Not the copy. Not the storyboard. The actual frames. Here's the thing most people don't know. There's a library called Remotion that lets you build video in React. Every scene is a component. Every animation is code. You hit render and an MP4 comes out the other end. Which means a model that's good at writing React is now good at making video. 🎬 So I pointed Opus 5.5 at it and had it build a 38 second promo for one of my platforms. Three acts and a finale, each one its own file. Stat cards that count up. A pipeline view that slides in. A wall of tool logos that the camera dives through into a photo. Lock screen notifications stacking up. End card. No After Effects. No editor. No motion designer. No stock footage. The screenshots in it are real product, too. We wrote a small script that boots the actual app locally, swaps in a fake company name and fake customer data, hides the demo banners, and screenshots it. So the UI in the video is the genuine interface, just wearing someone else's brand. That last part is the bit I'd pay attention to if you run an agency. 💡 Because the whole thing is code, not a timeline in an editing app. Change the colors, swap the logo, point the capture script at a different app, hit render. You get a new video. Same build, different client. A promo like this from a studio runs a few thousand dollars and a couple of weeks of back and forth. The first one you build this way costs you a weekend. Every one after that costs you a render. That's not a cheaper video. That's a different business model. You're not selling a deliverable anymore, you're selling output from a system you own. Three things so nobody thinks this was magic. This was version two. The first attempt had timing that felt off and a scene that didn't land, and I rebuilt it. You're still directing, the model is just doing the production. If you've been outsourcing video or avoiding it because of what it costs, that math changed and most people haven't noticed yet.

Anthropic shipped Opus 5.5 a couple weeks back. I waited for the independent numbers before saying anything, and they're in. Artificial Analysis, who test everyone on the same harness instead of taking vendor word for it, now rank it first on their Intelligence Index at 58. GPT-6 Astra and Fable 5.1 are tied behind it at 53. They called it the highest score they've measured by several points. It also got cheaper. $4 per million input, $20 output, down from $5 and $25. Remember the Grok post a couple weeks ago? Same price per token, but it burned double the tokens, so the real cost per job went up. 📉 This is the opposite case, and it's worth understanding properly, because the honest version is more interesting than the headline. Opus 5.5 also uses more tokens than Opus 5 did. About 1.6 times more at maximum effort. But the price came down enough to absorb it, so Artificial Analysis found cost per task landed roughly level with Opus 5, while the score went from 51 to 58. Flat cost, meaningfully better output. That's the actual trade, and it's a good one. Just not the same thing as "40% cheaper," which is Anthropic's number for typical workloads at default settings, not what you'll see if you run everything at max. Now the number I'd actually act on. 💸 Cache reads dropped from 50 cents per million tokens to 20 cents. That's a 60% cut. If you run agents, that's your biggest line item and you probably don't think about it. Every turn an agent takes, it re-reads the same system prompt, the same tool definitions, the same conversation history. A 500,000 token context that cost 25 cents to re-read on Opus 5 now costs 10 cents. Nothing about that gets offset by token burn. It's just cheaper. Two more things worth knowing. Anthropic raised the five-hour limits on Pro, Max and Team at the same time, and gave subscribers a one-time rate limit reset they could trigger themselves. If you read my post about the weekly limits dropping in September, this softens that a bit. And there's a Fast mode in Claude Code at $8 and $40 for up to two and a half times the speed, if you're sitting there waiting on it.