environment analogies ### 📊 Accessibility Assessment #### Screen reader semantics - Present: Clear heading structure, descriptive alt text for screenshots - Improve: ARIA labels for complex visual elements, table captions for benchmark data #### Keyboard navigation - Present: Standard video controls, logical tab order - Improve: Skip links for timestamp navigation, keyboard shortcuts for sections #### Contrast and readability - Present: High contrast text in video, clear visual hierarchy - Improve: Subtitle styling improvements, consistent color coding #### Text resizing and reflow - Present: Responsive transcript formatting - Improve: Scalable UI elements, flexible layout design #### Media and captions - Present: Full transcript available, timestamp references - Improve: Audio descriptions of visual elements, downloadable caption files #### Forms and interactions - Present: Simple link navigation - Improve: Clear focus indicators, accessible button labeling ### 🌍 DEIB Assessment #### Cultural breadth - Present: Global perspective on AI competition and market dynamics - Improve: International regulatory considerations, diverse company examples #### Inclusive language - Present: Technical jargon defined, gender-neutral references - Improve: Plain language alternatives for complex concepts, acronym explanations #### Neurodiversity support - Present: Clear section breaks, consistent structure - Improve: Executive summary for key points, visual content organization #### Representation - Present: Multiple company perspectives, diverse technical approaches - Improve: Varied voice representation, inclusive imagery choices #### Psychological safety - Present: Respectful analysis of company challenges - Improve: Community guidelines for speculation, constructive discussion frameworks ### ✅ Logical Validity #### Strong reasoning - Deductive strength: Clear cause-effect relationship between competitive pressure and delays - Inductive strength: Pattern recognition from deprecation schedule changes - Causal reasoning: Organizational structure impacts on coordination effectiveness - Analogical strength: Comparison to other large tech company release patterns #### Fallacies - Appeal to authority: Relying on insider knowledge without verification - Location: friend testimonials - Fix: Acknowledge limitations of anecdotal evidence - False dilemma: Suggesting only competitive pressure or organizational issues cause delays - Location: delay explanation - Fix: Present multiple contributing factors #### Tone of voice - Overall tone: Analytical and frustrated, balancing skepticism with evidence-based reasoning ### 🛡️ DARVO Analysis - Deny: No significant denial patterns observed - Attack: Mild criticism of Google's organizational structure, not personal attacks - Reverse victim/offender: Not applicable to this technical analysis - Emotional cues and rhetoric: Frustration expressed appropriately about communication gaps ## 🎯 Conclusion - Gemini 3.0 Pro appears to exist internally at Google but faces release delays due to competitive positioning concerns and organizational coordination challenges - Developers should prepare migration strategies for deprecated APIs while maintaining realistic expectations about release timelines - The situation highlights broader industry challenges around balancing speed-to-market with performance benchmarks in competitive AI landscapes ## 🧱 Refactor Appendix ### 📚 Literature Split Pad ### [H1] Google's organizational fragmentation impacts AI releases > [!quote] [Where the f*** is Gemini 3?](https://www.youtube.com/watch?v=eUG6r8ixngs) by [[Theo - t3․gg]] via [[YouTube]]. Published 2024-12-19. Accessed 2024-12 <!-- deep-read:sections end --> ## 📺 Video  --- # 📄 Full Transcript *Auto-generated from YouTube captions · timestamps roughly every 30 seconds · [watch on YouTube](https://www.youtube.com/watch?v=eUG6r8ixngs)* **[00:00]** The Gemini 2.5 Pro preview dropped back in March earlier this year. Since then, we've had Llama 4, Quen 3, Claude 4, Claude 4.1, Cloud 4.5, GPT5, 03, 04 Mini, GPT5 Codeex, and many, many more models. Which leads us to the obvious question, where the [ __ ] is Gemini 3? Seriously, I've been hearing rumors that this model's coming out this week for about 3 months now. I have no idea where it is. I have a bunch of friends at Google and it seems like they don't know **[00:30]** either. There are leaks every week, half of which turn out to be fake, the other half turn out to be questionable. There's a lot of snapshots floating around, but none of them seem real. And Google's not shy to test their models like they did with Nano Banana for the image gen stuff earlier. But none of this answers the question, where the [ __ ] is Gemini 3.0? I am so uncertain and I've been doing my best to comb through all of the results, leaks, and everything else people are talking about to try and figure out where this model is and when we can expect it. From the weird other things that Google is **[01:01]** shipping in the interim to the chaos of people trying to hijack the Gemini CLI to get it to give them Gemini 3 Pro access to the list of things they're about to deprecate. There is a ton of stuff going on here and none of it really answers the question. We're going to do our best to figure it out right after a word from today's sponsor. AI has gotten really smart, but it's way smarter when it has access to more information. You know, like the entire web. The problem is that is way harder every day to actually get info from the web as more and more companies crack down on your