FLUX 3 blends real & digital worlds

Welcome back, Superhuman. Black Forest Labs just bridged the gap between the physical world and the digital world. Its new model FLUX 3 creates outputs so realistic they can be used to train real-world robots. Plus, voice AI keeps getting better and better.

Today: Health in ChatGPT rolls out, how to evaluate business ideas quickly with AI, and get the latest prompts and trending social posts.

1. Black Forest Labs launches one model for image, video, and robotics: The German lab just released FLUX 3, a multimodal foundation model that combines image, video, and audio elements to gain a better understanding of how the real world works and create results the lab calls “truer to life.” A version of the model is already powering real-world robots deployed at Audi. FLUX 3 Video is available now in early access. See what the model can create.

2. OpenAI and Anthropic make it easier to command agents with your voice: You can now direct agents on your computer using ChatGPT’s desktop app, powered by the recently updated ChatGPT Voice. It’s available now on most paid plans on macOS and Windows. Anthropic also pushed an update: Claude’s voice mode now runs on Opus/Sonnet, giving the voice agent a higher level of reasoning. It can also connect to tools like your email or calendar mid-conversation.

3. Health in ChatGPT gets a broader release to US customers: Health in ChatGPT is getting a wider rollout after launching in January. The platform connects to Apple Health and other tools to offer personalized insights, like comparing old test results to new ones, summarizing appointments, or helping you navigate physical therapy after an injury. OpenAI says 300M people ask health-related questions each week, but stresses that AI is not a replacement for professional medical care. Watch how it works.

Yesterday’s most-clicked story: OpenAI launched Presence, a new platform that helps companies deploy voice and chat agents across their enterprise.

Shipping agents to production is the easy part. Keeping them reliable, governable, and improving over time is where most enterprise AI programs stall.

How do top teams do it? They use an Agentic Operating Model (AOM), a step-by-step framework for aligning people, process, and technology so enterprise agents improve as they scale.

Download LangChain’s guide to learn how.

Claude Fable 5 and GPT-5.6 Sol continued to push AI’s limits. Claude Fable 5 helped mathematician Levent Alpöge find a counterexample for a math problem that’s stumped human researchers for nearly a century: the Jacobian Conjecture. Meanwhile — in another unprecedented event — OpenAI’s GPT-5.6 Sol autonomously broke out of a closed environment during a training session to hack the AI community, Hugging Face.

The White House accused a Chinese lab of copying Anthropic's models. Michael Kratsios, the administration's top technology advisor, called out Moonshot AI for allegedly distilling US models to mimic their performance in its own products. The accusation comes as Chinese models have increasingly matched US leaders on top benchmarks. Critics have pointed out the irony, as many US models were built on scraped data.

Google posted strong Q2 earnings and released three new models. Alphabet’s revenue jumped 24% to $119.8B, while cloud revenue soared 82% to $24.8B. That’s a positive sign that Big Tech’s massive AI infrastructure spending is translating to revenue gains, at least for now. Google also shipped three new models designed to make agents cheaper and more dependable: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber.

Decart brought AI editing features to live video. The startup's latest model, Lucy 2.5, can edit footage in real time with minimal lag. It lets you add or remove objects, change the visual style, or layer in effects as the camera rolls. The explainer video went viral (4M views) and signals a future where live broadcasts could carry the same visual polish as post-produced film.

🤯 Making History: AI is getting mighty close to solving math. Shouqiao Wang just solved 6 open Erdős problems in 5 days, while Dmitry Rybin bested the Dinitz-Garg-Goemans conjecture — and they did it with shockingly simple prompts.

🤖 OpenWorker: Andrew Ng rolled out OpenWorker, an open-source AI assistant that can handle tasks across your desktop and lets you easily swap between LLMs to prevent vendor lock-in (5K bookmarks).

😅 Nightmare Fuel: These throwback clips show just how much AI video has improved since 2023. A top comment calls them “a bit horrifying” in retrospect (3K upvotes).

🍎 Redefining Education: An AI-powered white board has racked up 5M views for the impressive way it spins up custom graphs and visualizations on the fly. The product demo offers a glimpse into what education might look like in the AI era.

😆 You’re Absolutely Right: The internet has finally coined a name for AI’s speech pattern that we all recognize instantly — confidently useless.

That's one finding from Clarinet's mid-year State of AI report: what's working with AI in the People function, what's stalling, and what to watch in H2. Built from 200+ leaders with real-world expert advice.

Get the report

Grow customers & revenue: Join companies like Amazon, HubSpot, and Salesforce. Showcase your product to our 4 million+ readers and followers on socials. Get in touch.

Until next time — Zain, Theodore, & the Superhuman AI team.