Plans and AI
Limits and AI models
How long a recording can be, what happens past that, and which AI does the work.
How long a recording can be
Every Mullo app records speech at 32 kbps, so the limit is about the same everywhere. Mullo's server takes one recording up to 44 MB.
| Where you record | Written down in one go |
|---|---|
| iPhone app, Android app | About 3 hours. The phone can be locked. |
| Apple Watch, Wear OS watch | About 3 hours. A Wear OS watch keeps recording up to 4 hours, but past about 3 hours it can't be written down in one go. |
| A browser | About 3 hours, if the browser keeps to Mullo's quality setting. On a phone, keep Mullo on screen while recording. |
| With your own Gemini key | About 2 hours: past that Gemini may not write it all down, and Mullo marks the note as incomplete. |
| With your own OpenAI key | 23 minutes with OpenAI's speaker-labelling model (the default). Other OpenAI models take longer notes; a longer note also offers Mullo's plan. |
Past the limit
Nothing is lost. The recording is saved on your phone (up to 200 MB, about 14 hours) and, when you're signed in, copied to your account with its audio. Mullo says it's too long to write down in one go; you can still play it back. For long meetings, stop and start a new recording every 3 hours.
Which AI does what
| Who pays | Transcripts | Titles, summaries, tasks, Ask, drafts |
|---|---|---|
| Mullo's plan (trial, Pro, Max) | Chosen by Mullo for each recording, by its language. Mullo's main speech model writes about 60 languages, several in one note too, and recordings up to about 3 hours. Other languages, a transcript in another language, and a recording the main model can't finish go to a second AI model. Every recording labels who said what, and every plan gets the same transcription. | Mullo's default, or the model you pick in Settings, Summaries. A trial picks from the trial models, Mullo Pro adds more, Mullo Max adds the most capable ones. |
| Your own Gemini key | Gemini 3.1 Flash-Lite, or any Gemini model you pick | Any model in the list, Max ones too |
| Your own OpenAI key | gpt-4o-transcribe-diarize (speakers labelled), or any OpenAI transcription model you pick | GPT-6 Luna, or any OpenAI model you pick |
- Mullo's server picks the model, not the app, so a better model reaches every phone without an update. This page says what's used now.
- The AI services Mullo uses work on paid terms that don't let them train their models on your recordings or notes (the kinds we use).
- On any recording, ⋯ then Summarise again can use another model for that note.
- Languages: about 110, about 60 of them on Mullo's main speech model (which ones, and when to pick one). Speakers are told apart within one recording; Mullo never recognises a voice from one recording to the next.