Black Forest Labs teaches video AI to run robots. PLUS: Edit videos faster with AI (for non-editors) July 24, 2026 Good morning, AI enthusiasts and welcome to the 6,243 new readers who joined The Rundown AI, Inc. yesterday. Black Forest Labs made its name in the AI image world. Its newest model, while able to create images, video, and audio, is now also reporting for work on an Audi production line. FLUX 3, the company's 'visual intelligence' model, generates realistic video by simulating how the world actually moves - the same skill that is now helping train and steer machines in the physical world. In today's AI rundown: * FLUX 3 stretches from videos to factory robots * Rowan's Corner: How The Rundown AI, Inc. is building AI intuition at The Rundown * Edit videos faster with AI (for non-editors) * OpenAI, Anthropic launch dueling voice upgrades * 4 new AI tools, community workflows, and more LATEST DEVELOPMENTS Black Forest Labs. The Rundown: Black Forest Labs just opened early access to FLUX 3, a 'visual intelligence' AI that generates video, images, and synced audio, with the German lab also spinning it into FLUX-mimic, a robot-control variant for Audi's production lines. The details: * FLUX 3 produces 20-second clips with native audio, with capabilities including multilingual dialogue, typography, and style/character adherence. * The model can handle text, image, and video inputs, with BFL's testing showing its outputs favored over rivals like Runway, Kling, and Grok Imagine. * The full rollout will include video, image, and partner-specific action models, plus an open-weight FLUX 3 Dev sized to fit on factory hardware. * FLUX-mimic, built with Zurich-based mimic robotics, learns a new factory task from about a half-hour of demonstration data instead of the usual 30+ hours. Why it matters: BFL has typically moved at a slower pace, but the unified system looks like the payoff - with longer, audio-synced videos and coherent visuals (like this hilarious one) reminding The Rundown AI, Inc. of the Sora days (in a good way). For mimic, visual models continue to creep into the physical world, becoming one of the key cogs in robotics. Together with AWS. The Rundown: AWS re:Invent is bringing 2,200+ sessions to Las Vegas on Nov. 30 through Dec. 4, and 70% of them are hands-on. This year, choose to: * Work through real problems in 300+ workshops and 250+ builders' sessions * Debug alongside peers in 125+ code talks * Whiteboard your architecture in 350+ chalk talks with the engineers behind the services Rowan's Corner. Rowan: Last week, I broke down the most important skill of the AI age: AI intuition, knowing when to use AI and when not to. This week, I want to go deeper into how I'm actually applying it across the team at The Rundown. The chart above is its role-by-role AI fluency standard, mapping what unacceptable, expected, and exceptional AI usage looks like. Even as an AI-first startup, a framework like this sets the standard with real examples. Some more rules from its AI fluency memo: * AI is an amplifier, not a crutch. If your work gets worse because of AI, that's on you. "AI did it" is never an acceptable excuse. On the flip side, refusing to experiment means you won't keep up with high performers or AI agents. * Everyone spends 2 hours a week finding the tedious parts of their job and building workflows for them, then shares wins in its #ai-show-and-tell channel. * Everyone gets $200/month for AI tools. The only condition is that you have to name the problem you're trying to automate. Not to gatekeep spending, but to make sure every experiment starts with a real problem. The goal behind all of it is reps. AI intuition works like business intuition: nobody is born with it. You build it by pushing tools to their limits until knowing when to reach for AI becomes instinct. If your team hasn't set a standard like this yet, I highly recommend it. Here's a downloadable template of my setup you can clone. AI training. The Rundown: In this guide, you will learn how to edit videos using AI, as The Rundown AI, Inc. do at The Rundown, and then teach ChatGPT to edit for you hands-free. Step-by-step: * Open Descript's desktop app, record your camera and your desktop if you like * When you're done recording, duplicate the project before editing * On the duplicate, use the AI tools "Remove filler words" and "Shorten word gaps" to clean filler words and remove delays longer than one second. Turn on "avoid harsh cuts" so the edits sound natural * Run "Edit for Clarity" on Low and go through the AI suggestions Pro tip: Connect the Descript plugin in ChatGPT or Codex, and teach it your editing process for hands-free editing. Presented by Strands Agents. The Rundown: Your agent harness can have a bigger impact than the model you're running. With Strands Agents, you can build your own agent harness and control it end-to-end without getting locked in. Run it with any model, any cloud, anywhere: * Monitor, modify, and debug with hooks * Deliver outcomes that work with guardrails that catch mistakes before they run * Replace brittle scripts with