I audited every model the pipeline calls apart from the script writer and brought them all up to date. Where a stage only needs a good fast model, it now asks OpenRouter for the latest version of its family, so it moves to a new generation without anyone touching the code. Titles, descriptions and the blog post come from Gemini Flash, as structured output with length limits, and the blog post is much shorter. Covers are drawn by a newer image model that is shown earlier covers as a style guide.
- Title, description, URL and image prompt: one Gemini Flash call that returns fields rather than free text, with limits on each (title up to 60 characters, description 50 to 90 words). A miss is regenerated. Before this, one DeepSeek call wrote all of it plus the blog post in labelled text.
- Blog post: a separate call, about 100 words instead of 150 to 220.
- Voice-memo transcription: Gemini Flash, latest version, instead of a Gemini 3.1 Flash Lite preview. Whisper is still the fallback.
- Search embeddings: the released Gemini Embedding 2 instead of its preview. The vectors are identical, so nothing had to be re-embedded.
- Covers: Krea 2 Turbo Style, given three earlier covers as style references, about a cent per cover. Flux Schnell, the previous model, is the fallback.
- Embeddings stay on a fixed version, because a silent change there would invalidate every stored vector.
- The experiment and interview episode types were removed: the guest models they relied on had been retired.