Descript Doesn't Have Its Own AI. Here's Proof.
Every generative AI feature in Descript — voiceover, video, images, even the in-app editing assistant — turns out to be a labeled, swappable third-party model. It's not hidden. It's sitting in a Settings page most users never open, right next to the exact credit cost of every locked, paid alternative.
What I Was Actually Testing
Descript markets itself as an all-in-one AI video and podcast editor: AI speech, AI-generated video and images, and "Underlord," an in-app AI co-editor. The obvious assumption is that Descript built these models the way it built its actual original tools — transcription, Studio Sound, filler-word removal, Eye Contact.
So I went looking for exactly which model does what. The answer wasn't hidden — it was one settings page away the whole time.
Settings → AI Models Names Every Vendor
Under Settings, Account, AI models, Descript lists the exact model behind every generative feature, split into two sections: Core and Media generation.
- AI speech — exactly two options, both real ElevenLabs models: ElevenLabs V3 (the default) and ElevenLabs Multilingual v2. No in-house Descript text-to-speech exists at all.
- Video — PixVerse v5 free by default, with Gemini Omni Flash, Kling O3 Pro, and Kling v3.0 shown locked behind paid tiers.
- Image — Nano Banana free by default, with Flux Kontext also available and GPT Image 2 plus Nano Banana 2 shown locked.
- Avatar — Kling Avatar v2, confirmed by actually running a real avatar generation to completion.
- Underlord, the AI co-editor itself, is a model picker too — Claude Haiku 4.5, Claude Sonnet 5, and Claude Opus 5 are available on Free, with Claude Fable 5 and GPT 5.5 shown locked, and Gemini 3.5 Flash also available.
The Locked Models Come With Real, Visible Prices
Opening the Video model dropdown doesn't just show which model is selected — it shows the exact credit cost of every locked alternative: Gemini Omni Flash at 15 credits, Kling v3.0 at 22 credits, and Kling O3 Pro at 29 credits. Free plan users can see precisely what a premium generation would have cost, model by model, before ever subscribing to unlock it.
This is worth saying plainly: Descript isn't hiding this. It's not a leak or a bug. It's a settings page most users have no reason to ever open, listing exactly which outside company is doing the actual generation, and exactly what Descript charges in credits to let you use it.
This Isn't True Of Descript's Older, Original Tools
It would be inaccurate to say Descript has never built anything. Transcription, Studio Sound, filler-word removal, and Eye Contact are Descript's own long-standing features and were not tested or disproven here. The finding is specifically about the newer generative layer — voice, video, image, and chat-assistant generation — which is where every single model turned out to be an outside vendor's, selected from a dropdown.
The One Thing You Can Actually Do
Open Settings, Account, AI models yourself. It takes about ten seconds, and it tells you exactly which company is generating your voiceover, your video, and your images the next time Descript's interface says "AI-powered."
Tightening your own script before you film?
Paste it into Viral Analyzer Pro, see your hook score, get a rewrite suggestion — no credit card required.
Get Started Free →Free plan available · Setup in 2 minutes
Watch more real tool tests on the channel.
This post reflects hands-on testing on free-tier accounts as of August 2026. Descript's model lineup and pricing may change after publication.
Last updated: August 29, 2026.