Issue #49 · Wednesday, October 21, 2026 · 6 stories

🎵 Spotify Pulls 50 Million Slop Tracks 📸 Instagram Won't Throttle AI Posts

A 15-second Claude Code edit for TikTok, and the AI tools people actually pay for.

Mass-producing AI content is getting less profitable, and this week's news shows why. Spotify has pulled 50 million slop tracks from recommendations, and Instagram ranks AI posts by whether people want them, with more labels coming. What still pays is taste: Brooke Lacey's Claude Code setup edits a TikTok in 15 seconds, but the real work was teaching it what's funny. Spending follows the same logic, with a16z's card data showing creators settling on one bundle, and tools like Eddie AI and Mistral's Voxtral competing on what each step costs.

NEWS

Spotify pulled 50 million slop tracks from recommendations, and likely-AI tracks are 10 times more likely to miss royalties · 4 min read

Spotify's one-year update on its AI rules says it has taken more than 50 million slop and spam tracks out of recommendations in 2026, on top of the 75 million it removed outright, a figure it has not updated since September 2025. The detail for anyone releasing AI-assisted music: tracks flagged as probably AI-generated are 10 times more likely to fall under the 1,000-stream royalty threshold, so they earn nothing. Detection combines listener skips, upload patterns, metadata and third-party AI detectors, and Spotify admits those detectors get weaker as models improve. Thousands of AI Persona artists are now labelled and kept out of recommendations unless a listener follows them, and artists file tens of thousands of AI Credits a day. Disclosure, not a ban, is the line it is drawing.

Instagram says it does not throttle AI content for being AI, and wants creators to label it themselves · 3 min read

Adam Mosseri said in his weekly Q&A that Instagram ranks AI posts like any other: reach depends on whether people want to see the post, not on how it was made. He said more AI labels are coming and that Instagram wants to reward creators who tag their own AI work. The problem he called tricky is AI accounts copying comedians' and dancers' routines, where copyright claims do little because most performers never registered the material. For AI creators the practical read is to label your own work and build on your own ideas, not someone else's bit.

TECHNICAL

A creator with 290,000 TikTok followers edits each video in about 15 seconds with Claude Code, FFmpeg and Whisper · 2 min read

Brooke Lacey shoots on her phone and hands the raw file to Claude Code, which transcribes it with Whisper and cuts it with FFmpeg on her own computer, about 15 seconds per video in the live demo. The hard part, by her account, was teaching it her jump-cut rhythm, her hooks and what counts as funny. She also has it read her TikTok comments to find the next script, and runs a once-a-week posting schedule with no editor. A segment covers what she does when Claude gets the edit wrong, which is the part to watch before you copy the setup.

ANALYSIS

Card data says creators pay for Higgsfield and Suno more than Runway or Kling · 5 min read

a16z's new list of the 50 AI products people spend the most on, built from US card data, includes 22 tools for images, video, music or voice: Canva 3rd, Higgsfield 5th, Suno 7th, ElevenLabs 10th and Midjourney 11th, though Midjourney is not among the 50 most visited AI sites. Heather Cooper's read is that bundles win the wallet: Higgsfield and OpenArt put many models behind one subscription, while Runway (25th), Kling (26th) and MiniMax (42nd), built around their own models, sit lower. Only 13% of people who pay for one AI product pay for a second. Veo and Seedance are missing because their spending sits inside Google and ByteDance platforms.

TOOLS

Eddie AI now shows the plan, and which steps cost credits, before it spends anything · 5 min read

Eddie AI's latest update, out the day before NAB Show New York, makes any prompt that would spend credits return a step-by-step plan first, marking the paid steps. You approve, change or drop steps, and nothing runs or bills until you do; choices like a colour grade come back as stills from your own footage. It also adds a DaVinci Resolve Studio panel that sends edits back as a new timeline, fills After Effects and Premiere motion-graphics templates, and lets teams bring their own Claude, ChatGPT or Gemini keys. A priced plan before the run is the feature to ask your current AI editor for.

Mistral's Voxtral Speech 3 costs $13 per million characters, under ElevenLabs Flash · 14 min read

Voxtral 3 arrived on 20 October: transcription models for 50+ languages and a text-to-speech model, Voxtral Speech 3, at $13 per million characters, which Mistral says undercuts OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. Zero-shot voice cloning is included, and the Pro Voice tier copies a speaker's character and range from 30 minutes of audio. Check the language list before you plan around it: the launch post claims 35+ languages, while Mistral's own model docs still list nine for TTS. Mistral pitches it at call-centre agents, but the per-character price is the number for anyone voicing narration at volume.