AI News: Gemini 3.5 LEAKED, GPT-5.3 CONFIRMED and DeepSeek R2?

OpenAI just confirmed GPT 5. 3 is in development. Deepseek dropped hints about a mystery model in their code. And Google's Gemini 3. 5 leaked through an AB test. And it's not looking great. The AI race is heating up. Models are leaking left and right. And we're breaking down what's actually happening versus what is just hype. So, let's get into it. Sam Elman just confirmed GPT

5. 3 is in the works. He tweeted asking for feedback on what's working and what needs fixing, which basically confirms that this version exists and people are losing their minds over this. But let's actually talk about what this means. The confirmation matters because Open AI usually doesn't say much about what they're building until it's ready. The fact that it's version 5. 3 tells us something. This is probably more of a incremental update, like how GPT4 got better over time with different versions and not a complete reinvention of the wheel. What are people hoping for, though? They're hoping for better reasoning, fewer mistakes, and maybe it actually follows instructions properly this time. Improved coding ability, better at handling long conversations without forgetting what you said 10 minutes ago. The usual stuff that would make these models actually more useful day-to-day instead of just impressive in demos. Here's the thing, though. Alman's asking what needs improvement, which means they're still working on it. This isn't a launch announcement and the. 3 suggests we're talking about refinements and not some massive breakthroughs. Every time a new model drops, people start talking about AGI being right around the corner. I don't think it's there yet. These are still sophisticated autocomplete systems. They're getting better for sure, but we're not suddenly getting science fiction AI in the next quarter. What you should pay attention to, though, is that when beta testing starts, that's when we'll get real information, actual benchmarks, and not marketing claims. and most importantly whether people building stuff with it can do things they genuinely couldn't do before. Also, better models usually cost more to run. That matters if you're actually trying to build something with this. So, yeah, GPT 5. 3 is happening and it'll probably be better, but temper your expectations and we'll know more when we actually see it. While

everyone's talking about GPT 5. 3, Deepseek just quietly dropped hints about something new in their code repository, and people are trying to figure out what exactly they're cooking up. Is it version 4? Is it R1. 5? Or something completely different. Testing catalog spotted a repository update mentioning a new model one model in Deepseek's Flash MLA library. For context, Flash MLA is the optimized attention kernel that powers DeepSeek version 3 and version 3. 2. The code changes show stride multiples of 576 billion for sparse FP8 decoding, which basically means they're building this for way more efficient large scale inference. That's not a small tweak, that's infrastructure for something big. So, what are people thinking? The obvious guess is Deepseek version 4. But here's where it gets interesting. Some are speculating it could be something new entirely. Remember, Deepseek's R1 model beat OpenAI's 01 in certain benchmarks and topped the iOS app store last year. They've proven they can compete at the highest level. The fact that they're optimizing for sparse FP8 decoding suggests their focus on making whatever this is run faster and cheaper. That matters because Deep Seek's whole thing has been efficiency, getting GPT4 level performance at a fraction of the cost. Here's why you should care. Deepseek operates differently than the US labs. They move fast, they're less hyped up, and they've been quietly eating everyone's lunch on efficiency. If this is version 4, it could be another leap forward. If it's a new R series model, we might even see better reasoning capabilities than R1. And the timing is interesting. Right as Open AI confirms they're working on GPT 5. 3, DeepSeek drops breadcrumbs about their next move. The AI race isn't just happening in San Francisco anymore. There are no official announcement yet, just code changes and speculation, but Deepseek doesn't usually update their repos for nothing. Could be version 4, could be something else. Either way, worth watching. Google's also got a leak

situation. A checkpoint for what looks like Gemini 3 Pro, possibly called Gemini Advance or GA, is being AB tested on AI Studio right now. And people are already testing it and posting results. Google started AB testing again in AI Studio. And one of the checkpoints being tested might be the Gemini 3 Pro G8. according to the leakers is far better, apparently good enough that Google can call it 3. 5 and people will accept it, which is a pretty interesting way to describe it. Testing catalog and others have been running it through benchmarks, including the robot SVG test. The outputs look solid, clean cyberpunk robot designs with good detail and consistency. Early testers are saying this feels like a meaningful step up from the current Gemini Pro models. better instruction following, cleaner outputs, more consistent results. The robot SVG bench in particular shows it can handle structured creative tasks well. Something previous Gemini models have been hit or miss on. But here's the thing about these AB tests. You're seeing a checkpoint, not necessarily the final version. Google could be still tuning this. The fact that they're calling it 3. 5 instead of 4. 0 suggests this is more of a refinement than a complete overhaul. This is becoming a pattern. Model leaks through AB tests. People benchmark them, post results on Twitter, and suddenly everyone knows what's coming before the official launch. It's not exactly the control narrative these companies want, and the timing matters. OpenAI just confirmed GPT 5. 3. Deepseek dropped hints about their next model, and now Google's testing Gemini 3. 5. Everyone's pushing updates at the same time. The race is heating up. There are no official announcements from Google yet, but the checkpoint's out there and people are testing it. If you have access to AI Studio, you might get randomly assigned to it. Worth checking if you want to see what Google's working on before they're ready to talk about it. So, here's

another example of somebody who's doing the early checkpoint testing. And he got that AB prompt as well in AI Studio. So, his verdict on the SVG generation, same performance as current Gemini 3 Pro. It's not better or worse, it's just the same. So, if you're hoping for a leap forward in code generation or visual outputs, that's not what this checkpoint delivers, at least at the moment. The robot SVG test from earlier looked decent, but when you compare it side by side with current Gemini, you're not seeing meaningful improvements yet. Then there's a Minecraft clone test. This is where it gets worse. The prompt result actually came out worse than the current checkpoint. You can see in the output is generating a basic Minecraft style world, but the quality and detail are below what current model produces. So for a checkpoint that's supposedly good enough to call 3. 5, having it perform worse than 3. 0 on certain tasks is not a good look. He also notes that the performance is worse than the initial Gemini 3 Pro checkpoints like ECP and K0 and the model is still very lazy, meaning it's taking shortcuts, not following through on complex tasks or giving incomplete outputs. He couldn't test agentic performance or tool calling because of AB test limitations, which is frustrating because those are the key areas where you want to see improvement in. But what he could test doesn't inspire confidence that this is ready for prime time yet. This is the problem with leak checkpoints. You're seeing Google's work in progress and is clearly not polished. Maybe they know these issues exist and they're using the AB test to gather data before fixing them. Maybe there are multiple checkpoints being tested and this is the worst one. But if this is what they're planning to call Gemini 3. 5, they've got work to do. You can ship something that performs worse than your current model and call it an upgrade. The checkpoints still live in AI Studio if you want to try it yourself, but based on early testing, you might want to stick with what you've got. Google's got some fine tuning to do before this is ready for sure. Make sure

to subscribe to our channel. We do real tests, not just headlines. Make sure you're also subscribed to the world of AI. And don't forget to check out our newsletter for deeper breakdowns you won't see on YouTube. And I'm growing my Twitter following, so make sure you follow me on Twitter as well. Hope you guys enjoyed today's video and I'll see you in the next