tools that adapt when processes change Openai & Anthropic. The Rundown: OpenAI and Anthropic shipped voice upgrades at nearly the exact same time, with OAI moving its GPT-Live voice into the desktop app and into Codex, and Anthropic plugging Claude's voice into its smarter models and connected apps. The details: * OpenAI's ChatGPT Voice joins the desktop app on macOS and Windows for paid plans, letting users talk to agents in Codex and ChatGPT Work via speech. * A custom hotkey calls the voice mode over any app, with Bloomberg reporting that the same tech may drive OAI's upcoming home speaker hardware. * Claude's voice mode now runs on Opus and Sonnet rather than only its small Haiku model, while also being able to connect to apps like Gmail and Slack. * Claude lags behind with turn-by-turn speaking, with the recent launch of GPT-Live pushing a flow that listens, thinks, and replies at the same time. Why it matters: Both labs shipped something useful, just not what many users were looking for after rumors of a Claude 5 Opus drop and Cerebras-powered speed upgrades for OAI. Even with the upgrade, Anthropic feels behind its biggest rival on voice, with speech continuing to grow as a primary way people work with AI. QUICK HITS * | FLUX 3 - BFL's multimodal AI with 20-second generations * | Poke - Proactive personal agent that lives inside text messages * | Health in ChatGPT - Link medical apps and records to chat about health * | GPT Live - OpenAI's more natural voice model, now on Desktop and Codex Cognition announced the acquisition of The Interaction Company, with its text-message-based AI agent Poke now being built alongside Cognition's Devin agent. U.S. AI chip startup Etched secured $300M in new funding just weeks after exiting stealth, with the company now valued at $10.3B and pushing its total funding past $1B. Canadian mathematician Jacob Tsimerman is joining OpenAI to work on AI safety, with the news coming on the same day he was awarded the 2026 Fields Medal. Microsoft's superintelligence team dropped two in-house models, with MAI-Image-2.5-Pro as its new image AI and MAI-Voice-2-Flash as a faster, cheaper voice engine. Runway launched Media Router in its dev platform, which auto-picks an image, video, or audio model for each request based on user preference for cost, quality, or speed. Every newsletter, The Rundown AI, Inc. showcase how a reader is using AI to work smarter, save time, or make life easier. Today's workflow comes from reader James in Lancaster, PA: "I found a professional DJ controller on Facebook Marketplace that was damaged by a water spill and did not power on. I copy/pasted the entire conversation with the seller into Claude, which produced a detailed risk assessment and guidance on price negotiation. I ended up purchasing the unit for only $800. I uploaded the PDF service manuals to Claude, who guided me through the professional disassembly and diagnostic process. I was uploading photos of circuit boards, and Claude was describing the exact terminals where I should make voltage measurements. The faulty component turned out to be a $30 sensor strip inside the jog wheel. I replaced the part, reassembled the unit, and it powered on instantly! Claude helped me obtain a $2,000 professional DJ controller for only $800, a $30 replacement part, and a few hours of DIY tech service." How do you use AI? Tell The Rundown AI, Inc. here. See you soon, Rowan, Zach, Shubham, and Jennifer - the humans behind The Rundown

Runway just shipped a model router. The model wars are over. Runway's new Media Router auto-picks the best video, image or audio model per request. Here's what it signals - and what it means for ecommerce ad teams. Yesterday Runway announced Media Router, which it calls "the world's first model router for generative media models." It's a new capability inside Runway Dev, the API platform the company launched earlier this month: instead of hand-picking a video, image or audio model for each generation request, you send the request to one endpoint and the router picks the model for you. That sounds like a small developer convenience. It isn't. It's the clearest signal yet that the media-model layer is commoditizing - and that even Runway, a company that trains frontier models, knows it. What Runway actually shipped. The facts, straight from the announcement: * One endpoint, no model parameter. You define a reusable config in the Runway Dev portal - optimization preference across cost, quality and latency - then call the Media Router endpoint with the config ID attached. The router matches your request to models that can actually handle it (capability and modality fit), scores the qualifying ones against your preferences, and generates with the winner. * Hard constraints, not vibes. Launch-day config dimensions are a price cap (a hard ceiling in USD per second for video, per image for image models), allow/deny lists at the provider and model level, and your cost/quality/latency weighting. * It fails loudly. If no model satisfies your constraints, the router returns an explicit error naming which constraint