00:00All right, so a lot just dropped in the AI space and honestly it's one of those
00:06weeks where everything is shifting at once. We've got OpenAI finally launching
00:10codecs inside ChatGPT, Manus AI stepping into image generation in a way that
00:15actually feels intelligent, Google opening up more about its search future,
00:20and Anthropic silently laying the groundwork for what looks like their
00:24biggest Claude upgrade yet. There's a lot going on so let's talk about it.
00:29All right, so let's get into Codex, OpenAI's new software engineering agent that
00:35just dropped as a research preview. If you're using ChatGPT Pro, Team, or
00:40Enterprise you've probably already seen it pop up in the sidebar. Codex works more
00:44like a full-stack dev running inside a secure cloud-based sandbox. It operates
00:49entirely within its own isolated environment. No internet access, no
00:53external API's, nothing leaking out. You just connect it to your github repo and
00:58from there it starts handling actual engineering tasks without needing
01:03constant supervision. You can throw at tasks like writing new features, fixing
01:08bugs, running tests, cleaning up messy code, or even digging through your code base to
01:12answer specific questions. It spins up a virtual environment, loads everything in,
01:17and then handles the task from end to end. That includes setting up and running
01:21tests, applying linters, checking types, the whole routine. And the best part is you can
01:26watch it work in real time. Perminal logs, test results, status updates, they're all
01:31right there so you always know what's going on. The model running behind all
01:35this is Codex 1, which is a specialized version of OpenAI's O3, fine-tuned
01:41specifically for software development. They didn't just train it on general
01:45purpose data, they used reinforcement learning on real coding tasks, pull
01:50request patterns, and team workflows. The result is a model that writes clean,
01:54structured code, understands project layouts, and mirrors the way human
01:59engineers actually work. In internal benchmarks, it hit 75% pass-at-1 accuracy
02:06on SWE bench verified tasks. That's a noticeable jump over O3 high, which
02:11landed at 67%. And what's nice is it doesn't need a ton of configuration to be
02:17useful. Sure, you can give it agents MD files to help it navigate your repo more
02:22efficiently, but even if you don't, it still figures things out. It respects your
02:27architecture, follows your naming patterns, and can juggle multiple tasks without
02:32stepping on its own toes. Once Codex finishes whatever you've assigned, it
02:36doesn't just give you some output and call it done. It commits the changes
02:41directly within its sandbox and includes logs and references so you can trace
02:45exactly what it did and why. From there, you've got options. You can review the output,
02:50make tweaks, turn it into a pull request, or pull it down locally and keep working
02:54from there. If you prefer working from the terminal, Codex CLI is probably the
02:59better fit. It's the open source version you run locally, and now it uses Codex
03:03Mini by default. That's a smaller, faster model based on O4 Mini, optimized for low
03:09latency workflows. It's great for everyday tasks, renaming variables, writing test cases,
03:14refactoring functions, stuff that takes time but doesn't require deep focus. You
03:20can leave it running in your terminal almost like a quiet assistant who's
03:24always ready to help out when things get repetitive. For API usage, pricing is
03:29pretty straightforward. Codex Mini costs $1.50 per million input tokens and $6 per
03:36million output tokens. Plus there's a 75% discount on cached prompts. So if you're
03:42repeating similar tasks, the costs drop significantly. Codex is part of OpenAI's
03:48push to turn ChatGBT into a workspace built around task-specific agents. There's
03:54operator for browsing, Sora for video, deep research for analysis, and now Codex for
04:00software development. Access to Codex is generous for now, though rate limits are
04:04coming. The idea is simple. You assign real coding tasks and Codex handles them
04:08while you stay focused. It's built to feel like part of your team, understanding
04:13your project, following your standards, and quietly taking care of what slows you
04:17down. Alright, now while OpenAI keeps expanding its agent lineup inside ChatGBT, over
04:23in China, something wild just happened. Maness AI, the autonomous agent from
04:28Monica, also known as Butterfly Effect AI, has introduced an advanced image
04:33generator that's on a completely different level. And it's not just another model that
04:37turns prompts into pretty pictures. It's a full-blown visual problem solver
04:41built right into an autonomous agent framework. Let's say you ask it for a
04:46modern Scandinavian living room. Maness won't just throw together random
04:51furniture. It first analyzes your intent. Are you designing a catalog, creating ad
04:55visuals, or drafting a room layout? Then it builds a strategy. It uses layout engines to
05:01arrange space, style detectors to match the look, and browser tools to pull design
05:06trends or brand guidelines. It might even select real IKEA furniture, consider
05:10spatial relationships, apply color theory, and ensure everything fits the purpose.
