Episode summaries are now written by Claude Opus 5.5, episode pages link to the episode itself, and long M4A episodes now transcribe. This release also cuts down on failed episodes from Substack-hosted shows.
What’s new
- Summaries are written by Anthropic’s Claude Opus 5.5. In side-by-side tests on real transcripts it was more specific, stayed closer to what was actually said, and kept nearer the 400 to 600 word target than the GPT model it replaced.
- Acronyms are spelled out the first time they appear in a summary or deck, so a guest’s “UGC” or “ICP” comes with its full name.
- Episode pages link to the episode on Apple Podcasts, plus the publisher’s own episode page and the podcast’s website where those exist. Most episodes used to link to the show’s Apple page, and few had a website link.
Fixes
- M4A episodes larger than 25 MB transcribe. They used to fail every time, because the file was split in a way that left every piece after the first unreadable. Each piece is now a complete audio file of its own, cut on audio frame boundaries.
- Episodes published as raw AAC get wrapped in an M4A container before transcription, since OpenAI’s transcription API won’t accept raw AAC.
- Audio that still can’t be handled (fragmented MP4, non-AAC audio inside an MP4, a video file, a truncated download) fails right away with a message that says what’s wrong. It used to retry three times and end with “corrupted or unsupported”.
- Substack-hosted podcasts fail much less often. Substack throttles requests for the audio, so the pipeline now follows the full redirect chain to the CDN and retries a throttled hop.
- When a publisher reissues its back catalogue with new IDs, the monitor marks the old episodes as seen. Previously it worked through the archive one old episode per run.
- If the last part of an episode can’t be transcribed, the transcript is kept (and marked partial) only when that part is a short tail of at most 10% of the audio. A bigger loss fails the episode, so a summary isn’t written from half a transcript.
Under the hood
- Editorial decks and pull quotes run on gpt-5.6-terra, and tags on gpt-5.6-luna.
- MP4 and AAC chunks are rebuilt in the Worker without transcoding, and capped at 15 minutes each as well as 15 MB.
- New maintenance scripts: a one-command transcript rescue for episodes the Worker can’t fetch, a local check that every chunk of a file decodes, and Anthropic support in the summary A/B harness.