• AI Emergence
  • Posts
  • đź’¸o3-Pro Is Here - More Reasoning at Lower Cost

đź’¸o3-Pro Is Here - More Reasoning at Lower Cost

Along with: Apple Holds Back on AI at WWDC

Hi there đź‘‹

You open ChatGPT, type in your prompt, hit enter… and nothing. Maybe it’s your Wi-Fi? Maybe a refresh will fix it? Nope. This time, it was ChatGPT itself. A traffic spike knocked multiple services offline for hours - a small outage, but a reminder of how quickly these tools have become part of daily life.

While that was happening, Sam Altman dropped what he’s calling his last handwritten post, laying out where AI is headed next: smarter agents that don’t just answer questions but handle full tasks, make decisions, and improve themselves over time. The goal? A world where AI quietly runs more of what we do behind the scenes - but how ready are we for that kind of dependence when things still break this easily?

Here’s what else moved this week. 🚀

What would be the format? Every week, we will break the newsletter into the following sections:

  • The Input - All about recent developments in AI

  • The Tools - Interesting finds and launches

  • The Algorithm - Resources for learning

  • The Output - Our reflection

Table of Contents

OpenAI has released o3-Pro, its newest and most advanced model for reasoning. It’s now available for all ChatGPT Pro and Team users. Enterprise and Education access is coming soon.

What’s New:

  • Smarter Tools: o3-Pro can search the web, read files, run code, and analyze images. It’s built to handle more complicated tasks than regular models.

  • Better Reasoning: In tests, o3-Pro did better than o3 at math, science, coding, writing, and problem-solving. Experts say it gives more accurate and detailed answers.

  • Slower but Smarter: o3-Pro focuses on giving better answers, even if it takes a little more time to respond.

  • Big Price Cut for o3: The regular o3 model is now 80% cheaper - $2 per million input tokens, $8 per million output tokens. o3-Pro costs $20/$80 for the same.

  • Some Features Still Coming: o3-Pro doesn’t yet support image generation, Canvas, or temporary chats, but these are planned for future updates.

o3-Pro gives developers a stronger reasoning model with better tools, while making regular o3 much cheaper for everyday use. It’s another step toward building AI agents that can handle more real work (source)

Apple’s WWDC 2025 is still happening, and so far Apple has taken a careful approach to AI. Instead of big new breakthroughs, they’re focusing on small improvements that protect privacy and make everyday features better. Most of the updates are about AI running on your device, keeping your data safe while helping you get more done.

What’s New(so far):

  • Developer Access: Apple’s foundation models are now available through the Foundation Models framework - on-device, privacy-first, Swift-native.

  • Real-Time Translation: Live translation added to Messages, FaceTime, and Phone - fully processed on-device.

  • Image Playground & Genmoji: AI-powered emoji customization and image generation with new ChatGPT-powered styles (oil painting, vector art, etc).

  • Visual Intelligence: Smarter, context-aware actions across apps - screen scanning, data extraction, and integrated search across platforms like Google, Etsy, and more.

  • Workout Buddy (Apple Watch): Real-time coaching using AI-generated dynamic TTS voices based on Fitness+ trainers.

  • Shortcuts & App Integrations: Deeper AI-driven automation in Mail, Reminders, Messages, Wallet, and more.

  • New OS Naming: Software updates are now aligned with calendar years - iOS 26, macOS 26, watchOS 26, etc.

Apple continues to expand AI across its ecosystem but remains focused on smaller, device-level updates rather than full-scale AI agents or foundational model breakthroughs. The bigger agentic moves seen elsewhere (OpenAI, Google, Mistral) aren’t Apple’s priority yet. (source)

French startup Mistral just dropped its first reasoning models, Magistral Small and Magistral Medium, marking a cool step forward for Europe’s AI scene. These models are built to handle step-by-step logic across math, physics, and business tasks, using more computing power during inference to boost accuracy.

What’s New:

  • Two Models:

    • Magistral Small (24B params) - open weights, Apache 2.0 license, downloadable via Hugging Face here.

    • Magistral Medium - more capable, now in preview on Mistral’s Le Chat, API, and partner clouds.