ability to scrape and get **[01:32]** info from their websites. That's why you need a proper browser for your AI to use. And there is no better browser than today's sponsor, Browser Base. On one hand, they're puppeteer in the cloud. On the other hand, they're an all-in-one suite of tools to unlock the web for your AI agents. Even companies that understand the complexity of scraping and infrastructure have went allin on browserbase, including Versel. Yes, really. The company that's not willing to adopt any external solutions is adopting browserbased because it solved so many problems for them. They did try rolling their own solution and they hit a combination of chaotic issues from how **[02:03]** unreliable Puppeteer can be at scale to CDN challenges blocking their access, data quality issues and limited parallelization. You have to spin up all these boxes. Browser base has handled all of this for you. You just make an API request and now you can talk to a browser from your AI agents directly. It's good to have someone reliable handling the browser management and CDN challenges so we can focus on the intelligence layer. If this sounds like you and you want your AI to have access to the web, look no further than soy.base. The first thing I want to talk about because it's the most suspicious and **[02:33]** potentially concerning is the aggressive deprecation schedule from the Gemini API. Remember, this is kind of how they do things at Google. They put out tons of different snapshots, and they'll often reroute them to other versions of the model when they deprecate the snapshot. Here are all of the models currently available in AI Studio. We got Nano Banana, which is 2.5 Flash image. Gemini Flash latest, which is still tagged new for some reason. Apparently, there was a flashlight preview in September. Forgot about that. Cool. And **[03:03]** then nothing else is new. The fact that they're labeling something from September as new still shows how behind they are. Like that's that's not how the AI world stuff goes. And Image Gen 4 is apparently new, too, even though there's no release date tagged here. They also did put out all the fun stuff with V3. So like they're moving on the media side, but they're not moving on the LM side at all lately. If we flag this down to just Gemini, you'll see we have 25 Pro flash light latest flash 2.5 flash. But a lot of places are still using the **[03:35]** old snapshots that are like specific numbered snapshots like 01-21 or 06-1705-20. Apparently all of these are being deprecated on November 18th. And a bunch of the snapshots for 2.0 Flash are being deprecated on the 2nd of December. And the flashlight preview, which I actually did use quite a bit. 20 flashlight was a very fun model. You're using this for title gen and T3 chat forever. That's being killed December 9th. Apparently, this schedule changed though. One of my **[04:05]** chatters had a screenshot from before where almost all of these same models and a few other ones were all listed as being deprecated on November 18th. And they've since changed that. They spaced it out a bit more. My honest guess here is that previously all of these were put on this week of November 18th because they planned to release it that week. And once again, I think it got delayed. I genuinely think this model is ready, but they keep delaying it for lots of different reasons. One of those reasons might be the recent models that have **[04:37]** been dropping. They might have had a model that was state-of-the-art for code and then 4.5 sonnet dropped which scared them enough to go do a bit more RL to hopefully get Gemini 3 over the line because Google does not want to release a second place model right now. If Google puts out a new model and the benchmarks are below the current state-of-the-art labs, god forbid if they are behind open models, like if Kimmy K2 is benching higher than Gemini 3 when it drops, Google's equity in stock is going to plummet. This is not financial advice. I am not a **[05:09]** professional trader. I invest in earlystage startups because I like losing money really fast, but this is probably a big part of how they're thinking there. If they have this model and it might be ready to go state-of-the-art and right before they hit the go button or they have a plan scheduled for it, something else happens that forces them to delay slightly. You really have to think about what Google has to win and what they have to lose in any of these scenarios. They really want the Gemini models to win. They're the only company I know of that's competing on both sides here because Google has a **[05:42]** platform that you can host other models with. We use them for cloud models a lot because you can use that in the Gemini Vertex like hosting. Google's cloud hosting for models lets you use other models. So, they'll still make money on infrastructure even if they're not able to make Gemini really good. But, they also had the craziest hype cycle around 2.5 Pro. I'll never forget how hyped everyone was when 25 Pro first dropped. It was it felt like a monumental step up at that time. And since then, it's just kind of sat there. Google's kind of been **[06:12]** the the least active lab in the LLM world for a bit. Kind of feels like they hit their peak earlier this year and haven't been able to maintain it since, but they only have one way to fix that. Show us something better. And if they've been this quiet for 8 or 9 months now, and what they drop isn't monumental and mindblowing, that's going to be really, really bad for