Welcome back!

OpenAI has unveiled Operator, an AI agent that autonomously navigates a web browser to handle tasks like booking tickets and grocery shopping. Still in its preview stage, this type of agent could reshape how we interact with the web. Let’s unpack…

Today’s Summary:

• 🤖 OpenAI unveils “Operator” agent

• 🎯 Gemini “Flash Thinking 01-21” tops benchmarks

• 🧩 DeepSeek R1 leads open-source reasoning

• 🖼️ Runway launches Frames cinematic image AI

• ✍️ Anthropic Claude introduces Citations feature

• ✨ Hunyuan3D 2.0 generates 3D assets

• 🛠️ 2 new tools

TOP STORY

Meet Operator, OpenAI’s autonomous agent

The Summary: OpenAI has launched Operator, an AI agent designed to perform tasks autonomously through a browser. Powered by the new Computer-Using Agent (CUA) model based on GPT-4o, Operator navigates web interfaces like a human, handling actions such as clicking, filling forms, booking, shopping, and researching. While still in preview, Operator introduces AI capabilities that merge automation with human oversight.

Key details:

• Leverages GPT-4o vision and reasoning to control a browser, handling complex tasks like booking tickets and ordering groceries

• Anthropic Claude Computer Use introduced similar functionality 3 months earlier, but stayed low-profile due to API-only availability

• Early tests show a 38.1% success rate on OSWorld benchmark, still trailing human performance

• Partnerships with DoorDash, Instacart, and Priceline integrate Operator into real-world workflows

• Currently exclusive to US-based Pro users, with plans to expand to international users as well as to Plus, Team, and Enterprise tiers

Why it matters: Operator builds on Anthropic’s Claude “Computer Use” concepts, achieving improved benchmarks and smoother usability. It shows how web browsing could evolve, with AI handling some tasks autonomously. Despite its early limitations, Operator demonstrates the potential for browsers to function as task-driven agent workspaces.

GOOGLE DEEPMIND

Google Gemini 2.0 “Flash Thinking 01-21” tops benchmarks

The Summary: Google’s Gemini 2.0 “Flash Thinking 01-21” sets new performance standards, surpassing competitors on key benchmarks, with features like transparent reasoning, a million-token context window and built-in code execution. The model is available for free during the beta.

Key details:

• Achieved 73.3% on AIME benchmark and 74.2% on GPQA Diamond, surpassing OpenAI o1 on reasoning-heavy tasks

• Handles up to 1 million tokens, enabling the analysis of extensive data or multiple long texts simultaneously

• Includes built-in code execution, letting users to run and test code directly

• Offers transparency by showing its step-by-step thought process

Why it matters: Gemini 2.0 “Flash Thinking 01-21” challenges premium reasoning models like OpenAI o1. It raises the bar by combining cost efficiency, detailed reasoning steps, and integrated code execution.

Try it in AI Studio

DEEPSEEK

DeepSeek-R1 leads open-source reasoning

The Summary: DeepSeek has released DeepSeek-R1, an open-source reasoning model on par with OpenAI o1. Trained through reinforcement learning, the model achieves standout results in math and coding while being fully open and extremely cost-effective.

Key details:

• Achieved 79.8% on AIME and 97.3% on MATH-500, matching o1

• API is priced at $0.55 per million input tokens, 90% cheaper than OpenAI o1

• Distilled models range from 1.5B to 70B parameters, with smaller versions running locally on laptops

• MIT open license lets researchers and developers fine-tune and deploy models freely

Why it matters: By achieving state-of-the-art reasoning through reinforcement learning and open source methods, DeepSeek-R1 offers a rare alternative to proprietary AI models like OpenAI o1. It provides researchers and developers with a new blueprint for accessible, high performance reasoning AI.

Technical paper

QUICK NEWS

Quick news

• Runway releases Frames cinematic image generation model

• Anthropic launches Citations, a new API feature for Claude models

• Hunyuan3D 2.0 generates high resolution textured 3D assets

TOOLS

🥇 New tools

• Needle - Enable AI search across all your data

• Shimmer - ADHD coaching, AI-enhanced