  • Enterprise Angle: Targeted at structured calculations, programmatic logic, decision trees, and rule-based systems.

  • Speed Play: Mistral claims 10x faster response times vs competitors on Le Chat.

  • Multilingual Support: Covers Italian, Arabic, Russian, Simplified Chinese, and more.

  • Benchmark Check: Magistral Medium still lags leaders like Gemini 2.5 Pro and Claude Opus 4 on GPQA Diamond, AIME, and LiveCodeBench.

Backed by €1.1B in funding (General Catalyst, others), Mistral is steadily building out its enterprise AI stack, following recent launches of Mistral Code and Le Chat Enterprise. (source)

Google just dropped an upgraded preview of Gemini 2.5 Pro, its most capable model yet - ahead of full general availability in a few weeks. The latest version sees a 24-point Elo jump on LMArena (now 1470) and leads WebDevArena at 1443, with standout performance in coding (Aider Polyglot), GPQA, and Humanity’s Last Exam.

Key improvements include better style, formatting, and creativity based on prior user feedback. Developers can start building today via the Gemini API in Google AI Studio, Vertex AI (now with thinking budgets for cost/latency control), and the Gemini app.

OpenAI’s ChatGPT platform experienced a major outage Tuesday, with widespread errors and latency across ChatGPT and the API.

What Happened:

  • Outage began Tuesday morning, impacting ChatGPT and API services globally.

  • Root cause was identified and mitigated within hours.

  • Full recovery is now complete across all impacted services.

 OpenAI will publish a detailed Root Cause Analysis (RCA) within the next 5 business days.
Live status updates remain available on OpenAI’s status page.

Hugging Face just dropped Sheets, blending spreadsheets with AI-powered data transformation - designed for working directly with unstructured data using open models.

What’s New:

  • AI-Powered Spreadsheets: Run prompts and models directly on your data.

  • Agentic Search: Pull in real-time, accurate information.

  • Simple Interface: Clean, minimal UI built for data interaction.

  • Human Feedback 2.0: User input refines and improves outputs.

  • Open Model Access: Tap into hundreds of open models and inference providers. (source)

Try it out here: hf.co/aisheets

🎥 Veo by Google DeepMind

Google's new Veo lets you turn text prompts into short AI-generated videos - perfect for creators, demos, or pure fun. Currently available to Gemini Pro & Ultra users via gemini.google.

How to use it:
Step 1: Sign in at gemini.google (Pro/Ultra access required).
Step 2: Paste your prompt (swap in any names or details).
Step 3: Hit the video button - done!

Example Prompt (used to generate this video):
A marmot sits on a green grassy mound. It speaks with a distinct, slightly gravelly English accent, then begins to narrate. The accent is quite pronounced, with clear, clipped consonants and a fairly relaxed vowel pronunciation and never rolls their R sounds or has any presence of other accents. The speaker's tone is somewhat dry and contemplative at first but increases its franticness, urgency and pitch toward the end. "Robert! Rob! Robt! Rob! Roberto! Robby! Robert! Robert! Rob!". Each time the name is delivered with increasing speed, urgency and volume and comes out all at once, frustrated that Robert can't hear it.

  • In The Gentle Singularity, Sam Altman lays out a candid, long-term vision of AI’s accelerating progress - from today's cognitive agents to a future of abundant intelligence, recursive self-improvement, and massive productivity shifts. A thoughtful roadmap on how the singularity might actually unfold.

  • Google DeepMind CEO Demis Hassabis visited the Institute for Advanced Study (@the_IAS) for a deep dive into AI’s role in advancing mathematics, highlighting their ongoing collaborations and future potential. Great fireside chat with IAS director David Nirenberg. Watch here.

  • This week, we’re spotlighting a full Data Science Program-a free collection of 9 courses designed to build real-world skills across the entire data stack. Perfect for anyone looking to break into data science or sharpen their technical edge.

The capabilities keep improving - models like o3-Pro are getting better and cheaper. But outages and platform caution remind us that being able to do more isn’t the same as being fully ready to rely on it. The real test isn’t just how smart these systems get, but how stable and dependable they become as we keep building on top of them.

đź‘‹See you next week!

Reply

or to participate.