them. One of the things that makes Google so unique is the stuff that they own and how they operate as a business. A company like Deepseek, for **[06:42]** example, doesn't really have presence in the application world. Like, sure, you can go to the Deepseek website and use their AI chat. It was so bad that it inspired me to make T3 chat. That's the actual origin story. V3 was an impressive model. The Deep Seek site for trying it was horrible. So, I made my own to play with it more. And that became T3 Chat, which is now its own very successful business. Thank you all for the support, by the way. So, a lot of companies here do or do not have an application surface area. Foundation models are another thing that varies. There are companies like Meta that do **[07:13]** have models, and there's companies like Grock with a Q, not a K, with a Q that don't. Grock with a Q has their own inference provider though. So you can use Grock with a Q to generate things from lots of different models and they host a lot of these different models. These four categories are really interesting way to break up the different AI companies and how they're thinking about what they build. Do they have applications? Do they have their own models? Do they do inference over API? And do they have hardware that they're building themselves? Almost none **[07:44]** of these companies have hardware. There's a few that do. Like obviously Nvidia does. Grock and Cerebrris and SambaNova all do trying to build their own chips for this. And almost everyone has cloud inference except for Meta which I find really interesting. Meta is the only company that you can't pay to use their models. You have to use Meta's models for free or on another host that will charge you money. Everyone makes money on Llama except for Meta, the creators of Llama, which I find pretty funny. Meta's also somewhat investing into making their own chips, but not super seriously it seems. Thus far, **[08:15]** Amazon is taking it a little more seriously, but has had mixed results. The only company that's truly firing on all cylinders here is Google. They have applications people use every day. They have models that are really good that they made themselves. They have inference where you can pay them to use those models and generate things for you as well as with other models. And they're building their own chips that are going very, very well for them. Not just their quantum [ __ ] which is also going insanely well, but their own chips that are focused almost entirely on doing inference. and they're pretty good. It's a big part of why Google's **[08:46]** models fly the way that they do. I've also heard rumors that they don't have anywhere near enough of said chips and they're actually relying on CPUs for a lot of work they're doing internally, as funny as that sounds, does not seem like Google's a big Nvidia house beyond hosting for things like Claude. But they are building their own chips that are really good. They have crazy vertical integration here, but that also means that their investments are spread thinner than at other companies. There's just less alignment between the different pieces. And I've heard of a lot of drama between like the AI dev **[09:16]** side of Google versus the deep mind side versus AI studio versus Vert.Ex. These are all different teams with different metrics with different OKRs with different goals in different buildings in different states sometimes and that's making it harder for them to intersect and overlap and build things together which is more and more essential. Now, Enthropic's new model releases are cool and all, but they're a lot cooler once you see how they work in cloud code. OpenAI's model releases are cool as **[09:46]** well, but they're way cooler when you see the cool integrations they're building with things like Sora in the Sora app or in the crazy features they're putting into chat GPT. There is some amount of this going on at Google. They've been doing a lot of weird side quests like the new file search tool, which is a rag solution they built for the Gemini API that will simplify grounding models with your private data to deliver more accurate and verifiable responses. This is a new thing they're offering for their models that uses their internal Gemini embedding model in order to find all of your data and **[10:18]** manage that via API so that an LLM can use it to be more likely to be correct. That's really cool. There's a lot of demand for this. I know that because Raggy is one of my favorite sponsors and these guys have been building this for a bit now. So, if you want to have access to all of your users private data for the LLM when they want it without having to manage all of that safely and securely, Raggy's a good bet. They're not paying me for this spot. I just wanted to mention them because I think they're interesting in this regard. Check them out for sure. Yeah, cool to see Google doing these types of things. But what's really interesting, this **[10:48]** person's role is AI dev experience at DeepMind, huh? So what's going on there? My suspicion is that this sub team within DeepMind that's doing developer experience stuff, their resources aren't for building APIs or implementing crazy features, they don't have all that edge effort, but they do have people that know how to train models. I suspect this is a product that is happening as a result of the language model training they're doing with this embedding model. I honestly think that when your team **[11:20]** only has hammers, everything