05:14The system is built on a multi-agent architecture, where separate modules
05:20handle planning, execution, and verification. They run independently but
05:24collaborate like a design team, allowing Maness to work through complex
05:28workflows, not just one-off prompts. That's why it can deliver things like
05:33product campaigns, architecture mock-ups, or platform-ready visuals. All consistent,
05:38brand-aware, and usable. It's already being tested for e-commerce, product
05:42visualization, marketing content, and even architectural planning, like generating
05:47full interiors from blueprints. The big limitation? It's still in closed beta and
05:53available by invitation only. So, unless you're part of a select test group,
05:57you can't use it yet. Now let's shift to Claude. Anthropic's been relatively quiet,
06:02but behind the scenes, they're cooking up something huge. There's been a bunch of
06:06internal leaks about a new model possibly named Claude 3.8 or Claude 4. We're seeing
06:12names like Neptune appear in their config files. And yeah, Neptune being the eighth planet is
06:18probably not a coincidence if we're talking versioning. Publicly, they denied the rumors
06:23after that win four months of mass contest went viral, but come on. There's too much back-end
06:27evidence pointing to something real. One of the standout leaks even showed internal tools with
06:32redacted model names and Easter eggs tied to upcoming versions. And the information confirmed.
06:38Anthropic is prepping upgraded versions of both Claude Sonnet and Claude Opus. The big thing with this
06:45new wave of Claude models is what Anthropic calls true agentic behavior. That means the model can
06:52switch autonomously between reasoning and acting without user prompts. It doesn't just generate
06:57an answer in one shot. It breaks down the problem, builds a plan internally, then switches into action
07:03mode to call tools, search data, or run code. If something goes wrong mid-tack, it backtracks,
07:09rethinks, and tries again. That's a real agent, one that reasons like Gemini, but potentially with more
07:15precision in task delegation. And this actually mirrors what OpenAI's O3 model already does inside
07:21ChatGPT, where it can browse, run code, and iterate before showing you the final result. But Anthropic's take
07:27might offer better transparency or control depending on how they deploy it. For example, with the upcoming Claude
07:33update, developers might see the full breakdown of thoughts, tool calls, and revisions in the
07:39background, not just the polished final response. And they're not stopping there. Anthropic is also
07:45investing in making these agents work better with complex tool chains, possibly building integrations
07:50with search, databases, and APIs all inside one flow. That's a direct response to what Google's doing
07:56with its own AI mode inside search. DEO Sundar Pichai was recently on the All In podcast and the big
08:03question came up whether Google is being disrupted by ChatGPT, perplexity, and other AI native tools that
08:11are rapidly eating into traditional search behavior. Pichai didn't seem rattled at all. His take was that
08:18disruption isn't inevitable unless you ignore it. He sees it more as a shift, one that Google is
08:25actively adapting to rather than resisting. And the numbers show they're already moving. Over 1.5 billion
08:32users have engaged with Gemini-powered AI overviews inside Google search. These aren't just summaries
08:40or snippets. It's an AI layer baked directly into search results designed to give more context, answer
08:47follow-up questions, and reduce the need to click through multiple pages. It's a way to keep users in
08:52Google's ecosystem while still giving them something closer to an AI chat experience. But they're not
08:59stopping there. Google is preparing to launch something called AI mode which will turn search
09:04into a full-on conversational experience. It's not just query and result anymore. You'll be able to ask a
09:10question, get a response, follow up with more context, refine your query, and get deeper answers
09:15all inside the search interface. Basically, it turns search into a Gemini-powered assistant with memory
09:22across turns. And this isn't some distant roadmap item. It's already been confirmed and will be showcased
09:28in more detail at Google i slash o. That said, Google's position isn't bulletproof. Apple recently signaled
09:36it may replace Google search in Safari with a more AI native system, possibly its own or powered by
09:42another provider like OpenAI. That kind of move could hit Google hard, especially on mobile, where
09:47Safari holds massive market share. The moment that new surfaced, Google stock took a noticeable hit.
09:54Investors clearly understand what's at stake here. Still, Pichai isn't new to these shifts. He pointed out
10:00that people had similar concerns when mobile search took off and again when TikTok started pulling younger
10:06audiences away from YouTube. In both cases, Google adapted, integrated features from rising platforms,
10:12and kept their core products alive. Pichai's betting they can do the same again, this time by making
10:17search smarter, more conversational, and more useful than anything that exists in AI chat apps today.
10:25So while yes, Google is under pressure, especially from Apple and OpenAI, they're not standing still.
10:31They're building, integrating, and reshaping how search works to stay relevant in an AI-first world.
10:37Now the question is, with Cloud, Gemini, Codex, and Manus evolving fast, who's actually building the
10:43smartest agent right now? Let me know in the comments, drop a like if this gave you something to think about,
10:50and subscribe if you want to stay ahead of where all this is going.
10:53Thanks for watching, and I'll see you in the next one.
Comments