OpenAI Just Released a Mind That Can Think on Its Own + Gemini 2.5 Flash, Embed 4...

So OpenAI dropped its brainiac duo 03 and 04 mini plus the slick codec cli, google rolled out budget, tuned gemini 2.5 flash cohere unleashed embed 4 for next gen multimodal search, and Microsoft made Copilot Vision free in Edge. We're diving into how each one works, why they matter, and which you need to try first. So first up, OpenAI just unveiled O3 and O4 mini. Now, if you've been using ChatGPT, you might have noticed that sometimes it really feels like it's thinking longer before it speaks. And that's by design. O3 is their most powerful reasoning model yet. It's been trained to think deeper, combine tools, agency and deliver highly detailed answers in under a minute. We're talking web search, Python code execution, file analysis, you name it. And on top of that it reasons about when to use each tool so you get precise, thoughtful answers without having to babysit it. The step change shows up all over the benchmarks for complex math problems like the Aime O4 mini, despite being smaller and cheaper, hits 99.5% pass at 1 on the 2025 exam when it's got Python access and 100% consensus at 8. O3 isn't far behind at 98.4% pass at 1. Think about it, with just a bit of code to work with, that's nearly perfect. On CodeForces, the Elo rating for 0.4Mini high comes in at 2,719 compared to O3 high at 2,000 706. And if you're into PhD level science questions, deep research tasks, multimodal benchmarks like MMMU or Scientific Figure Reasoning on Charzive, these models are leaving their predecessors in the dust. O3 cuts down major errors by about 20% compared to O1. Visual perception tasks especially light up O3 nails 86.8% accuracy on MMMU versus O1s 71.8% and math vision puzzles see a jump from O1's 55% to O3's 78%. What's really wild is the agentic capabilities. Imagine asking how will summer energy usage in California compare to last year? And watching the model chain together a web search, fetch public utility data, write Python to forecast usage, generate a graph, then explain the key factors all autonomously. It can loop searches, pivot as it sees new info, and keep thinking with images, rotating and zooming them as needed. Talk about next level under the hood. OpenAI scaled up the reinforcement learning compute by an order of magnitude and they retraced that scaling path for inference time reasoning more compute still means better performance. They also rebuilt their safety training data from the ground up new refusal prompts around BioRisk malware jailbreaks and they've layered on a Reasoning LLM safety monitor that flags suspicious behavior with 99% success in red teaming tests. Plus they've run both models through their preparedness framework across bio, cyber and AI self improvement risks and they're still below the high threshold. So as these models get sharper, the safety foundations are getting firmer too. You can try O3.04 Mini and the O4 Mini High variants today if you're on ChatGPT Pro or Team Enterprise and EDU users get it in about a week and free tier users can even dabble with O4 mini by hitting think before your prompt. Developers can call them via the Chat Completions API and Responses API complete with reasoning summaries and and soon to come built in tools like web search and code interpreter. And just when you think you've got it all, OpenAI drops Codec CLI, a minimalist coding agent you run locally. Picture a terminal interface that can reason with your code, take in screenshots, lo fi sketches and hook into your machine directly. It's Open Source on GitHub, and they're teeing up a $1 million grant program to get community projects going, handing out 20 $25,000 increments in API credits. So whether you're an enterprise architect or an indie dev, you've got some serious new toys to play with. Switching gears let's talk about Google, because yesterday they rolled out Gemini 2.5 Flash. The headline here is Thinking Budgets. You can now dial in how many reasoning tokens you want the model to use, anywhere from zero up to a whopping 24,576 token tokens. Why? Because deep reasoning costs more compute, and compute costs money and time. So for simple stuff like translations, you turn thinking off and pay just 0.6 cents per million output tokens. But for heavy lifting, complex engineering questions, multi step logic you crank the thinking back on and it's 3.5 cents per million. Input tokens stay at 0.15 per million. That six fold price swing is no accident. Google