In partnership with
Hey folks,
The Anthropic CEO admitted his company is under "incredible commercial pressure" and that all the safety work they do "makes it even harder." But the line that made me listen was "If my revenue is not $1 trillion... there's no force on Earth, no hedge on Earth that could stop me from going bankrupt if I buy that much compute." This is the guy who left OpenAI because he thought they weren't focused enough on safety and now publicly wrestling with the tension between mission and survival.
Today in AI:
→ Anthropic's budget model just beat its flagship model
→ Apple's building three AI wearables
→ Prompt of the Day: Get Claude to automate your browser workflows
Let's dive in..
🧠 The $3 Model That Beats the $15 One
Anthropic just released Claude Sonnet 4.6, and it's doing something that shouldn't be possible at its price point: matching Opus-level performance at one-fifth the cost.
The headline capability is computer use. Sonnet 4.6 can fill out multi-step web forms, navigate spreadsheets, and coordinate across browser tabs, performing at near-parity with Opus 4.6 on real desktop tasks. That's the $15/million-token model's job being done by the $3 one. 16 months ago, no Sonnet could do any of this reliably.
In internal testing, users preferred Sonnet 4.6 over Opus 4.5 .They said it was "less prone to overengineering," showed "fewer hallucinations," and "more effectively read context before modifying code." People working with the model daily chose the cheaper one. So why would anyone pay Opus prices? That's the question Anthropic just made very hard to answer.
What you actually get:
1M token context window, a first for Sonnet-class models. That's roughly 750,000 words of memory
$3/$15 per million tokens, one-fifth the cost of Opus 4.6
Now the default model for free and Pro users on claude.ai
Computer use at near-human level: filling multi-step web forms, navigating spreadsheets, coordinating across browser tabs
For your team: If you've been paying Opus prices for coding or automation tasks, test Sonnet 4.6 against your actual workloads. You might be overpaying by 5x for comparable results.
Together with 1440 Media
Smart starts here.
You don't have to read everything — just the right thing. 1440's daily newsletter distills the day's biggest stories from 100+ sources into one quick, 5-minute read. It's the fastest way to stay sharp, sound informed, and actually understand what's happening in the world. Join 4.5 million readers who start their day the smart way.
📱 Apple's Building Three AI Wearables
Examples are AI generated
The first is smart glasses. They've got a dual-camera system: one for photos and video, one that feeds visual context to Siri so it can answer questions about what you're looking at. Prototypes are already with hardware engineering teams.
The second is an AI pendant. AirTag-sized, clips to your shirt or hangs from a necklace. It's essentially a camera and mic that act as "the eyes and ears" of your iPhone. Some Apple employees are calling it exactly that.
The third is camera-equipped AirPods, and these are the furthest along. Apple's been exploring camera earbuds since 2024, and they're planned for as early as this year.
All three devices share the same core idea: give the upgraded Siri visual input so it can understand the world around you. But remember Humane? They tried to build a standalone AI wearable and it flopped. Apple's doing the opposite: three different form factors, all tethered to the iPhone. Is that the right call, or is it just Apple being Apple? Either way, they're not building one AI device. They're building an AI sensor network you wear.
For your team: If you're building products that rely on voice-only AI assistants, the shift to multimodal (voice + vision) is coming from Apple's direction. That changes the UX bar for everyone.
🎯 Prompt Of The Day - Automate Repetitive Tasks
Get Claude to automate repetitive browser workflows by describing the task the way you'd explain it to a new hire.
I need you to [describe your repetitive browser task] e.g., check three competitor websites for pricing changes, fill out a weekly status report form, gather data from multiple dashboard tabs.
For each step:
(1) take a screenshot first to confirm what you see,
(2) describe what you're about to do before doing it,
(3) verify the result after each action before moving on.
If anything looks unexpected (a popup, a changed layout, a login wall), stop and tell me instead of guessing. When you're done, give me a summary of what you completed and anything that needs my attention.The key with computer use is being explicit about verification. Without the "take a screenshot and check" instruction, Claude sometimes assumes its clicks landed correctly when they didn't. Adding that one line dramatically improves reliability.
🐝 AI Buzz Bits
🤖 Alibaba drops Qwen 3.5: 397 billion parameters, only 17 billion active per pass. It's built for agents. The model can independently execute tasks across apps, not just generate text. Supports 201 languages (up from 82), runs 60% cheaper than its predecessor, and it's open-weight.
🎮 Google I/O 2026: May 19-20 at Shoreline Amphitheatre. Google announced the dates with a Gemini-powered mini-golf puzzle because of course they did. Expect a wall of Gemini updates and Android announcements. Mark your calendar.
💬 Manus just put its full AI agent inside Telegram. Not a chatbot, the actual Manus with multi-step task execution, voice message support, and file handling. Scan a QR code, connect in a minute, and run complex workflows from your phone.
🛠 Tool Spotlight
Tonkotsu — Desktop app that turns you into a tech lead managing parallel Claude Code agents. Plan work, delegate dozens of coding tasks at once, review the diffs.
Vibe Pocket — Run Claude Code, Codex, and other CLI agents from your phone. Persistent cloud terminals with 10GB storage, close the tab and your session stays alive. Ship small fixes from your commute.
Goodfire — AI model debugger built by researchers. Their Ember API lets you reach inside a model's internals to steer behavior, prevent jailbreaks, and fix hallucinations.
Verdent Deck — Multi-agent coding platform that launches parallel AI agents in isolated Git-enabled codespaces. Each agent gets its own sandbox so they can't collide.
For a full list of 1500+ AI tools, visit our Supertool Directory
👉 Know someone drowning in AI news? Forward this to them or send your unique referral link
Cheers, Tim
