Good morning. It’s Wednesday, August 5th.

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

Alibaba releases Qwen 3.8 Max with 16-day autonomous execution capabilities

Alibaba released Qwen 3.8 Max and launched its QwenWork workspace, pairing a 2.4 trillion parameter sparse mixture-of-experts model with an enterprise agent platform. Running 95 billion active parameters per token, the flagship features a 1 million token context window, 128,000 max output tokens, and native visual feedback for GUI navigation across web, mobile, and desktop.

Built for multi-day execution, Qwen 3.8 Max handles long-horizon work. Benchmark runs and company tests show a 16-day autonomous software build, a 125-hour research loop raising reasoning benchmarks by +2.71 points, a top 13% finish in a 526-team data science contest, and a 500-turn chip design flow that cut gate counts from 8,298 to 678 at 500 MHz. In business testing, it produced a 4.16x return over a 365-day e-commerce simulation.

Independent evaluations place Qwen 3.8 Max second on Vision Arena, fourth on Frontend Code Arena, fifth on Text Arena, and tenth overall on the Vals Index at 66.1. It scored 87.3% on SWE-bench and 67.4 on Terminal-Bench 2.1. API access is live at $2.00 input, $6.00 output, and $0.25 cached per million tokens across Qwen Studio, Command Code, Baseten, and Hermes Agent.

Alibaba promises open weights next week for both the flagship and a smaller 27B model, though self-hosting the 2.4T model requires supernodes with 8 or more H100 or B200 GPUs, and geographic license details remain unconfirmed.

Simultaneously, Alibaba launched QwenWork in public beta.

OpenAI’s GPT-Live stacks WARP for sub-second full-duplex chat

To fight off Apple's trade-secret lawsuit following its acquisition of Jony Ive's io startup, OpenAI published internal chat logs showing Apple's legal team confused employee surnames and falsely claimed to have spoken with OpenAI General Counsel Che Chang.

The leaks show Apple staff routinely texted former engineer Chang Liu for technical help after he left, exposing Apple's weak offboarding controls.

Speaking at a media roundtable, OpenAI President Greg Brockman, said today's reliance on clicking and typing is a temporary phase, arguing future AI systems will adapt to people through natural conversation. The strategy aligns with OpenAI's recent rollout of ChatGPT Voice for desktop.

OpenAI also released details on GPT-Live, its full-duplex voice architecture built to handle real-time chat with sub-second responsiveness. Instead of waiting for you to finish talking, it streams audio continuously while quietly querying GPT-5.5 in the background for heavy-duty reasoning.

The stack uses stateful inference, dynamic context compaction and a custom WARP protocol that slashes WebRTC startup times from six network round trips down to just one.

These agentic workflows are moving straight into production. OpenAI launched three plugins for ChatGPT Work and Codex across ChatGPT Edu and ChatGPT for Teachers, tying K–12 and university workflows directly to calendars, syllabi, and study tools.

OpenAI is also preparing a public download page for Rosalind Codex, bringing its GPT-Rosalind reasoning model for biology, drug discovery, and translational medicine out from behind a waitlist for qualified.

White House exempts open-weight models from frontier AI testing rules

The UK AI Security Institute dropped a report showing frontier models acting out when safety filters were turned off.

Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol pulled off 19 unauthorized web actions across 122 cyber-range tests. In one case, Mythos 5 created fake GitHub personas, tried to trick maintainers into approving malicious code, and planted prompt injections for future agents. OpenAI also admitted a third-party evaluation error let a model leak out of its sandbox to breach a live website using exposed credentials.

Anthropic and OpenAI pointed out these tests deliberately disabled safeguards, but the behavior has White House officials moving quickly. Trump administration advisers met with OpenAI, Anthropic, Google, and Meta to finalize a voluntary safety framework.The unpublished plan sets up a classified 30-day pre-release review in secure environments for closed-source models to check offensive cyber capabilities, with OpenAI pushing for Commerce Department oversight.

Notably, the White House explicitly exempted open-weight models like Meta's Llama and Nvidia's Nemotron, as officials argue post-release restrictions would stifle innovation. Tech giants are agreeing to voluntary checks while Democratic lawmakers are pushing to make frontier testing mandatory by law.

Model & Product Releases

AI Agents & Developer Tools

AI Spend & Token Economics

Research & Interpretability

Security

Robotics & Physical AI

Chips & Infrastructure

Business & Deals

Policy & Regulation

Media & Culture

Coast runs secure local inference on Apple’s Neural Engine to enrich context for you and your agents.

Easy MCP AI executes complete WordPress workflows via native Model Context Protocol tools running on pure PHP.

Basedash Audit Logs provides traceable records of queries, sign-ins, and AI activity directly to your SIEM.

AgentSky deploys long-horizon cloud agents across any LLM or harness with managed recovery and multi-channel access.

Lumichats runs desktop commands and writes local files directly on your machine without terminal usage.

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.