is being super transparent about where the cost really lies in the thinking phase, where the model evaluates different solution paths in AI Studio's ui. You can even peek at those hidden internal thoughts on the API. You can't see the text, but you can watch the token count go up and down. Performance wise, Gemini 2.5 flash punches above its weight on humanity's last exam. It scores 12.1% ahead of Anthropic's Claude 3.7 Sonnet at 8.9% and Deepseek R1 at 8.6%, though it trails OpenAI's O4 mini at 14.3%. For technical benchmarks, it nails GPQA diamond at 78.3%, and Math Performance comes in at 78% on the 2025 AIME and 88% on the 2024 version. Google's pitch is that when you factor in speed and cost, it's the best value out there, especially for enterprise clients who need budget predictability. They're previewing it now in Google AI Studio and Vertex AI, and they've paired this with a couple of other moves. First, they just launched VO2 video generation for Gemini Advanced subscribers, eight second clips from text prompts. Second, US college students get free Gemini Advanced access until spring 2026, which is a clear play to lock in the next generation of AI talent. And for consumers, the Gemini app now lists 2.5 flash experimental in the dropdown, replacing the old 2.0 thinking option. It's Google's way of gathering feedback from real users while they fine tune things before general availability. Now, speaking of enterprise, Cohere just brought out Embed 4, a multimodal search engine that aims to be the foundation for any agentic AI app doing retrieval augmented generation. You know how Enterprises wrestle with PDFs full of charts, tables, code snippets and embedded images. Embed 4 lets you index up to 128k tokens of that roughly a 200 page annual report in one go without splitting it up. It's multilingual out of the box, with Support for over 100 languages and including Arabic, Japanese, Korean, French, you name it. And it's tuned for regulated industries, finance, healthcare, manufacturing, so it gets investor presentations, clinical trial reports, product spec docs, repair guides. You get the picture. The embed vectors come compressed binary, NT8, even FP32, so you can shrink your storage footprint by up to 83% and still hit top quartile 8 NDCG at 10 scores. Customers are already seeing big gains. Hunt Club said they saw a 47% relative accuracy boost over Embed 3 when searching complex candidate profiles. Agora, an AI powered shopping engine, said their product search got way better at surfacing the right items from tens of thousands of stores. And because it's robust to real world noise scan docs, handwriting, landscape pages, it slashes the need for those hacky pre processing pipelines that always break on weird PDFs. Embed 4 is live now on Cohere's own platform and you can spin it up in Microsoft Azure, AI Foundry, Amazon SageMaker, or even privately on prem in a VPC. It also plugs into north, their secure AI agent runner, powering the Compass search layer so you can build end to end agents that reliably fetch data from your own vault. Finally, let's talk Microsoft Copilot Vision, because they just made it free for everyone using the Edge browser. Previously, you needed a Copilot Pro subscription to share your screen content with Copilot Vision. Now if you're on the latest Edge, you hit the mic icon in the browser, point Copilot at Amazon or Target or Wikipedia or TripAdvisor, and it'll parse what you see and answer your questions. It won't work on paywalled or sensitive sites, and it's entirely opt in. Microsoft isn't harvesting your images, audio or conversation for model training. It's a privacy win. But that's not all. Earlier this month, they rolled Copilot Vision into their Mobile and Windows apps. On mobile, you can point your phone camera at, say, the coffee machine instructions or a weird street sign, and Copilot interprets the live video or your saved photos. On Windows, Insiders can share any app window via a little glasses icon in the Copilot Composer and ask questions. Before long, we'll probably see it roll out to more Windows users, and combined with Edge, it means that anyone for free can tap into an AI that sees and explains. That's a huge step towards seamless multimodal interaction for everyday browsing. And that's a wrap on today's AI overload. Tons of powerful tools landing in your hands. Dive in, experiment and let me know which one blows your mind first. Thanks for watching and I'll catch you in the next one.

Sam.