Good morning. It’s Wednesday, February 26th.

On this day in tech history: In 1991, Tim Berners-Lee, a British computer scientist, introduced the world to "WorldWideWeb," the first web browser and a WYSIWYG (What You See Is What You Get) HTML editor.

In today’s email:

• Claude Sonnet 3.7

• Perplexity Web Browser

• AI Audio Generation for Video

• OpenAI drops Deep Research access to Plus users

• DeepSeek Rushes For New AI Model

• 4 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

Claude 3.7 Sonnet Introduces ‘Thinking Mode’ to Balance Speed and Depth in AI Responses

Anthropic is raising the stakes in enterprise AI with Claude 3.7 Sonnet, a model that lets users fine-tune its depth of reasoning on demand. The ‘thinking mode’ toggle shifts between instant responses and deep analysis, challenging OpenAI’s modular approach and DeepSeek’s cost-cutting strategy. Benchmarks put Claude 3.7 at 78.2% accuracy on graduate-level reasoning tasks, outpacing DeepSeek-R1 in instruction-following. Rather than chasing raw speed, Anthropic is betting on AI that adapts to complexity without switching models.

Alongside Claude Code—a command-line AI developer assistant with built-in human oversight—the company is positioning itself as the alternative to fragmented AI stacks. If enterprises buy in, this could mark a shift away from model juggling toward AI that actually thinks before it speaks. Read more.

Perplexity teases web browser ‘Comet’

Perplexity AI is taking its search ambitions up a notch with Comet, an AI-powered web browser. While details remain scarce, Perplexity claims Comet will “reinvent” browsing, much like it did with search. This move follows the company’s streak of rapid expansions—an AI assistant for Android, a deep research tool, and an API—each reinforcing its bid to own the AI-first internet experience.

Despite its growth, Perplexity faces mounting legal scrutiny. As Perplexity scales its monetization efforts including advertising and enterprise offerings, Comet’s success may hinge on whether the company can navigate both technical innovation and legal headwinds without losing momentum. Read more.

Luma AI Adds Free Audio Generation to DreamMachine Video Tool

Luma AI has added audio generation to its DreamMachine video tool, allowing users to create soundtracks for their AI-generated videos. The new "Audio" feature enables automatic sound generation with a single click or customized audio using text prompts.

Currently in beta, the feature is available for free to all users. This update expands DreamMachine’s capabilities, integrating AI-driven visuals and audio within a single workflow. Read more.

OpenAI drops Deep Research access to Plus users, heating up AI agent wars with DeepSeek and Claude

OpenAI has expanded access to Deep Research, its AI-powered analysis engine, bringing ChatGPT Plus, Team, Education, and Enterprise users into the fold. Running on a specialized o3 variant, it autonomously pulls, structures, and synthesizes insights across text, images, and PDFs—delivering intelligence that mimics human analysis.

It’s a calculated move. DeepSeek’s open-source R1 is shaking up the AI market, undercutting OpenAI’s paywalled grip. Perplexity is running R1 for a fraction of the cost, while Anthropic’s Claude 3.7 Sonnet leans into transparency with explicit reasoning trails.

Subscription tiers, however, remain firmly in place. Plus users get 10 queries a month, while Pro subscribers—paying ten times more—unlock 120. Read more.

DeepSeek rushes to launch new AI model as China goes all in

Fresh off the market-shaking debut of R1, DeepSeek is now accelerating the launch of its R2 model, originally set for May. R2 is expected to enhance code generation and expand multilingual reasoning, further strengthening DeepSeek’s competitive position.

Once viewed with skepticism by Beijing over its massive chip acquisitions, DeepSeek now enjoys quiet state backing, with government-linked enterprises integrating its models at scale. Nvidia recently teamed up with DeepSeek to optimize R1 for Blackwell, driving a 25x revenue increase and 20x lower inference costs. Developers can now access an FP4-optimized DeepSeek checkpoint on Hugging Face.

The company has also resumed API access after a three-week pause due to capacity constraints. Developers can now top up credits, though server strain persists during peak hours. Its domestic rivals are accelerating development—Alibaba launched a preview of its own reasoning model, QwQ-Max, on the same day DeepSeek reopened API access. Read more.

• Gemini Code Assist free tier undercuts rivals with massive usage limits

• Microsoft’s Copilot Voice and Think Deeper now free with unlimited access

• Alibaba makes AI model for video, image generation publicly available

• Google AI Studio Launches Conversation Branching for Gemini Models

• Grok 3 is pulling Grok usage to record highs

• Huginn-3.5B—an AI that ditches words and thinks in pure computation

• Claude 3.7’s AI-generated Snake game gains self-awareness and evolves

• Unitree teaches its robot kung fu, warns you not to try the same

• Anthropic's Claude AI is playing Pokémon on Twitch — slowly

• DeepSeek drops DeepEP, an open-source library to streamline MoE model training 

• Grok’s uncensored AI voice mode lets users talk sex, therapy, and conspiracies

• Quora's Poe now lets users create and share custom AI-powered apps

• UK delays plans to regulate AI as ministers seek to align with Trump administration

• Robot watches itself like humans do in a mirror to master movements, self-repair

• Salesforce strikes multibillion-dollar deal to bring its AI agents to Google Cloud

• Nvidia's H20 chip orders jump as Chinese firms adopt DeepSeek's AI models

• Anthropic used Pokémon to benchmark its newest AI model

• Singapore's biggest bank DBS to cut 4,000 roles as it embraces AI

• 1,000 artists release 'silent' album to protest UK copyright sell-out to AI

📄 DICEPTION: A Generalist Diffusion Model for Visual Perceptual Tasks

📄 Thus Spake Long-Context Large Language Model ​​

📄 OmniAlign-V: Towards Enhanced Alignment of MLLMs with Human Preference 

📄 SpargeAttn: Accurate Sparse Attention Accelerating Any Model Inference

📄 VideoGrain: Modulating Space-Time Attention for Multi-grained Video Editing

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!