Welcome, humans.

Anthropic launched Sonnet 4.6 yesterday as a more or less drop-in replacement for Sonnet 4.5 for any Claude users out there. Swxy @ Latent Space put together a great technical recap you can read here, and we tested it live yesterday, so instead of focusing the newsletter on it, we’re going to focus on something more interesting below. Here’s our full recap deep dive, in case you’re curious.

Grant’s TL;DR thoughts: need to test this in Claude Code / agentic environments, because as a chat model, it’s not that impressive compared to Opus. I much prefer Opus 4.6 for everything; it’s just so reliable (I’m on the Max plan atm). But if you’re using Opus via the API or on the Plus plan, you’ll burn $$ (or get rate limited) super fast. So Sonnet 4.6 is a much needed upgrade for those scenarios.

The coolest part of the Sonnet release is that Sonnet 4.6 is now available for free Claude accounts. We compared it to GPT 5.2, which is the model that free ChatGPT users can use, and they’re pretty comparable for simple tasks.

Every free user looking to switch from ChatGPT to Claude gonna be like these Unitree robots…

Here’s what happened in AI today:

• Anthropic launched Claude Sonnet 4.6 with major upgrades to coding, agents, and a 1M token context window.

• Researchers found every major LLM gets dramatically worse the longer you talk to it.

• Apple is fast-tracking three AI wearables: a pendant, smart glasses, and AI-powered AirPods.

• The European Parliament banned AI chatbots from lawmakers' devices over security risks.

🤖 LATER TODAY: What does intelligence look like when it has to walk, grasp, and not fall over?

Today at 8:00 AM PT / 10:00 AM CT / 11:00 AM ET, we're going live with Nikita Rudin, Cofounder & CEO of Flexion Robotics, to break down the AI behind humanoid robots—and hopefully demo a robot live on stream. 

You want to ask questions to a guy building the brain that will power this AI? Then you definitely don’t want to miss this one.

Watch live: YouTube | LinkedIn | X

The Instagram co-founder thinks your next app won't be an app. It'll be an agent.

Mike Krieger, Instagram co-founder and now head of Anthropic Labs, just sat down with Hugo Barra and David Singleton (the team behind Dreamer) for a deep dive on what agent-native software actually looks like when real people use it. (Full interview)

Dreamer is a platform where anyone can discover, build, and remix AI agents; no coding required. Think of it as an app store where every app can talk to every other app, and you can build your own just by describing what you want.

The co-founders aren't first-timers. Barra and Singleton met 18 years ago at Google building mobile apps before Android existed. Singleton was CTO at Stripe. Barra led hardware at Xiaomi and VR at Meta. They see agents the same way they saw mobile in 2010: a page turn you only recognize in hindsight.

Here's what makes Dreamer interesting:

• Everyone builds, not just developers. The team originally expected a traditional marketplace (devs build, consumers use). Instead, almost every user ends up building their own agents. A "Sidekick" system agent guides you through it conversationally.

• Agents recruit other agents. Barra showed a workflow where a news agent sent an article to a read-it-later agent, which generated an audio summary, which the Sidekick then posted to Slack; all without anyone designing that chain in advance.

• Software that rewrites itself at runtime. Using Anthropic's Agent SDK, Dreamer agents can spawn sub-agents that are written on the fly. Barra called this "quite profound," comparing it to how Unix pipes originally made software composable, before desktop and web apps siloed everything again.

Krieger framed it well: there's so much intelligence in the models that it "deserves not to be trapped in a chat box."

If agents are going to become how we actually get things done (not just answer questions), they need surfaces, composability, and personalization. Dreamer is betting that's a platform problem, not a model problem. The open beta launched today.

FROM OUR PARTNERS

Migrate to Delve and get a $2,000 VISA card in your inbox

Delve is the AI-native compliance platform that actually does the work for you, auto-collecting evidence from AWS, GitHub, and your stack so you don’t have to chase screenshots or babysit integrations. Use AI security questionnaire tooling, AI copilot, and everywhere else to make compliance feel less, dreadful. Welcome to the new age.

The proof is in the pudding:

• Bland → Switched, got compliant, and unlocked $500k ARR in 7 days.

• 11x → Streamlined audits and moved faster on enterprise deals.

• micro1 → Scaled compliance without adding headcount.

Bonus: Delve will handle your migration for free. Zero-touch. No disruption. No starting over. If you’re dreading opening your current SOC 2 tool, that’s your sign.

Book a demo here and trigger a migration - get $2000 sent straight to your inbox as soon as you’re onboarded. 

Prompt Tip of the Day

Anthropic just dropped updated prompting best practices for its new Claude 4.6 models, and the biggest takeaway is counterintuitive: do less.

If you've been stuffing your prompts with phrases like "be thorough," "think carefully," or "you MUST use this tool," it's time to delete them. Those bandaid fixes for older models now actively hurt performance on 4.6. As one Anthropic engineer put it: "I saw huge leaps in intelligence as soon as I removed anti-laziness prompts."

Here's the new playbook:

• Remove "try harder" language. Phrases like "be thorough" or "do not be lazy" cause the model to overthink and loop. 4.6 is already proactive; telling it to try harder is like yelling "RUN FASTER" at someone already sprinting.

