In partnership with |  |
|
|
Good morning. It’s Wednesday, September 2nd. |
Cool drop this week: World Labs' new Atlas model generates a full minute of 1440p video with precise camera control, plus explicit 3D output for real-to-sim robotics work. Early access is enterprise-only for now, but the demos are worth a look. | -Jeff AI Breakfast |
|
|
You read. We listen. Let us know what you think by replying to this email. |
|
AI made PMs faster. Multiplayer mode is still broken. |
|
A PM can summarize research, draft a PRD, and mock up a prototype before lunch. The hard part starts when the team has to decide what actually gets built. |
Jira Product Discovery gives product teams one place to capture insights, prioritize ideas with consistent frameworks, and build living roadmaps stakeholders can rally around. |
And because it’s connected to Jira, the context behind every decision stays with the work—so developers and their agents know not just what to build, but why. |
AI helps PMs move faster. Jira Product Discovery helps the whole team build with confidence. |
Get Started for Free. |
|
OpenAI gates Astra after autonomous zero-day finds |
OpenAI is slamming the brakes on its upcoming Astra model after it became the first system to cross the Critical threshold under its Preparedness Framework. In internal testing, Astra scored a perfect 100 percent on ExploitBench and independently discovered two zero-day vulnerabilities in a V8 benchmark, chaining them into working exploits without human help. |
That sudden jump in power comes from a ‘recurrent depth’ design that loops inputs through model layers multiple times. While this approach boosts performance and slashes compute costs, it also hides the model's internal reasoning, making its decision-making much harder to trace. |
To prevent a repeat of last month's Hugging Face environment breach, OpenAI is layering on chain-of-thought monitoring, tougher jailbreak defenses, and strict prompt limits. For now, access to Astra's full cyber toolkit will stay strictly locked down to select partners in its Daybreak defense coalition. |
At the same time, OpenAI is pushing hard into hospitals and big business. The new ChatGPT for Healthcare platform adds a HIPAA-compliant Epic EHR hookup and a Healthcare Public Data plugin that pulls in nine public datasets, including PubMed and DailyMed. |
ChatGPT Ads also hit a $1 billion annualized run rate in under 200 days, so the company is expanding its self-serve ad manager worldwide. And to celebrate Codex hitting 25 million active users, OpenAI just reset usage limits for paid developer plans. |
|
Anthropic rolls out Fable 5.1 after sandbox escapes |
Anthropic just dropped Claude Fable 5.1 and its restricted sibling Mythos 5.1, bringing serious agentic upgrades alongside a price cut. Fable 5.1 posts major benchmark jumps, hitting 52.6% on Terminal-Bench-Science and 55.8% on Terminal-Bench 4.0. |
It also slashes cache-read costs by 75% down to $0.25 per million tokens. That should save heavy agentic workloads up to 45%, though independent testing notes max-effort tasks can run 20% pricier because the model generates more output tokens. |
On the research front, Mythos 5.1 achieved a near 50% hit rate on viable protein binders and helped map a third of Venus. The update loosens up overzealous guardrails, cutting false-positive triggers by 60% in cyber and 85% in biology while still blocking exploit creation. |
Anthropic is also introducing Enterprise Frontier Safeguards, letting business customers store monitoring data and encryption keys directly in their own cloud infrastructure. Plus, to meet EU AI Act mandates, Anthropic opened its text-watermarking API to regulators and media, though critics worry synonym-steering might degrade writing quality. |
All this comes right after some messy containment failures. Anthropic released new details behind the recent safety incidents and blamed misconfigurations and reward hacking. It lead to reassigning 150 engineers to security and deploying real-time tool-call blockers. But safety scares aren't slowing down its expansion: Anthropic just locked in massive cloud deals, including $35 billion with Lambda in Texas and $45 billion with Nscale in West Virginia. |
Related videos: |
Introducing Claude Fable 5.1 |
Claude Fable 5.1 runs the forecast overnight |
|
Google adds agentic video scanning to Gemini Flash |
Google is rolling out new agentic video understanding to Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Instead of processing full files at fixed frame rates, an internal agentic loop dynamically inspects relevant clips, audio, or transcripts. This cuts token consumption by up to 88% and slashes costs by up to 66% while enabling sub-second moment retrieval across multi-hour footage. The feature is live via the Gemini API and Enterprise Agent Platform, with consumer rollouts heading to the Gemini app and YouTube. |
On the developer side, Google is readying Gemini 3.8 Flash, code-named Skimaki. In internal testing using Google's Jetski coding environment, engineers preferred 3.8 Flash over Anthropic’s Opus. Google also showcased its Antigravity multi-agent framework, which solved seven open math problems with Lean-verified proofs and built a working RISC-V CPU simulator. To keep multi-day agent runs from bogging down, researchers created SKILL.state, a runtime system that swaps growing chat transcripts for a lean, mutable execution state. |
Google also launched Google Pics, a Workspace image editor powered by its Nano Banana model that offers object segmentation and inline text translation directly inside Docs, Slides, and Drive. Notebook also gained Interactive Reports, allowing users to build structured projects with embedded spreadsheets and web pages. |
|
|
|
Gauth AI Course turns any subject into interactive lessons featuring mindmaps, quizzes, and real-time tutoring. |
Radar by Particle makes thousands of transcribed podcasts searchable for users and AI agents via API. |
oMLX turns your Mac into a native LLM inference server that reduces agent wait times using persistent caching. |
ThunderPhone builds reliable AI phone agents starting at 2¢ per minute with models included across 47 languages. |
Computable GPU Index delivers an open-source, mathematically verifiable USD price per GPU-hour calculated from published provider rates. |
|
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. |