Presented by
Good morning, AI enthusiasts. Remember when controlling a computer with your mind was a sci-fi punchline? A startup just put that tech in a beanie.
Sabi's wearable doesn't require surgery or a lab setup. It's a hat with tens of thousands of sensors, and it's shipping to consumers by late 2026.
In today's recap:
Sabi's brain-reading beanie targets 30 WPM thought-typing
Anthropic's Opus 4.7 raises the bar on agentic coding
Build a desktop research pipeline with Perplexity for Mac
OpenAI Codex expands into full computer-use agent
4 new AI tools, prompts, and more
SABI INC
A brain-reading beanie you can actually wear
Sabi
Recaply: Sabi just unveiled a wearable beanie that reads brain signals and lets users control devices by thought, with shipments planned for late 2026.
Key details:
The Sabi beanie uses dry electrodes to pick up EEG signals from the scalp, letting users select text and control apps through focused thought.
The device packs 70,000 to 100,000 sensors, compared to 12 to 200 in most consumer EEG headsets, with a starting typing speed target of 30 words per minute.
Khosla Ventures, a firm with multiple BCI investments, is among Sabi's backers, according to Wired's report on the device.
Sabi plans to ship to early buyers by late 2026, with pricing not yet announced.
Why it matters: Brain-computer interfaces have lived in labs and operating rooms for years. Neuralink brought them into the spotlight, but most people won't do surgery. Sabi doesn't cut anything. It's a beanie you pull on. Most consumer EEG headsets use 12 to 200 electrodes. Sabi packs in tens of thousands. That density could make the difference between a novelty and a real tool. Real-world accuracy still needs to be proven.
PRESENTED BY HCAPTCHA
In a World of AI Agents: Intent > Identity
AI-powered bots aren’t just logging in anymore. They’re mimicking real users, slipping past identity checks, and scaling attacks faster than ever.
Thousands of companies worldwide trust hCaptcha to protect their online services from automated threats while preserving user privacy.
Now is the time to take control of your security.
ANTHROPIC
Claude Opus 4.7 arrives with stronger agentic coding
Anthropic
Recaply: Anthropic just released Claude Opus 4.7, a notable upgrade in software engineering that verifies its own outputs before reporting back, now available across all Claude products, the API, and partner clouds.
Key details:
Opus 4.7 runs at "xhigh" effort level, meaning it checks and retests its own code outputs before calling a task complete, reducing silent failures in long agentic runs.
On CursorBench, Opus 4.7 scores 70% against a 58% baseline, with SWE-Bench performance tripling compared to the previous model.
Anthropic added new cybersecurity guardrails to Opus 4.7, including stricter controls on assisted exploitation and unauthorized access attempts, according to the release notes.
Opus 4.7 is available now on Claude.ai, the API, and through Amazon Bedrock, Google Cloud Vertex AI, and other partner cloud platforms.
Why it matters: Benchmark debates are a fixture in AI model launches, and Opus 4.7 is already drawing some skepticism. But the partner results are harder to dismiss. Cursor's own CursorBench shows a real jump in coding accuracy. The self-verification behavior also addresses a common agentic failure mode, where the model confidently reports a task as done when it isn't. For developers using Claude in production, that fix matters more than any lab benchmark.
GUIDES
Build a desktop research pipeline with Perplexity Personal Computer
Perplexity
Recaply: In this tutorial, you will learn how to set up Perplexity Personal Computer on Mac to automate research across your local files, email, and calendar without switching tools.
Step-by-step:
Download the Perplexity Mac app from the App Store and sign in with your Max account. Personal Computer is live now for Max subscribers and waitlist users.
Open System Settings, then Privacy, then Files and Folders. Allow Perplexity to access the folders you want it to search and read.
In the Perplexity settings panel, connect iMessage, Apple Mail, and Calendar. This lets Personal Computer read and act across your native Mac apps.
Start a task in the app: "Find everything I got on [topic] this week in Mail and Calendar, then check my Downloads folder for related files." Personal Computer runs the task across all linked sources.
Enable "Run in background" in the Mac app. To send tasks from your phone, install the latest Perplexity iOS app, enable 2FA, and kick off tasks from your iPhone while Personal Computer runs on your desktop.
Pro tip: Create a Research folder in Documents and drop PDFs, screenshots, and notes into it. Then prompt "Summarize my Research folder on [topic]" to get a briefing without opening files one by one.
OPENAI
Codex grows from code editor to computer agent
OAI/Screenshot
Recaply: OpenAI just expanded Codex into a full computer-use agent for Mac, adding 90+ plugins, a memory preview, and the ability to run apps and generate images.
Key details:
Codex can now open Mac apps, run background tasks, and generate images using OpenAI's computer use capability, going beyond its original code writing and editing features.
More than 90 plugins are now available, and the platform serves 3 million weekly active developers, according to OpenAI's update.
A preview of persistent memory lets Codex carry context across sessions, so users don't have to re-explain their setup each time.
The expanded Codex rolls out to all users today, with no separate sign-up needed for those already on the platform.
Why it matters: Codex started as a code tool. Now it runs apps, stores context, and plugs into 90+ services. That's a new product category. On the same day Anthropic launched Opus 4.7, OpenAI's move isn't a better score. It's a bigger surface area. Developers in Codex now have a cross-app agent built into their workflow. Whether memory and computer use hold up in real work is still to be tested, but the scope has clearly shifted.
TOOLS
Trending AI Tools
🧠 Sabi - A consumer BCI beanie with 70K-100K sensors that lets users control devices through thought
🤖 Claude Opus 4.7 - Anthropic's new flagship model with self-verifying outputs and stronger agentic coding performance
⚙️ OpenAI Codex - OpenAI's expanded coding agent with macOS computer use, 90+ plugins, and a persistent memory preview
🎥 HyperFrames - HeyGen's open-source framework for AI-powered video hyperframe generation
NEWS
What Matters in AI Right Now?
OpenAI introduced GPT-Rosalind, a reasoning model for life sciences research covering drug discovery, genomics, and protein engineering, with a free Life Sciences plugin for Codex connecting to 50+ scientific tools and data sources.
Perplexity launched Personal Computer, a desktop AI agent for Mac that searches, reads, and writes local files while working across native apps like iMessage, Apple Mail, and Calendar, available to Max subscribers starting today.
Factory AI raised $150M in a Series C led by Khosla Ventures, valuing the company at $1.5B. The agentic software development platform reports doubling revenue month over month for six consecutive months, with customers including Nvidia, Adobe, and Palo Alto Networks.
Cloudflare launched Email Service in public beta, allowing agents built on its platform to send and receive email natively without API keys or manual authentication setup, alongside an Email MCP server and open-source reference app for building email-native agents.
xAI is supplying tens of thousands of GPUs to Cursor to train Composer 2.5, its next AI coding model, according to Business Insider. The deal comes as Cursor's valuation reaches $29B, signaling xAI's push to become a compute provider for AI developers.
Alibaba open-sourced Qwen3.6-35B-A3B, a sparse MoE model with 35B total parameters and only 3B active, under the Apache 2.0 license, with agentic coding performance claimed to match models 10x its active size.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
