Good morning. It’s Wednesday, April 22nd.

You read. We listen. Let us know what you think by replying to this email.

ChatGPT Images 2.0, powered by GPT-Image-2, is a model that “thinks” before it draws and currently sits at the top of the leaderboards. By using web-informed reasoning to plan layouts, it can generate up to eight consistent, 2K-resolution images per prompt. It is significantly better at rendering complex typography and infographics, though real-world testing shows its non-English text can still be hit-or-miss.

On the dev side, the Codex macOS app is getting a "Clippy" vibe with interactive pixel-art avatars. It is also gaining a "Chronicle" feature, a screen-aware memory that processes your activity into local Markdown summaries. To help parse all this data, OpenAI open-sourced Euphony, a dedicated visualizer for chat and Codex session logs.

The most ambitious move, however, is the Agent Studio (codename ‘Hermes’). This is OpenAI’s play for persistent, 24/7 autonomous agents that live in Slack and handle scheduled workflows. In a new interview, Sam Altman took a shot at Anthropic’s high-security Mythos model, calling their restricted-release strategy "fear-based marketing." Read more.

Watch: OpenAI’s founders sit down for their first joint podcast to talk about AGI and "Spud"

Bloomberg reports that Anthropic’s unreleased Mythos model, the cornerstone of its defensive Project Glasswing, was briefly accessed by unauthorized users via a third-party environment. To tighten its perimeter, Anthropic has begun requiring government-issued IDs and selfies for high-stakes accounts to block adversaries.

Anthropic is also making a hard pivot toward "persistent" AI with Conway, an unreleased system designed to run always-on agents in their own containerized environments. We’re already seeing the first signs of this in Claude Cowork through Live Artifacts, stateful, auto-updating dashboards that pull real-time data from external apps and maintain their own version history so your work doesn't vanish when the chat ends. To keep these background agents running, Anthropic locked in a massive $100 billion AWS deal for 5 gigawatts of power, fueled by over a million custom Trainium2 chips.

On the research side, the STEM Fellows program is recruiting PhDs to turn Claude into an "AI scientist" capable of long-horizon reasoning and experimental design.
The consumer side is getting a squeeze. Anthropic is A/B testing a shift where 2% of new Pro users may lose access to Claude Code, likely pushing that high-compute tool toward the $100 Max tier. Read more.

SpaceX is making a massive, grounded play for the developer market by partnering with Cursor, the AI coding startup that’s quickly becoming a favorite among engineers. The deal is essentially a "try before you buy" on a grand scale: SpaceX has the option to acquire Cursor for $60 billion later this year or pay a $10 billion development fee. In exchange, Cursor gets to plug into the Colossus supercomputer, a beast equivalent to about a million H100 GPUs, to train what they’re calling the "world's best" models for coding and knowledge work.

SpaceX’s latest pre-IPO filings reveal a much more skeptical view of its "Space AI" ambitions. Despite the earlier hype about "orbital data centers," the company is now warning investors that space-based compute relies on unproven tech and might never actually be profitable. Read more.

Moonshot AI’s Kimi K2.6 just broke the time barrier for autonomous work. While most agents flake out after a few minutes, K2.6 runs for days. In one test, it built a compiler from scratch in 10 hours—work that usually eats two months of a four-person engineering team's life.

An Agent Swarm manages 300 sub-agents across 4,000 coordinated steps, shifting AI from a simple script to persistent infrastructure. The problem is that current orchestration is too brittle for this stateful execution. We’re reaching a point where AI modifies code faster than humans can review it, creating a desperate need for agent-specific gateways and kill switches. Read more.

HyperFrames from HeyGen is an open-source, agent-native framework that converts HTML into MP4s through a simple terminal command.

Waydev tracks agent-generated code from IDE to production to measure real ROI and cost per shipped PR.

RankAI is an autonomous agent that handles SEO and GEO to drive buyers from Google and ChatGPT.

Dageno AI shows how your brand appears across major LLMs and uses agents to help you improve.

Dune is a context-aware Mac keypad that reconfigures its three keys in real time based on your active app.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.