How the AI works
Why a model at all
A rules engine can find the word “interesting”. It cannot find the best ninety seconds in an hour of conversation, and it cannot tell you whether a moment has the tension to make someone stop scrolling, because that judgement lives in the thousands of episodes the model has read, not in a regular expression.
The second half is harder. Rewriting a podcast moment as a hook for a vertical clip is a writing problem. The hook has to work in the first second, the caption has to earn the follow, and both have to sound like the creator who said it. A template cannot do that. A language model can.
So the model is the engine. Everything else in this product is plumbing around it: getting the transcript into one pass, routing each step to the model that does it best, and refusing to show you a moment that is not in your own recording.
The three steps
Before anything else, Verso reads a short sample of the transcript and works out what kind of recording this is: the topic, whether it is a solo episode or an interview, and which output mix suits it. This step is cheap, fast, and feeds the moment-detection pass that follows.
The full transcript goes to the balanced model, which finds the moments worth pulling and rewrites each one five ways: a clip script with a hook and caption, a thread, shorts captions, a LinkedIn post and a newsletter issue. The model is told to quote the transcript verbatim and to leave out any moment it cannot locate exactly. One run produces the whole board.
On Creator and Studio, any output on the board can be regenerated with one click. The same routed model, the same voice sample applied, a different pass over the same moment. Nothing is rerun unless you ask for it, and the original stays on the board until you replace it.
Anti-fabrication guarantee
The model is instructed to quote verbatim and to omit any moment it cannot quote exactly. Before a moment reaches your board, that quote is located in your transcript, first exactly, then allowing for whitespace and curly quote differences, then by the longest matching run that does appear. A quote that cannot be found is dropped and counted. This is the only part of the pipeline the model does not get a vote on.
Routing
This is not a diagram somebody drew. It is rendered from the same routing table the repurpose engine reads, so if it is wrong here it is wrong in production.
| Step | Model | Provider | Tier | USD per M tokens |
|---|---|---|---|---|
| Find the moments and write every format | gpt-4.1-mini | openai | balanced | $1.60 |
| Rewrite one format again in your voice | gpt-4.1-mini | openai | balanced | $1.60 |
| Work out what the recording is about | gpt-4.1-nano | openai | fast | $0.40 |
Above roughly 30,000 characters, about sixty minutes of transcribed audio, the repurpose pass escalates to the frontier model rather than the balanced one. That is where long-recording accuracy starts to matter more than cost, and where a single pass over the whole transcript is worth paying for.
If the routed model fails or returns nothing, the call walks a chain of candidates from other tiers and other labs before giving up. One provider having a bad afternoon should not cost a creator their project, and it does not.
Your transcript is sent to a model provider to produce your outputs and for nothing else, under agreements that prohibit training on that traffic. Delete any project, or the whole account, whenever you like.
Candidates
Every model below sits behind the same interface. Changing a route is one line in the routing config. Adding a lab is one case in one file, which is the entire point of building it this way.
openai · 1,000,000 token context
openai · 1,000,000 token context
openai · 1,000,000 token context
anthropic · 200,000 token context
google · 1,000,000 token context
meta · 128,000 token context
In this deployment the OpenAI and Anthropic adapters are written and the OpenAI one is live. The remaining providers are declared with their real model names and activate the moment their key is present. No model is trained on creator transcripts, by Verso or by its providers.