Welcome back!

Anthropic is pushing the limits on AI coding with its latest release. Claude 3.7 Sonnet combines quick answers and deep thinking in a single model, setting a new industry standard in coding performance benchmarks. This could be a game-changer for complex software engineering. Let's unpack...

Today’s Summary:

• 🚀 Anthropic unveils Claude 3.7 Sonnet

• 🤖 Helix robots learn via voice

• 🎥 Google's Veo 2 pricing

• 🧠 Ex-OpenAI CTO launches new AI lab

• 🧬 DNA-writing Evo 2 AI model

• 🎮 Microsoft's AI preserves classic games

• 🛠️ 2 new tools

TOP STORY

Claude 3.7 Sonnet is the first hybrid reasoning AI model

The Summary: Anthropic has released Claude 3.7 Sonnet, the first hybrid reasoning model that lets users control how long the AI thinks before answering. The model can provide both quick responses and extended thinking within the same system. Early testers have been highly impressed by its performance in coding. Alongside this release, Anthropic introduced Claude Code, a command-line tool that enables developers to delegate programming tasks directly from their terminal.

Key details:

• Achieved 62.3% accuracy on real-world coding tasks (SWE-Bench), outperforming o3-mini at 49.3% and Claude 3.5 at 49%

• 128K output tokens and 200K context window for complex projects

• Reduces refusals by 45% compared to previous versions

• API users can set a "thinking budget" trading speed for quality

• Early tests show Claude Code completing tasks in one pass that would typically require 45+ minutes of manual work

Why it matters: Anthropic has doubled down on AI coding with Claude 3.7 Sonnet. Its unified approach challenges the norm of separate models for different thinking modes. The integration of quick responses and deep reasoning into a single model gives it an edge, which could lead OpenAI to accelerate its plans for a similar architecture in GPT-5.

Extended thinking tips

ROBOTICS

Figure AI’s Helix robots learn by listening

The Summary: Figure AI has created Helix, a groundbreaking AI model that lets humanoid robots grab any household object through simple voice commands. This dual-system approach combines a vision-language model with a motor control system. In demonstrations, two robots worked together to put away groceries they had never seen before.

Key details:

• Helix controls 35 degrees of movement at 200Hz, including individual finger movements, head tracking, and torso positioning

• Trained on just 500 hours of data, 20 times less than usual

• Robots can understand abstract commands like "pick up the desert item" and identify a cactus

• Robots can collaborate without specific training

• Figure ended its partnership with OpenAI to develop its own AI

Why it matters: Home environments present robotics' greatest challenge with countless unique objects and unpredictable layouts. Helix's ability to understand natural language and manipulate objects never seen before changes how robots can learn. The traditional method of training on thousands of demonstrations per new task is replaced by instant adaptation through conversation.

GOOGLE

Veo 2 brings 4K AI videos at 50 cents per second

The Summary: Google has revealed pricing for its Veo 2 AI video generation model at 50 cents per second of footage. The system can create videos up to two minutes long in 4K resolution, positioning it as a competitor to stock footage and a tool for content creators. While this amounts to $30 per minute, Google DeepMind researchers target professional users and contrast it with traditional filmmaking costs at up to $32,000 per second.

Key details:

• Veo 2 offers pay-per-use pricing versus OpenAI's Sora subscription model at $200/month for ChatGPT Pro

• Google is testing Veo 2 to generate backgrounds for YouTube Shorts

• High-quality stock footage typically costs 50-100 times more

• Cheaper models like Kling AI 1.6 go as low as $0.07 per second, but Veo 2 achieves higher quality

• Benchmarks show Veo 2 outperforms other leading video models in human evaluations of quality and prompt accuracy

Why it matters: This pricing model is still costly for personal users but can bring professional-grade 4K video creation to small studios. As AI video generation becomes more affordable, we'll likely see a transformation in content creation workflows and new forms of visual media emerging from creators who were previously priced out of high-quality video production.

Google Cloud pricing

QUICK NEWS

Quick news

• Ex-OpenAI CTO Mira Murati launches Thinking Machines Lab

• Evo 2 biological research AI model writes DNA on demand

• Microsoft Muse AI model for gameplay ideation

TOOLS

🥇 New tools

• Tanka - AI Messenger with long-term memory for teams

• Riley - Your smart parenting companion