ChatGPT just became your AI employee

Good morning, AI enthusiasts. OpenAI just launched ChatGPT Agent, giving their AI the ability to use its own virtual computer and complete complex tasks from start to finish.

With AI now capable of doing actual work instead of just answering questions, are we witnessing the birth of the first true AI employee?

In today’s recap:

• ChatGPT becomes your virtual assistant that actually works

• Nvidia releases free reasoning models that rival OpenAI's best

• AI model achieves picomolar protein binding

• OpenAI's secret model wins gold at math olympics

• 5 new AI tools & prompt

LATEST DEVELOPMENTS

OPENAI

🤖 ChatGPT gets virtual computer powers for real work

Recaply: OpenAI just launched ChatGPT Agent, a major upgrade that lets the AI assistant use its own virtual computer to complete complex tasks from start to finish. The system combines web browsing, research analysis, and conversational abilities into one unified tool.

Key notes:

• ChatGPT can now navigate websites, run code, create spreadsheets, and build slide presentations without human help.

• Users activate "agent mode" through the tools dropdown and maintain full control with permission prompts for important actions.

• The system scored 41.6% on Humanity's Last Exam and 27.4% on FrontierMath, beating previous AI models significantly.

• Rolling out to Pro users first (400 messages/month), then Plus and Team users (40 messages/month) over the next few days.

• OpenAI implemented extensive safety measures against prompt injection attacks and treats it as high-risk for biological capabilities.

Why it matters: This changes everything. ChatGPT isn’t just answering questions anymore it can now do the actual work. From browsing to building full projects, it’s the first AI agent that might really replace parts of your job especially for office and computer-based work.

NVIDIA

🧠 Nvidia releases powerful reasoning AI models that rival OpenAI's o3

Recaply: Nvidia just released OpenReasoning-Nemotron, a set of 4 advanced reasoning AI models. They’re open-source and can handle complex math, science, and coding problems without using expensive training methods like reinforcement learning.

Key notes:

• Four model sizes released (1.5B, 7B, 14B, 32B parameters) based on Qwen2.5 architecture.

• The models were trained using data from DeepSeek’s R1 0528 671B model through simple supervised fine-tuning.

• The largest model (32B) beats OpenAI’s o3 (high) on math tests when using “heavy” GenSelect mode.

• All models are free to download on Hugging Face, and the training dataset will be released soon.

• Even though they were trained mostly on math, the models perform well on coding tasks too.

Why it matters: This release gives more people access to high-quality reasoning AI. Nvidia shows that you don’t need costly training to build strong models. By making these tools open-source, it becomes easier for developers and researchers to create powerful AI systems.

LATENT LABS

⚗️ AI model achieves picomolar protein binding with 91-100% success rates

Recaply: Latent Labs just unveiled Latent-X, a super-precise AI that designs protein-based drugs at the atomic level. It creates molecules that stick to their targets with incredible accuracy, solving complex 3D puzzles inside the body.

Key notes:

• Perfect 100% success rate across all tested therapeutic targets.

• Achieved picomolar binding strength, surpassing AlphaProteo and RFdiffusion benchmarks.

• Reduces candidate testing from millions to just 30-100 designs per target.

• Available through accessible web platform requiring no coding skills.

• Supports both macrocycle and mini-binder therapeutic modalities.

Why it matters: This could totally change how we discover new medicines. Old methods take months and cost thousands per test. Latent-X makes it fast, cheap, and accurate by using AI to design the best molecules instantly. Plus, it’s free to use, so even small labs can try cutting-edge protein design.

AI BREAKTHROUGHS

OPENAI

🏆 OpenAI's experimental AI achieves gold medal at world's toughest math competition

Recaply: OpenAI's latest experimental reasoning model just made history by earning gold medal-level performance at the 2025 International Mathematical Olympiad, solving world-class math problems that challenge the brightest human minds.

Key notes:

• The AI solved 5 of 6 problems, scoring 35/42 points - enough for gold medal status.

• Only 9% of human participants typically earn gold at this prestigious competition.

• Model used general-purpose reasoning, not specialized math programming.

• Solutions were graded by former IMO medalists following official competition rules.

• This breakthrough required sustained creative thinking over 100+ minute problem sessions.

Why it matters: This shows how far AI has come. Instead of solving basic math, it can now tackle full-on proofs and complex logic just like top human students. The model isn’t public yet, but it shows we’re getting closer to AI that can truly understand and reason through hard math problems.

FAST TRACKS

📰 What Matters in AI Right Now?

Alibaba Qwen released Qwen3-235B-A22B-2507, splitting it into two parts: one for following instructions and one for deep thinking. This makes it work better than their old all-in-one version.

Musk hints xAI is working on Baby Grok, a friendly version of their AI chatbot made just for kids. It will focus on safe and fun content for younger users.

CapCut alternative OpenCut launched Alpha, a free video editing app. It lets you cut, trim, and edit with a timeline, no watermarks, and no cost.

DuckDuckGo added a new setting so you can hide AI-generated images in search results. It uses open-source lists to block them.

Composite Sciences released Composite Agent, an AI assistant that connects to your browser and can do tasks for you using your own accounts.

OpenAI’s ChatGPT now handles 2.5 billion prompts every day. Out of those, 330 million come from users in the US.

🧰 AI Tools of the Day

🧠 Kiro: AI IDE that turns prompts into specs, tasks, code, and docs.

🤖 Kimi K2: Open‑source trillion‑parameter MoE model for coding and autonomy.

🎯 Mintly: Clone proven ads into high‑converting static/video creatives instantly.

🗣️ Talkio: AI voice tutors for practice and pronunciation in 40+ languages.

🐝 beehiiv: All‑in‑one newsletter platform built for growth, monetization, SEO.

AI PROMPTS

📊 Create Business Model Canvas

#CONTEXT:
You will act as a business model canvas architect. Your role is to help the user build a detailed business model canvas. This canvas should clearly outline the main parts of their business. Ensure these parts work well together, support the business goals, and match current market conditions.

Important Note: Click the button below to see the full prompt in Notion.

Have a favorite prompt? Tell us about it or rate today’s prompt here

That’s a wrap for today!

We hope today’s insights saved you time and brought more clarity to your day.

👉️ Got a second, Your feedback help us make Recaply even better.

Have an idea for Recaply AI?

We’re all ears. Send your thoughts to [email protected] — we read every message and love turning your ideas into reality.

See you in the next recap,

— Jason @ Recaply AI