Good morning. It’s Friday, June 26th.

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

OpenAI builds first in-house custom AI chip ‘Jalapeño’ as White House delays GPT-5.6 launch

The AI race is shifting from who builds the smartest software to who controls the actual infrastructure, and Washington wants a front-row seat. OpenAI is building a vertically integrated empire, teaming up with Broadcom and Celestica to build Jalapeño. This is a custom inference chip designed using OpenAI’s own models to optimize memory bandwidth and slash data bottlenecks in massive data centers.

On the software side, OpenAI is rolling out an upgrade to GPT-5.5 Instant focused on long-context usability, sending engineers into businesses through its DeployCo subsidiary, and testing a peer-to-peer credit economy on Codex, which just passed four million weekly users.

This expansion comes right as OpenAI’s own staff goes all-in on automation. According to the company’s new Economic Research paper, employee use of Codex systems exploded between November 2025 and June 2026. Median token usage skyrocketed 56-fold in Research, 32-fold in Customer Support, and 27-fold in Engineering. This sudden pivot toward "tokenmaxxing" shows employees are heavily relying on AI for research synthesis and legal analysis, even though they have had unlimited model access for years.

But the federal government is throwing a wrench into the rollout schedule. The Trump administration just forced OpenAI to stall the release of GPT-5.6 over security concerns, shifting it to a limited enterprise preview where federal agencies vet buyers case by case. This follows even harsher restrictions that blocked foreign nationals from accessing Anthropic’s Mythos 5 and Fable 5 models.

Watch: OpenAI's Chief Research Officer Mark Chen explains the next frontier of AI work

Anthropic claims Alibaba operators used 25,000 fake accounts to siphon Claude secrets

Anthropic is sounding the alarm on what looks like a massive AI heist. The company claims Alibaba-linked operators used nearly 25,000 fake accounts to hammer its Claude chatbot with over 28 million prompts between April and June 2026. This wasn't just casual scraping; it was a heavy-duty model distillation campaign.

By analyzing Claude's responses, the operators allegedly tried to reverse-engineer advanced capabilities like coding, agentic reasoning, and long-horizon task execution. Anthropic told lawmakers this shortcut lets rivals bypass millions of dollars in research and safety guardrails, threatening US tech leadership.

At the same time, the battle for AI dominance is moving well beyond raw compute. Anthropic is rapidly expanding its global footprint, hiring Steve Jarrett, the former Chief AI Officer at French telecom giant Orange, to lead its push into Europe and Africa from a new Paris office.

While Anthropic plans to triple its international headcount and open a office in Milan, Silicon Valley is also investing heavily in real-world infrastructure. Anthropic, OpenAI, and Stripe recently poured money into Intercept, a $500 million bio-defense initiative using AI-driven biology and smart air tech to protect human workers from illness-driven downtime.

Meta poaches Virtue AI founding team to fortify autonomous agents

In a major talent move, Meta Superintelligence Labs recruited Virtue AI co-founders Bo Li, Dawn Song, and Sanmi Koyejo, along with key engineers specializing in AI security. Rather than acquiring the startup outright, Meta absorbed the operational core of the company to accelerate deployment of runtime guardrails, automated red teaming systems, and AI governance frameworks designed to prevent model abuse, hacking vulnerabilities, and sensitive data leaks inside large-scale agentic systems.

The hires place AI safety directly inside Meta’s highest-priority research units. Song and Li will report to Nat Friedman within Superintelligence Labs, while Koyejo joins FAIR under Rob Fergus. Meta described trustworthy autonomous systems as essential for deploying AI products to billions of users.

At the same time, Facebook is testing a standalone AI companion app for creators that transforms Creator Studio into a conversational analytics and engagement platform.

Frontier models and product moves

Agents and the agentic stack

Business, labor, and institutions

Hardware and infrastructure

Research

Science and medicine

Robotics

Developer tools

External Agents in Notion connects Claude and Cursor directly to your workspace to manage team tasks automatically.

Oxlo.ai provides a single API to access over 35 frontier AI models under a predictable monthly subscription.

Zaro aggregates scattered workplace data to let you build custom apps and agents from a single prompt.

Tencent EdgeOne Makers provides an edge platform to deploy modern web applications and AI agents globally in minutes.

CanIRun.ai evaluates your hardware to instantly show you which local AI models your machine can run.

Thank you for reading today’s edition.

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.