A lot happened under the hood this week - and your episodes will feel it. We upgraded the core AI engine, taught Claude to search the web while building episode outlines, and systematically hunted down a series of Gemini audio issues that were causing silent truncation and dropped dialogue. Episodes should generate more reliably and completely than ever before.
New Features
Live Web Research During Episode Creation
When you create a prompt-based episode, Claude now searches the web in real time while building the outline. Instead of relying solely on its training data, the AI can pull in fresh information, recent developments, and current context before writing a single word of your script. The result is outlines that are more grounded, more timely, and more relevant to what's actually happening in your industry right now.
Social Content Planning
We've completed a detailed design for an upcoming Social Content feature that will let you turn podcast episodes into short-form videos for TikTok, Instagram Reels, YouTube Shorts, and LinkedIn. The feature is in the design phase now - expect more on this soon.
Improvements
Upgraded AI Engine (AI SDK 6)
The underlying AI framework powering all of PodcastAI's generation features has been upgraded to its latest major version. This brings better compatibility across OpenAI, Anthropic, and Google providers, more consistent structured output, and a foundation for new capabilities going forward. The upgrade touched every part of the generation pipeline and was validated with thousands of tests.
Upgraded to the Latest Gemini Model
Script generation, brand extraction, and prompt processing now use Gemini 3.1 Pro Preview, the latest available version. This means better reasoning quality across the board for all Gemini-powered steps.
More Reliable Gemini Audio - Multiple Rounds of Fixes
We spent a significant portion of this week tracking down and fixing a cluster of related issues with Gemini text-to-speech that were causing audio to be cut short or dialogue to be silently dropped. The fixes include:
- Audio now processes in smaller, smarter chunks that stay within Gemini's actual limits, preventing silent truncation on longer episodes
- The system now detects when audio has been cut short and automatically retries, with smarter splitting strategies when chunks are too large to process in one pass
- Tiny "runt" chunks that were too small to process reliably are now merged with adjacent content before synthesis
- False positives that were triggering unnecessary retries (and adding minutes of wall time) have been eliminated with more accurate duration estimation
- Gemini's content safety filters, which were silently blocking podcast news topics and truncating audio, are now properly configured for podcast content
- Episode generation now runs faster with higher concurrency and tighter timeouts tuned to real production data
Host Voice Assignment Fixed
Gemini was sometimes swapping which voice belonged to which host, particularly during incremental section regeneration. This has been fixed with a more reliable calibration approach that consistently assigns voices to the right hosts. If you have older episodes with swapped voices, regenerating the audio will now produce the correct result.
Accurate Episode Duration
Silence padding between episode segments used to contain malformed audio data that confused duration-reporting tools, causing errors in downstream processing. This is now replaced with properly structured audio frames, bringing duration accuracy to within a few milliseconds. This also resolves crashes that were occurring when processing certain episodes.
AI Model References Stay Current
Hardcoded, dated model version strings have been replaced with stable alias versions throughout the codebase. This prevents "model not found" errors that would occur as specific model versions are retired by AI providers.
Bug Fixes
Fixed a compatibility issue where episode suggestions were failing due to schema constraints that the Anthropic API rejects. The suggestions pipeline now generates reliably again.
Fixed a bug in article relevance scoring where a URL format annotation was causing OpenAI's structured output API to reject the request. Article filtering now works correctly.
Fixed an issue in the legacy episode pipeline where extra AI context being sent with each audio chunk was pushing requests over Gemini's input limits, causing audio to be cut off at roughly 38% of expected length.