• Soften tool instructions. Swap "You MUST use this tool" for "Use this tool when it would help." The model triggers tools appropriately now without being threatened into it.

• Be explicit about actions. "Can you suggest changes?" will get you suggestions. "Make these changes" will get you changes. Say what you mean.

• Use effort settings as your main dial. Instead of prompt-engineering your way to better results, adjust the effort parameter (low, medium, high) to match your task.

Want more tips like this? Check out our Prompt Tip of the Day Digest for February.

Treats to Try

1. Tasklet just dropped 32 new integrations — native Outlook, Teams, SharePoint, plus Google Chat, BigQuery, Canva, Stripe, and more; just describe a workflow (“reconcile Stripe payments with Ramp expenses every Monday”) and the AI agent handles it autonomously (demo, blog).

2. Clam is a YC-backed security layer for OpenClaw that scans every message your AI agent sends or receives for personal info leaks, prompt injections, and malicious code; your API keys never touch the AI.

3. OpenClaw shipped a major update with Sonnet 4.6 support, a 1M context window beta, native Slack streaming, an iOS share extension, and chat-spawned subagents.

4. Happycapy turns your browser into an agent-powered computer; no setup, no terminal, just describe what you want built and it handles coding, design, and file work in a cloud sandbox (#1 on Product Hunt this week).

5. Cursor, the AI coding tool, now has a plugin marketplace.

6. WordPress added an AI Assistant that edits layouts, styles, content, translations, and generates images via Google Gemini in the block editor. 

Around the Horn

1. Anthropic expects to pay Amazon, Google, and Microsoft at least $80B through 2029 to run Claude on their cloud servers; on top of that, cloud providers get a cut of Anthropic's AI sales revenue, projected to jump from $1.3M in 2024 to $6.4B by 2027.

2. NVIDIA and Meta expanded their partnership, with Meta deploying millions of Blackwell and Rubin GPUs plus its first large-scale Grace CPU-only rollout.

3. Microsoft Research and Salesforce analyzed 200K+ AI conversations and found every major model (GPT-4, Claude, Gemini, Llama) gets 39% worse the longer you talk to it.

4. The European Parliament banned AI chatbots like Claude and Copilot from lawmakers' devices, citing cybersecurity risks from cloud data uploads to US firms.

5. Figma partnered with Anthropic to launch Code to Canvas, letting you convert Claude Code-generated interfaces into editable Figma designs.

6. Apple is reportedly fast-tracking an AI wearable pendant, smart glasses (code-named N50, targeting 2027), and AI-powered AirPods to compete with Meta and Snap.

NEW: Want more? Check out our new Around the Horn Digest for February here 

FROM OUR PARTNERS

Stop typing prompt essays. Dictate full context and Wispr Flow turns it into a clean, structured prompt you can paste into your AI assistant in seconds.

Start Flowing for free today

Midweek Wisdom

1. Dan Hockenmaier says there’s now four types of people at work: slop cannons, turbo brains, steady hands, and dead weight; Dan’s full blog breaks down what that means and why good taste is the key.

2. Speaking of taste, Y Combinator co-founder Paul Graham’s new essay argued good taste in design is objective and learnable, following principles like simplicity, symmetry, and redesign, which is applicable to programming and using AI to create enduring, practical work.

3. John Coogan of TBPN diving deep into “why is no one talking about the Cournot Equillibrium”, the 187-year-old concept that came up in Dario’s interview with Dwarkesh and explains why AI labs will eventually profit: when only 3-4 players can afford to compete and products are differentiated, margins don't race to zero. Econ 101, meet the GPU age.

   1. P.S: If you haven't read our Weekend Podcast Deep Dive on Dario and a bunch of other AI podcasts, definitely check it out! 

4. Morning Brew co-founder Alex Lieberman came away from the Dario/Dwarkesh interview with a framework he's calling "The Repositioning Gap": the distance between knowing AI is transforming your industry and actually changing how your company operates.

5. Is the SaaSpocalypse’s build vs buy paradigm real? A Retool survey of 817 builders found that 35% of teams have already replaced at least one SaaS tool with a custom-built alternative, and 78% plan to build more in 2026.

6. Wharton professor Ethan Mollick just published his updated guide to which AI to use, and his big point is that "which AI should I use?" is now the wrong question. The right question has three parts: which model, which app, and which harness.

   1. Models are the AI brains (GPT-5.2, Claude Opus 4.6, Gemini 3 Pro).

   2. Apps are what you talk to them through (ChatGPT, Claude, Gemini).

   3. But harnesses are the new variable most people don't think about; they're what let the AI actually do things, like browse the web, write code, or build you a spreadsheet.

   4. The same model behaves completely differently depending on what harness it's running in.

   5. His single biggest piece of advice: always manually select the smartest model available (and our tip: add thinking if not obvious!). The defaults on every platform are optimized for speed, not accuracy. When someone posts an AI doing something dumb, they're almost always using the default. Funny; we said the same thing in yesterday’s Sonnet 4.6 live!

P.P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.