starts looking like nails. And when you're on DeepMind and the majority of the team is there to train new models, even the APIs you're giving out are kind of models disguised as an API. There's no other way for things like this to be made though because who would make it? Would this be the AI Studio team's problem? Would this be Vertex's problem? Is this general Google Cloud? Is this something that the Jules team should touch? Like who owns this is a very consistent problem that Google has right now. And it's gotten bad enough that I know a handful of people that got tired and **[11:50]** left because it's just such chaos lately. There are a couple people I still trust at Google though, like Logan, who is all over the place. He's the lead for product at AI Studio, but he also tries to bridge all the gaps between these different things. The fact that AI Studio and Vertex are competing internally is very, very annoying, and I could do a longass rant about that. It seems like they're trying to bridge the gap a bit more recently. Someone earlier today leaked that 3.0 0 Pro is now somewhat available on Vertex to a few limited regions in accounts as Gemini 3 **[12:21]** Pro preview 11 2025. And he put this screenshot says pay as you go. Has all this random information in it. Logan replied, "This is fake. I am inclined to trust him. There's no reason for him to lie about this. I see a lot of people saying that there's no way. It's so easy to fake a screenshot like this. I mostly trust him. But there is a catch which is again AI Studio which is his job and Vert.Ex which is what the screenshot is from are two different teams working from two different buildings in two **[12:52]** different states doing entirely different [ __ ] which makes it really hard for all of this to overlap properly. All of that said there is actually good evidence that Gemini 3 can be accessed now by some people in some ways. Here's one of those examples. This is somebody on the Singularity subreddit saying that it's real that you can access Gemini 3 Pro in the CLI and their evidence is honestly pretty compelling. If you enter a random model name when you do the Gemini-model command, it doesn't let you prompt and you'll get an error code that says requested entity **[13:22]** not found. But if you do Gemini-model Gemini 3 Pro preview 112025, it does let you prompt. The screenshot below shows the model thinking and then me being rate limited on it. So you can't, which like you wouldn't be able to do for a fake model. You know what? Let's try this. Gemini 3 Pro preview 112025. Let's ask it to generate an image studio for me. Requested entity was not found. What if I use a different model name? I'm changing it to 2027 to see what **[13:52]** happens. Same error. Okay, it appears as though this is not the case anymore. This person either faked it or they have since restricted this. But where things get much more interesting is when you look at the results of generating this example page. This is the generation it did with 2.5 Pro, which like looks fine. It's not the worst thing. It's not good. When you compare this to the results you **[14:22]** get with three pro, you get an entirely different look and vibe. And like these little pills like that, not something I've ever seen Gemini 3 do. I happen to believe that a lot of this new style stuff is coming from a set of training data that a specific data company is selling to a lot of these labs and that this data didn't exist in the previous Gemini training set which is a big part of why the UI should meaningfully improve with this **[14:54]** release. But I want to be clear, this is all skepticism edit. In the time it took me to post this, it seems that Google has blocked it, hearing the same from other people, too. Yeah, I think it's there. I would be really surprised if there wasn't a Gemini 3 snapshot or 10 floating around across Vertex and AI Studio and all the other places internally. And I honestly believe they're just running into random [ __ ] that's causing the releases to be delayed. Be it fear from other benchmarks, weird behaviors they're **[15:26]** seeing, copyright issues, which is an increasingly common one that a lot of these model companies are running into before they do their releases. There's a lot of reasons that a release could be delayed and that a company the size of Google with the amount of split ownership that Google has. The likelihood that this model is ready but isn't ready across everything else. Yeah, it makes sense. Quick fun find from my chat. There is a recent update to some JSON test files for the Gemini CLI that includes Gemini 3.0 Ultra is **[15:57]** one of the models that is available. Just thought that was interesting. It's definitely coming and different people have access to different things. No one knows what snapshots exist and which one's going to ship. But uh this is weird. I have a really bad feeling that the moment I hit publish on this video, the model's going to appear. But honestly, that's kind of why I'm making it. Hopefully by saying this model is in a weird state and doesn't exist. The moment I hit publish, I am influencing the whole universe to make it happen. So, uh fingers crossed this is about to **[16:28]** drop. Probably isn't. We'll know in a couple weeks, maybe a couple months. No one knows for sure, even my friends at Google. But I'm curious, what do you guys think? Is this weird? Is this coming? Or are we just getting hyped for something for no reason? Let me know what you'all think. And until next time, peace nerds.