Welcome back. Compute is the only real moat in AI now. For months, Cursor had to rent infrastructure from rivals like OpenAI and Anthropic. SpaceX just flipped the equation with a two-track deal: a $60 billion buyout path or a $10 billion partnership that locks Cursor onto Colossus, the world's largest training cluster.

Today: OpenAI’s new image model, the 56 laws every engineer should know, and why Altman called a rival lab's safety pitch "fear-based marketing."

OpenAI's new image model brings a major leap in reasoning: The ChatGPT maker just dropped ChatGPT Images 2.0, describing it as a major leap forward in instruction following, object placement, and rendering dense text across any aspect ratio. It’s also the first image model with "thinking" capabilities. Developers can test the new models and find more technical details here.

Claude Code can now recap your terminal sessions: The AI coding tool can now generate a one-line recap of your last session whenever you step away for a few minutes. It only triggers after three turns to keep short exchanges clean, and it never fires twice in a row. It is perfect for developers juggling multiple terminals on different tasks. You can also run /recap to get one on demand.

Unauthorized users tap Anthropic's restricted AI model: Bloomberg just reported that a private Discord group accessed Anthropic’s unreleased Claude Mythos Preview, a model capable of exploiting flaws in every major OS and browser. The group allegedly found the endpoint through a third-party contractor, using clues from the Mercor breach. Anthropic is investigating, but says there is no evidence the leak extended beyond that single vendor.

Everyone has an opinion on where agentic analytics is headed, but few can tell you what it’s actually like to run in production.

Cube’s Agentic Analytics Summit 2026 is the flagship conference forecasting this future, focused specifically on the intersection of semantic layers and agentic AI.

Register for the no-cost summit to learn from data leaders at companies like Brex, Jobber, and more:

Save your spot.

The user has just changed. AI agents are completely shifting how we think about developer tools. Just this past month, Cursor engineer Eric Zakariasson shared a post on designing CLIs tailored specifically for these agents. The post struck a chord — almost every major dev tool company is quietly rebuilding its CLI for this exact reason.

Stuck at the keyboard. When AI agents use CLIs built for humans, things break fast. They stall on interactive prompts, guess at missing flags, and get stuck on help pages without examples. Even worse, without safeguards, an agent might retry a failed deployment and accidentally create a duplicate.

Your internal tools are invisible. Even if an agent knows every public CLI, it’ll still trip up on your internal deploy scripts. Your custom tools are complete strangers to it. That documentation you've been putting off is now the only way for agents to actually navigate your tech stack.

Agents become the user. Designing CLI patterns is the easy part. The real challenge is the documentation layer. That’s what determines if agents are actually useful within your stack or just a flashy demo. If your team is currently building one, former Microsoft PM Trevin Chow outlined seven principles worth following.

Tweet Hunter helps 5,600+ founders stay consistent on X.

In one session, you get:

All built on the official X API, so your account stays safe.

Start your free trial for 7 days

If you're tired of conversations going off the rails, Codex CLI has a hidden fix that lets you fork the chat from any previous point. Just double-tap “Esc” with an empty composer to jump into edit mode on your last message.

You can keep hitting “Esc” to go further back through the transcript, fix the prompt where things went south, and hit “Enter” to restart the thread from there.

Anything after your edit is swapped out for a brand-new response starting from that point.

How Intercom's principal engineer ships 2X faster: In this tutorial, you’ll find practical ways to integrate AI into mature codebases. Discover how to build custom AI skills and track productivity with telemetry. The guide also covers automating tech-debt fixes and shifting toward a faster, agent-first engineering workflow.

Routa (740 ⭐): A workspace-first platform for multi-agent coordination in software delivery. It surfaces goals, tasks, sessions, and review status on a visual board so everything stays clear and accessible, rather than getting lost in a long chat thread.

LeWorldModel (by Yann LeCun, Ex-Chief AI Scientist at Meta): Most AI world models have a hard time learning from raw video without using complex training shortcuts. LeWorldModel changes that with a straightforward two-rule system that enables fast, stable training and matches the performance of far more massive models.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 240K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team