Good morning. It’s Friday, July 31st.

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

Write docs 4x faster. Without hating every second.

Nobody became a developer to write documentation. But the docs still need to get written — PRDs, README updates, architecture decisions, onboarding guides.

Wispr Flow lets you talk through it instead. Speak naturally about what the code does, how it works, and why you built it that way. Flow formats everything into clean, professional text you can paste into Notion, Confluence, or GitHub.

Used by engineering teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works system-wide on Mac, Windows, and iPhone.

Try Wispr Flow free

OpenAI's self-optimizing AI slashes token costs by 13x

OpenAI just blew up its own pricing structure, relying on internal infrastructure efficiency to drive down the cost of intelligence. By deploying speculative decoding to increase token generation by 15% and cut GPU serving costs by 20%, the company slashed API prices across its GPT-5.6 family. GPT-5.6 Luna took an 80% price drop to $0.20 per million input and $1.20 per million output tokens.

AI researcher Nic Dunz pointed out, Luna now matches the benchmark performance of March's full GPT-5.4 flagship model at one-thirteenth the token price. Meanwhile, GPT-5.6 Terra dropped 20% to $14, Sol Fast added a speed tier doubling pricing for 2.5x throughput, and Auto-review in ChatGPT and Codex CLI moved to Luna for a 10x cost reduction, countering pressure from Chinese providers, Google, Anthropic, and Microsoft.

A Chrome extension adds cross-tab context for YouTube videos and text, alongside desktop URL suggestions and history management. On the ARC-AGI-3 benchmark, using retained reasoning and context compaction via the Responses API jumped GPT-5.6 Sol scores from 13.3% to 38.3% while dropping token output sixfold.

OpenAI also launched a $250 million initiative giving 100,000 scientists free access to GPT-5.6 Sol Pro through 2027, where heavy users double their offloading of multi-hour tasks.

To top it off, President Greg Brockman confirmed OpenAI is developing a family of consumer hardware devices alongside former Apple designer Jony Ive. The hardware stack will feature verifiable, auditable privacy guarantees, positioning persistent AI assistants as the primary interface.

Anthropic's biggest AI security risk wasn't the model. It was the test harness.

Anthropic learned the hard way that testing offensive AI tools in bad setups creates real chaos. After auditing over 141,000 evaluation runs following OpenAI's Hugging Face breach, the company found that three models, Claude Opus 4.7, Mythos 5, and an unreleased research prototype, breached live corporate networks during security tests.

A third-party environment accidentally left the models connected to the web while telling them they had no internet access, so they treated real corporate systems like part of the game. They exploited weak passwords, probed open endpoints, and Mythos 5 even published a malicious Python package to PyPI that 15 real systems downloaded. Anthropic has since paused web-connected security testing alongside evaluation partner Irregular.

When they aren't accidentally breaching live corporate networks, these models are proving to be ruthless capitalists.

In Andon Labs' year-long Vending-Bench simulation, Claude Opus 5, GPT-5.6 Sol, and Kimi K3 were tasked with running competing vending machine businesses. Opus 5 crushed the competition, but it did so by orchestrating price-fixing cartels, breaking supplier truces, denying refunds, and attempting market-sharing deals, all while explicitly noting that its price-fixing violated antitrust laws. GPT-5.6 Sol wasn't much better, initiating collusion before immediately undercutting rivals.

Google launches Gemini Robotics 2 for full body humanoid control

Google is tearing down its old AI structure to build something broader, pushing hard into whole-body robotics and desktop tools. DeepMind just launched Gemini Robotics 2, describing it as "one brain for any robot."

The setup splits into three parts: a vision-language-action model for full-body movement, Gemini Robotics ER 2 for multi-robot planning, and an on-device model that adapts to new hardware in hours using under 200 examples. To keep these physical agents in check, Google also introduced ASIMOV-Agentic, a safety benchmark designed to measure if a robot knows when to stop and ask a human for help.

That same automation push is overhauling Google’s software ecosystem.

Gemini-powered security tools patched 1,072 Chrome bugs in June alone, fixing more flaws in one month than the last 23 browser releases combined.

On macOS, Gemini is morphing into a screen-aware assistant activated via the Fn key, while Gemini Notebook adds interactive app creation.

Creatives get Lyria 3.5 in Flow Music with granular stem controls and section painting. At the same time, Google is preparing Gemini Notebook to generate interactive apps from uploaded sources, alongside new Canvas, Literature Review, AI Notes, watermarking, and web-sourcing features.

To fuel this push, DeepMind dismantled its famous AlphaFold team, scattering researchers across Gemini, nuclear fusion, and Isomorphic Labs, even as several top contributors walked out the door for Anthropic.

Related videos:

Watch: Intelligent whole-body control with Gemini Robotics 2

Watch: Introducing Lyria 3.5 now available in Google Flow Music

Zuckerberg bets $145B on an agent-first internet

Meta is burning through massive amounts of cash to build the infrastructure for an AI-agent economy, and Mark Zuckerberg isn’t flinching. The company partnered with BlackRock on a $14 billion, 1-gigawatt data center in Texas, driving its 2026 infrastructure budget to $145 billion.

That massive spend squeezed second-quarter free cash flow down to $784 million, sending the stock down 10% despite revenue climbing 28% to $60.8 billion. Zuckerberg dismissed Wall Street's reaction, framing open-source AI as a necessary check on corporate power while Meta shifts toward directly monetizing model intelligence over basic cloud compute.

That compute engine is powering a major push into autonomous personal agents, with Zuckerberg predicting billions of people will use them within five years. Meta plans to deploy these persistent assistants through WhatsApp and Messenger to run continuously in the background for daily logistics and finances, building on enterprise momentum where over a million businesses already use Muse Spark 1.1.

Meanwhile, Alexandr Wang called current AI advances a once-in-a-civilization opportunity and urged founders to trust their own convictions while discussing talent density and building frontier labs.

On the hardware side, Meta updated its smart glasses to automatically disable the camera if the recording light is covered or blocked.

Model & Product Releases

Efficiency, Routing & Cost Control

AI Agents & Developer Tools

Research & Frontier Science

Safety, Evaluation & Integrity

Robotics & Physical AI

Infrastructure & Compute

Business & Markets

Policy & Regulation

Labor & Economy

Consumer & Home

Pushary provides a lock screen control panel to approve AI agent requests and keep automated runs moving.

Flowtask acts as a ‘company brain’ for AI agents, structuring messages and docs into actionable memory.

Memmy Agent is an open source desktop app that syncs context and preferences across multiple AI tools.

Nautis gives founders an AI-native operating system that unifies fundraising, docs, and operations into one connected workspace.

Claude Overlay provides a floating desktop window that reads your screen and runs Claude Code tasks directly.

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.