GPT-5.3 Release This Week, GLM-5 Model LEAKED, Claude Cowork Mobile!

OpenAI might be dropping a new model this week. According to internal messages from Sam Elman that were leaked to CNBC, the company is preparing to launch an updated chat model within days. Now, Elfman didn't explicitly say GPT 5. 3, but the timing lines up with something interesting. Last week, OpenAI quietly released GPT 5. 3 Codeex, a new coding focused model along with this standalone app for Mac users. if they're launching an updated chat model this week, there's a good chance we're looking at GPT 5. 3 for general use. So why now? Open AAI is under pressure. Chad GPT has 800 million weekly users, but Anthropic and Google have been gaining ground. In December, OpenAI declared a code red and pulled developers off other projects to focus on improving chat GPT. They needed to respond to Claude's recent improvements, particularly in coding, and it seems to be working. Elman told employees that chat GBPT is back to exceeding 10% monthly growth. More importantly, their coding product, Codex, apparently grew 50% in a single week. Elman called the growth insane in the internal Slack message. That 50% jump matters because Codeex competes directly with Entropics Cloud Code, which has been eating into OpenAI's market share over the past year. According to internal charts that OpenAI has been showing investors, Codeex is now clawing back market share from cloud code. That's significant because coding has become one of the most important use cases for these models. Here's what we know about GPT 5. 3 Codex so far. It launched last week specifically for coding tasks. And if the pattern holds, a general purpose GPT 5. 3 chat model would follow shortly after. That's been OpenAI's release strategy before. specialized models first, then the broader version. The timing also makes sense from a business perspective. Open is in the middle of closing what could be a hundred billion dollar funding round. Sam Elman and CFO Sarah Frier have been meeting with investors and they need momentum. Launching a new model this week would give them concrete evidence that they're not losing ground to competitors. The funding itself is massive. Microsoft, Nvidia, and Amazon are all expected to participate with Amazon reportedly in talks to invest up to $50 billion. Soft Bank is discussing another $30 billion contribution. The round could close in two parts over the next few weeks, and having a new model launch right now would strengthen OpenAI's negotiating position. There's also the Anthropic factor. Last week, Enthropic ran Super Bowl ads that made fun of OpenAI's decision to put ads in chat GPT. Alman fired back on social media calling the commercials deceptive. OpenAI is starting to test ads in chat GPT on Monday. So launching a new model the same week would shift the narrative back to product quality rather than monetization. What could GPT 5. 3 actually improve? Based on the codeex release, we would expect better reasoning, longer context handling, and faster response times. The model would likely build on whatever architecture improvements they made for the coding version. And if they've been in code red mode since December, they've had about two months to iterate. One thing to note, Alman said in his message that this was a great week, referring to the codeex launch and growth numbers. Launching the chat model this week would keep the momentum going and give OpenAI a solid narrative heading into the next phase of fundraising. We'll know soon enough. If Elman's internal message is accurate, we could see the announcement within next few days. Whether it's called GPT 5. 3 or something else, OpenAI clearly needs a win right now and a new model launch is their most direct way to

get one. GLM5 from Zoo AI has been spotted on GitHub. References to the model showed up in ZA's repositories and according to the leaks is expected to drop sometime in February. The model is around 745 billion total parameters, which is roughly double the size of GLM 4. 5, which sits at 355 billion, but it only uses about 64 billion active parameters per token because it's running a sparse mixture of experts architecture. What's interesting here is that GLM5 appears to be using Deep Seek's architecture directly. A poll request number 34124 shows it. You're reusing Deepseek's version 3 and version 3. 2 setup, including the sparse attention mechanism and multi-token prediction. They essentially took Deepseek's design and scaled it up. The architecture uses dense sparse attention with 78 hidden layers, which is more than Deepseek's versions 3. 2 at 61. hidden size is 6,44 and it runs 256 routed experts with eight active per token which is how they keep the active parameter count at 64 billion while having that 745 billion total capacity. For comparison, DeepSeek version 3. 2 is 685 billion total with 37 billion active. Kimmy K2 is around 1 trillion total with 32 billion active. And GLM 4. 5 is 355 billion total with 32 billion active. So GLM 5 sits between Deepseek and Kimmy in scale, but has more active parameters than both. According to the leaks, there's no multimodal support yet. So no vision capabilities at launch, just language processing and long context handling. What this tells us is that Chinese AI labs are iterating quickly. They're taking proven architectures like Deep Seeks, scaling them, and releasing at a faster pace than the American models. GLM 5 in February means more options in the open weight space and more competition overall. I'll be watching for the official release and testing it when it drops. So, this model is actually being tested on Arena right now underneath the name Pony Alpha. And we see this because it kind of proves itself by showing a Chinese phrase called pour vegetable oil into the pot and heat it which was something specific to GLM4. And I guess that is why people figured out it's GLM5 and people are actually been testing this model itself in like kilo code and everything like that. So some people are saying that the quality is opus level stuff happening here. So people are actually excited about this model. What other people have been saying is that the inference is a bit May, but the output is second only to Codeex right now. My excitement is through the roof. So, this model is going to be very capable. So, it's going to be another model from China that might take the leaderboard like Kimmy K2 has been currently. There are some criticism as well. People like this model, but unfortunately, it's much more censored than version 4. 7. Otherwise, though, it's a great model. If I had more context, it would be absolutely fantastic. So, that's another complaint that people are noticing. And then some people are also saying that this model is really good in my testing, too. And it can replace Kimmy. So, the Kimmy model right now is probably the leading model from China and open-source model that is from China that is, you know, taking the leaderboards and is making people consider switching to Chinese models. So if this model GLM5 is that good, we might see another option for people who are looking for a cheaper alternative to American models and want something that's open sourced to build on. So GLM5 might be the model that takes away all this attention. However, we know that Deepseek was supposed to drop something near Chinese New Year which is coming up as well. So let's see what happens. Does GLM5 take the lead or does Deep Seek come in and completely overtake the market? Finally, Enthropic is working on

a task mode for cloud mobile apps. This comes from testing catalog news, who posted a screenshot showing the feature in development. The screenshot shows a new task tab in the mobile interface sitting alongside existing artifacts and the code tabs. The main area displays your task will show up here, which shows that this is still early in development. What's interesting is the mention of mobile co-work is coming. Co-work is Enthropic's desktop tool that lets Claude automate file and task management. It's currently in beta for desktop users, but bringing that functionality to mobile would be a significant expansion. Task mode would likely let users assign multi-step workflows to Claude on their phone. Think things like research these three topics and compile a summary or draft an email and schedule a follow-up. The kind of work that requires Claude to remember context and execute several actions in sequence. Who knows, we might also get a demo of mobile co-work cleaning up your photos for you or organizing them on your phone. The fact that it's showing up in the mobile app is very important. Most people interact with AI on their phones now, not just desktop. If Enthropic can make Claude handle complex tasks from mobile, that's a completely different use case than just answering questions in a chat interface. There's no word yet on when this will actually ship. And this screenshot shows it's in development, but there's no timeline. Given that desktop co-work is still in its early stages, mobile task mode is probably weeks or months out, not days. But I'll keep an eye out on this one. 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