đŸ˜șThe new ChatGPT: o3-mini

PLUS: GPT's new research mode is WILD (30 min searches?!)

Welcome, humans.

Can you believe it’s only February? IDK about you, but January felt forever long. Maybe it was the onslaught of AI news, let alone news in general; maybe it was the new year blues; maybe it was the Kappa effect (otherwise known as perceptual time dilation), Chronostasis, or even the oddball effect, where encountering novel stimuli makes time seem to slow down; who’s to say?

Also, who knew there were so many explanations for this? Thanks, Perplexity!

Here’s what you need to know about AI today:

  • We tested out o3-mini + rounded up what the internet thinks of it.

  • OpenAI launched Deep Research, which lets Pro users conduct long searches.

  • The EU can now ban AI systems it deems harmful under the AI Act.

  • Google spun out startup that uses AI for crop genome analysis.

Initial thoughts on o3-mini AND OpenAI’s weekend live-stream surprise


As promised, OpenAI’s new reasoning version of ChatGPT, o3-mini, was released over the weekend, so of course we spent all three days trying it out and combing through the internet’s top demos of it in action.

First, the TL;DR on the model itself:

OpenAI says o3-mini is a specialized alternative to o1 for “technical domains requiring precision and speed” (think STEM, coding, and math).

  • Available now for ChatGPT Plus, Team, and Pro users (Enterprise access coming in February).

  • Works with search to find up-to-date answers (with source links!).

  • Free users can try it out by clicking “Reason” in the composer.

Now, what’s unique about o3 is that you actually get to choose how hard it “thinks” across three reasoning levels (low, medium, and high). These different thinking levels actually make a huge difference, allowing o3-mini to best o1 on certain benchmarks.

Speaking of benchmarks, this thing is impressive: 24% faster than o1-mini and 77% accuracy on PhD-level science questions.

Matt Berman compared o3-mini to DeepSeek on pricing and found them to be “extremely comparable”—what he calls the “DeepSeek effect”, where open source competition pushes prices down across the industry.

Here’s what the people think:

There was a lot riding on o3-mini being good, and by most accounts, it is. According to OpenAI, 56% of testers preferred it over o1-mini. Here are a few things it did:

  1. Create a water simulation in Blender (prompt).

  2. Solve hard Suduko problems.

  3. Generate procedural clouds (in one shot).

  4. Build an autonomous snake game, another with 100 snakes, and another where DeepSeek + o3 face off.

  5. McKay Wrigley replaced all his agents + workflows with o3-mini, and not only do they all still work—some even work better, for 9x cheaper and 4x faster.

And here’s the minecraft benchmark, comparing o1 mini vs o3 mini:

For those curious, here’s the prompt used.

Besides Minecraft, this might be our favorite use-case: having o3 explain this weekend’s OTHER wild news, the Luka Doncic trade, as if it happened in AI labs


Plus and Team users get 150 messages per day (up from 50 with o1-mini), while Pro users get unlimited access to both regular o3-mini and o3-mini-high. We’re also hearing free users get about 10 free searches a day.

Personally, we love o3-mini with search, though we’re still getting used to it. As we keep testing, we’ll provide more advice on how best to work with o3 based on what you need.

Because releasing o3-mini apparently wasn’t enough
 OpenAI also held a livestream on Sunday to announce the launch of Deep Research, their version of Google’s feature of the same name.

It works basically the exact same way—except once you ask a question and answer its follow up questions, you can then watch GPT “reason” as it conducts multiple searches to answer your question (sorta like an Operator session happening the background, without you watching your screen).

Oh, and the searches can take as long as 30 minutes to finish up, so make sure you have something else to do while you wait!

If you have ChatGPT Pro, you can use Deep Research with 4o starting today. Check out the launch video here to learn more.

What would you want to ask OpenAI Deep Research?

We have ChatGPT Pro, so we can test out Deep Research for you...what are some sample reports you'd like us to try?

Login or Subscribe to participate in polls.

FROM OUR PARTNERS

This FREE guide will help you win with AI


Top sales teams are saving 10+ hours per week with AI. So what’s their secret?

In our new, TOTALLY FREE guide (created in partnership with Attention) we’ll show you how to get similar or BETTER results—complete with 10+ ready-to-use prompts.

Plus, you’ll discover how Attention's AI assistants take it further—analyzing your calls in real-time, catching competitor mentions, and automating follow-ups.

Top-performing sales teams are getting amazing results, eliminating 90% of manual tasks and improving win rates by 70%, with these exact tools.

Prompt Tip of the Day

If you haven't used Canvas' new tools lately, you might want to give it another look—OpenAI just added support for o1 in Canvas as well as the ability to render HTML and React apps directly in chat.

That means you can now create interactive apps right in ChatGPT (think: quizzes, games, data visualizations) and everything renders directly in the chat window.

Quick example to try: Click “Canvas” in the bottom tool bar of your GPT window and ask: “Create an interactive quiz app with 5 multiple choice questions about [your topic]. Include a progress tracker and a final score display. Style it using shadcn/ui components with a clean, modern design.”

Treats To Try.

  1. Gemini AI can now help you create charts and uncover data insights about your spreadsheet data.

  2. Leonardo helps beginners edit AI images with simple tools to touch up, enhance, and make canvas edits without Photoshop or technical expertise.

  3. Rootly handles your technical emergencies by automatically gathering the right people and tools in one place.

  4. Open WebUI and AnythingLLM both help you chat with documents offline, but Open WebUI adds voice calls while AnythingLLM focuses on easy setup.

  5. Pimosa can edit all media files in one offline desktop app—compress a video, resize images, or merge audio tracks without switching between different tools—gotta buy a license, FYI.

  6. OpenRCP turns recipe websites into clean, ad-free pages—just paste any recipe URL and instantly get a minimalist, readable version you can cook from.

  7. ScamAI checks your screenshots and photos to instantly spot fake profiles and scam messages.

Around the Horn.

  • The EU can now ban AI systems it sees as containing “unacceptable risk” of harm.

  • There’s a new US bill in consideration that could make it illegal to download Chinese AI models like DeepSeek with a penalty up to 20 years, though it was only just introduced + has very little support at the moment.

  • A researcher discovered two instances of R1 (DeepSeeks reasoning AI) speaking to each other in a language completely made up of symbols.

  • Google spun out Heritable Agriculture, a startup using AI to analyze plant genomes for improving crop yields and developing climate-friendly traits, with successful tests across three states so far.

FROM OUR PARTNERS

Keep your social security number off the web.

Every day, data brokers profit from your sensitive info—phone number, DOB, SSN—selling it to the highest bidder. 

The question is, who’s buying

  • Best case: companies targeting you with ads. 

  • Worst case: scammers and identity thieves. 

If you want to protect your sensitive data, you need Incogni. It scrubs your personal data from the web, confronting the world’s data brokers on your behalf. 

Monday Meme

We just love this one so much


A Cat's Commentary.

That’s all for today, for more AI treats, check out our website.

The best way to support us is by checking out our sponsors—today’s are Attention and Incogni.

See you cool cats on Twitter: @noahedelman02

What'd you think of today's email?

Login or Subscribe to participate in polls.