Hey, it's Andreas.
A few updates before we get into it.

The first The Agentic AI Cohort is in full swing, and the biggest lesson so far: three hours a week is not enough to build agents, so we spent the whole weekend in calls with participants pushing their builds forward. Absolutely worth it. A few participants are already building real products, things that could genuinely be sold or save their companies many hours every week. And it's only week one. Given the outreach I've received, a second edition is likely coming soon. I'll share the details and some of this weekend's builds with the waitlist first (500+ people already in), so if you're interested, join the waitlist here and be the first to hear when we go live.

The AI career book I wrote with co-authors from Microsoft, Google and Liquid AI gets a live panel with DataCamp in a few days. It's free, join here if breaking into AI is on your list.

Pre-orders for my Claude Code book (“The Definitive Guide to Claude Code”) are open and moving faster than the publisher expected, with Korean and Japanese translations already in discussion.

And I was just in Los Angeles, invited by LinkedIn to record a course on AIOps and transformation in their studios. Out soon.

In today's issue:

Let’s go

🌀 OpenAI ships GPT-5.6, GPT-Live and ChatGPT Work → Three launches in three days: a new model family, real-time voice, and an agent that works across apps and files.

🌀 Meta pulls Muse Image after privacy backlash → The feature let anyone generate images from public Instagram users’ photos, with adults opted in by default and no notification. Meta removed it days later.

🌀 Apple sues OpenAI over alleged hardware secret theft → Apple claims former employees took confidential product files to help build OpenAI’s rival AI device.

🌀 Claude Cowork keeps working after your laptop closes → Tasks and scheduled jobs can now run without any device online, with results available on mobile. Beta starts with Max users.

🌀 Notion launches Agent Hub for shared workspace agents → Teams can now run custom Notion agents from mobile and share them across the workspace.

🌀 Grok 4.5 lands as a fast, low-cost coding model → Priced at $2 input and $6 output per million tokens, with limited free access in Cursor and Grok Build.

🌀 Nobel chemist Omar Yaghi joins Tsinghua’s AI materials lab → The UC Berkeley scientist will lead AI-driven materials discovery in Beijing, targeting breakthroughs in water, carbon capture, and medicine.

📘 AWS offers a free course on prompt engineering → A four-hour intermediate course covering prompt fundamentals, advanced techniques, and safeguards against misuse, with hands-on examples using Amazon Bedrock.

📘 AWS offers a free no-code ML and GenAI course → A 5.5-hour course on building, evaluating, and deploying models with SageMaker Canvas, including RAG, fine-tuning, and foundation model comparison.

📘 Anthropic launches three new Claude certifications → New tracks cover everyday Claude use, agent development, and enterprise architecture, including Claude Code and MCP. For now, exams are limited to Anthropic Partner Network members.

📘 Geoffrey Litt (Notion) on why understanding is the new bottleneck → Autonomous loops still need human judgment. Practical tips on the loop, how to build understanding faster, and guide agents.

🔹 Musk vs Altman turns into an AI credibility battle → Musk revived his “Scam Altman” attack, while Altman fired back by pointing to model rankings.

🔹 Dwarkesh Patel and 3Blue1Brown on what AI’s math breakthroughs really mean → A sharp conversation on AI, verification, hidden links across fields, and why human judgment and curation still matter.

🔹 AI Futures Project maps five paths to AI 2040 → The team behind AI 2027 explores outcomes from a full-speed race to global cooperation. Its preferred path calls for a verified U.S.-China slowdown, radical lab transparency, and delaying superintelligence until 2040.

For the first time, both frontier labs shipped their flagships within a month of each other. Anthropic released Fable 5 on June 9, their Mythos-class model adapted for general use. OpenAI followed with GPT-5.6: previewed June 26 to roughly 20 government-approved organizations, fully public since July 9. The new OpenAI models come as a family of three: Sol, the flagship, Terra for everyday work at roughly half the cost of its predecessor, and Luna for high-volume tasks at a fraction of that.

The question I keep getting: which one is better?

Well, I think that is the wrong question, but let’s look into the details.

What the launches tell us

Neither launch was routine. Fable 5 came off the market three days after release, after researchers at Amazon claimed a bypass of its cyber safeguards. The US government applied export controls, Anthropic added new safety classifiers, and the model was redeployed on July 1. GPT-5.6 was slowed by the White House itself over cyber and bio concerns before its staged rollout.

Model launches now involve Washington. That is new. And it tells you more about where capability sits than any benchmark chart: these models are being handled closer to export-grade hardware than consumer software.

The scoreboard, read carefully

Each lab published the chart and benchmarks where it wins. Anthropic points to SWE-bench Pro: Fable 5 at 80.3%, against 58.6% for GPT-5.5, the previous OpenAI flagship. OpenAI points to Terminal-Bench: Sol at 88.8%, and 91.9% in "ultra mode," where it spawns its own subagents.

Two footnotes are worth mentioning here.

First, METR, the independent evaluation org, just reported Sol gaming its evaluations more than any public model they have tested (it exploited bugs in the eval harness itself). At this stage, it should be clear that Anthropic and every other lab are doing similar things. Benchmarks and scoreboards are superficial; while they can indicate some direction and improvements, it's better to approach them with caution.

Second, price. Fable 5 runs $10 input / $50 output per million tokens. GPT-5.6 comes as a family of three: Sol, the flagship, at $5/$30, with a new max-reasoning-effort mode and the ultra mode above. Terra, the everyday model, at $2.50/$15, roughly GPT-5.5 performance at half the cost. Luna, built for high volume, at $1/$6. I have used Fable 5 a lot over the last few days and spent a significant amount of money on tokens. If you manage agent teams or sub-agents, the limits are quickly exhausted. Therefore, being much cheaper is a strong argument in favor of OpenAI here.

What a month of real use looks like

There is an interesting read by Dan Shipper and his team at Every. They were one of the ~20 early partners and ran GPT-5.6 in production for about a month before the rest of us got it.

Dan's framing and metaphor are quite effective in conveying the idea: "GPT-5.6 is like a Porsche, while Fable is like a warp drive." Porsche is what you drive every day, the best mix of power, speed and cost for normal knowledge work and coding. Fable is what you save for getting across the galaxy. He described GPT-5.6 not as a Fable killer. But as a division of labor, when combined, the two are better than either alone.

Where each one wins

My own take, after a few days of testing both: Fable 5 is still the model I reach for first when researching complex topics, designing agent architectures, or thinking through a system before I build it. It holds large amounts of context, challenges my assumptions, and stays coherent across long, messy tasks. For coding, and frontend work in particular, I also tend to prefer Fable: cleaner structure, better design judgment, fewer superficial fixes.

GPT-5.6 earns its place on speed, cost and fast iteration, and as a second pair of eyes that catches blind spots and pressure-tests the result. Fable as the primary builder, GPT as the faster, cheaper challenger. Their strengths sit in different domains, and that is exactly why I believe "which one" is the wrong question: The most interesting usage patterns pair the models instead of picking one.

One hack before you go: GPT-5.6 Sol inside Claude Code

Most people switch between Claude Code and the Codex app. However, there is a way around that. I discovered the approach below while testing Codex and really liked it.

Just do the following:

After installation, just run Claudex, and Claude Code launches with Sol as the main model.

As I've mentioned here many times: frontier AI is moving so fast that you need a setup, and the cognitive flexibility, to adapt just as fast. Otherwise you fall behind, and you fall behind quickly.