Infinite Studio

What a creation actually costs

You pay a subscription, not tokens. Each week opens an envelope, and every creation takes a share of it, by what it asks of our machines. The tables below give what it costs you, in cents, plan by plan. Every figure is computed by the code that really charges the week, not by a brochure.

What each one can make

Price is not the whole story: the assistant still has to be able to make the thing. Observed in September 2026 on the consumer apps.

The price shown is the plan that carries what is ticked. Cheaper doors exist (ChatGPT Go $8, Google AI Plus $4.99, SuperGrok Lite $10) but they limit or exclude video and pictures. Here, all six rows are open on the first plan, and the first week is free.

Video, sound included

Sound is included (voice, ambience, music). Turbo+ at 768p, one notch below in quality: 6.3¢ for 5 s. Enlarging a video to 1080p: about 4.3¢, whatever the length.

What a second of video costs, everywhere

Everyone in one table, cheapest first. Ours are the Pro plan's prices, as each of them shows its cheapest tier here; count double on the entry plan.

At a comparable definition — our 768p (1344 x 768) sits a notch above a 720p — a second comes to 1.9¢ on the entry plan and 0.9¢ on Pro. The same second costs 10¢ at Alibaba (Wan 2.6 at 720p), 23¢ at ByteDance (Seedance 2.5 at 720p) and 40¢ at Google (Veo 3.1 at full quality): 5 to 21 times our price. Our 480p comes to 0.8¢ a second and Turbo to 0.5¢, at a lower definition.

Touching up a picture

What a touch-up costs, everywhere

Everyone in one table, cheapest first. Ours are the Pro plan's prices, as each of them shows its cheapest tier here; count double on the entry plan.

A quick touch-up comes to 1.6¢ here, 0.8¢ on Pro. Elsewhere, a picture produced runs from 2¢ to 13¢.

The quick touch-up, the everyday one, runs on our own cards with Qwen-Image-Edit; the careful one goes to OpenAI. Cutting an object out costs nothing: only the edit that follows takes from your week. Nobody else bills an edit separately — elsewhere it costs the price of a picture produced, and that is the rate shown above.

Songs

What a song costs, everywhere

Everyone in one table, cheapest first. Ours are the Pro plan's prices, as each of them shows its cheapest tier here; count double on the entry plan.

A three-minute song at the careful quality comes to 5.9¢ here, 2.9¢ on Pro. ElevenLabs, the only public music API, charges 45¢ for the same length — 15 times more. Suno publishes no per-song price: its subscription starts at $8 a month, on top of yours, for music alone, and its credits do not roll over from one month to the next.

Voices

What a minute of speech costs, everywhere

Everyone in one table, cheapest first. Ours are the Pro plan's prices, as each of them shows its cheapest tier here; count double on the entry plan.

A minute of text said by a kept voice comes to 2.9¢ here, 1.4¢ on Pro, against 4.5¢ at ElevenLabs on their fastest model. A minute of speech counts here as nine hundred characters.

Slides, sites and quizzes

The quick quality is written by our own model, on our machines; the careful one by a large cloud model. None of the four assistants bills a presentation by the unit: they sell tokens, and a presentation eats as many as the model writes.

What one week holds

A table of cents is exact and says little. Here is the same thing the other way round: how many creations fit in one week, if you did nothing else.

A week starts at your first creation and lasts seven days; the next one starts at the creation that follows. Nothing is lost while you do not create.

Where the difference comes from

Three reasons, and none of them is a trick.

Models other people opened

Most of what the studio does runs on models their teams made public: ACE-Step for music, VoxCPM2 and CosyVoice 3 for voices, Qwen-Image-Edit for touch-ups, SAM 3 to cut objects out, the Nomos models to enlarge, Wan for part of the video. The main video model is MiniMax H3, which we run on our own machines under MiniMax's written authorization. Those teams — MiniMax, ACE Studio and StepFun, OpenBMB, FunAudioLLM and Alibaba, Meta, the authors of the upscalers — did the expensive part, the training, and then shared it. We pay no per-second licence for that. A good share of what you save comes from their choice rather than ours, and it is only fair to say so.

Our own cards

Those models run here, in Europe, on machines we operate. We pay for the card hour and the electricity — not a model vendor who has a margin of their own to make. That is what explains the order of magnitude, not a launch discount.

The cloud, only where it earns its place

The direction (rewriting your request, painting a first frame), the careful touch-up, transcription and the jury of critics go through OpenAI or Alibaba Model Studio. They are the only steps that leave our machines, they are included in your week, and your account's privacy setting can switch them off: at the tightest level, nothing leaves at all.

What is not in the price

No credits to buy in advance and spend before they expire. No surcharge for sound: a video arrives with its voices, its ambience and its music. No paid queue: the Pro plan creates in priority at the normal lane's price. And the share of the week you do not use is not billed twice: our rates already account for nobody using everything.

Method

Our figures are produced by scripts/price_page.py, which calls the same functions the API calls when it charges your week: one estimate per trade, then the common formula. They are rounded and give an order of magnitude rather than a price to the decimal: what a creation really takes depends on what it asks of our machines and is measured afterwards. The others' figures are their published prices per unit produced, in dollars.

Sources: Alibaba Model Studio · BytePlus ModelArk · Google · xAI · ElevenLabs · Suno · OpenAI — read on 22 September 2026.

The plans · The studio inside your assistant