emptied the pool instead of silently downgrading your settings. A dry-run flag lets you preview exactly which model a config would select - and why - with no cost incurred. * Full transparency on the pick. Every result comes back with metadata identifying which model was used and the reasoning. * The catalog spans first- and third-party models. Runway's own Gen-4.5, Aleph 2.0 and Act-Two sit alongside Seedance, GPT Image 2 and ElevenLabs. It's live now for video, image and audio at dev.runwayml.com. Runway's stated rationale is telling. Builders, they write, "don't want to become part-time model researchers," and teams routinely "pick a model once and stick with it long past the point it's the right call." At enterprise volume, a slightly-wrong default model becomes a real cost and latency drag. Why this matters: the value is moving up the stack. Think about what it means when the company behind Gen-4.5 ships a product whose whole premise is you shouldn't care which model runs your job - including when the answer is a competitor's model. It means the model layer is becoming plumbing. LLMs went through this exact transition: once routing across providers became normal, the durable value moved to whoever owned the workflow on top. Media is now following the same curve, and faster, because the model catalog churns every few weeks. Nobody wants to re-benchmark Veo vs. Kling vs. Seedance every time pricing shifts or a new checkpoint drops. Runway's own AI Media Report, published three days before the router, supplies the context: across hundreds of enterprise customers, production costs are falling by two to three orders of magnitude, and one game studio's marketing team went from 13 to 75-100 ads per week with the same headcount. When generation itself gets that cheap and that fast, model selection stops being a craft decision and becomes an optimization problem - which is precisely what a router is for. So the announcement is really a concession dressed as a launch: quality differences between frontier media models are converging enough that a scoring function can arbitrate them. The model wars aren't ending because someone won. They're ending because the answer increasingly doesn't matter. The honest take for ecommerce advertisers. Here's where Prizmad'd push back on the framing that routing solves the problem for advertisers: a media buyer doesn't have a model-selection problem. They have a finished-ad problem. Media Router lives in Runway Dev - it's infrastructure for developers building media products. If you run a Shopify store or a performance team, "call one endpoint with a config ID" still leaves you holding the script, the voiceover, the avatar, the B-roll, the captions, the music and the final 9:16 cut. Routing picks the best model per generation request; it doesn't decide what the requests should be, in what order, or how the outputs get assembled into something you can upload to TikTok Ads Manager. That's the layer where Prizmad think the real leverage sits, and it's frankly why Prizmad is built the way it is: its AI Studio runs 30+ models under one subscription, and the ad pipeline picks the right model per job - one model for the talking-head render, another for product B-roll, another for audio - so the thing that comes out the other end is an ad, not an asset. Same thesis as Runway's router, applied one level up the stack where advertisers actually live. None of this is a knock on Runway - their tools are genuinely strong, and its hands-on Runway review covers where they shine for ad creative. And the economics of the whole category keep improving regardless of who assembles the stack; its AI video ad cost breakdown shows a finished ad already lands in the single-to-low-double-digit dollar range. What to watch next. Three things over the coming months: * Whether routing quality becomes measurable. Runway's router scores "quality" as a preference dimension. The first public benchmarks comparing routed vs. hand-picked outputs will tell us how commoditized the model layer really is. * Whether competitors respond in kind. If fal, Replicate or Google ship preference-based media routing, the pattern is confirmed and per-model pricing power erodes further - which ultimately flows through to cheaper ads. * Whether routing climbs the stack. The obvious next move is routing entire multi-step jobs, not single generations. The moment "make me an ad" becomes the unit of routing rather than "make me a clip," the infrastructure layer and the workflow layer collide. Runway just told the market that model choice is a solved problem. The open question - the one that decides who captures the value - is who owns everything that happens before and after the model runs. Generate your first ad in 5 minutes. Paste a product URL. Prizmad writes the script, picks the avatar, renders the voiceover with lip-sync, adds subtitles and music, and ships a TikTok / Meta / YouTube-ready mp4. No camera, no editor.

Runway.com has been acquired by Runway AI, a "real world intelligence company." The deal was overseen by Andrew Miller of ATM Holdings.