TARKOVSKY FIELD GUIDEOpen studio →

FIELD GUIDE · MODELS

Models:
use the ones you choose.

Tarkovsky is not one model. Writing, images, video and voices each use the model you choose, on Tarkovsky credits or your own key, and every project can change them. Each model's line says what it takes and what it costs a second; the makers that may train on what you send are marked before you choose one.

10 / MODELS

Use the models you choose.

Defaults for new projects sets the Writing, Images, Video, and Voices models that a new project starts with. Each project can change them.

  1. Create an API key with the provider. A consumer chat subscription does not necessarily include API access.
  2. Open Models and enter the key in the matching provider card.
  3. Save the key. Each key is checked with one provider call when it is saved; a rejected key is not stored. MiniMax keys cannot be checked.
  4. In the Story chat, select My API key, choose the connected provider, and select your model from the list.

OpenAI, Anthropic Claude, Google Gemini, OpenRouter, and MiniMax have separate key cards. The xAI card shows Connect your SuperGrok account · Coming soon. OpenRouter is a gateway, and Meta is a model developer. Meta Llama models are accessed through OpenRouter with an OpenRouter key. A key is sent only to its matching provider. Arbitrary API base URLs are not supported.

Keys are encrypted with AES-256-GCM and belong only to the account that saved them; collaborators cannot see or use them. A connected card shows the last four characters with Replace and Remove. Removing a key puts that provider back on Tarkovsky credits. It can also prevent an active job from being polled until the key is restored.

Provider references: OpenAI Responses, Gemini compatibility, OpenRouter, MiniMax H3.

Models: when one model or voice can't be reached (0.68.6)

  • When the studio's own account at one service is out, only the models on that service stop; every other model still works, and nothing is charged. Its credits are given back in full.
  • A take on Seedance 2.5, MiniMax H3 (MiniMax direct) or Grok (xAI direct) says: This model can't be reached from the studio just now. Nothing was charged. Try another model for now.
  • A voice line on an ElevenLabs voice says: This voice can't be reached from the studio just now. Nothing was charged. Try a Kokoro voice for now.
  • Music or a sound on credits says: Music and sound can't be reached from the studio just now. Nothing was charged. Try again later, or add your own ElevenLabs key in Models and choose Your key.
  • An upscale at Lock picture says: Upscaling can't be reached from the studio just now. Nothing was charged. Try again later.
  • With your own ElevenLabs key ("Your key"), you see ElevenLabs' own words, as before. When every model is out at once, the line is the one above: The studio can't reach its AI models right now.

Voices: Kokoro or ElevenLabs

  • Kokoro stays the budget option. Seven fixed voices on Tarkovsky credits, quoted from the OpenRouter tariff before it runs and charged at the quoted service price.
  • ElevenLabs is for voice-over and voice casting. Choose it in Models → Voices as "ElevenLabs · credits" or "ElevenLabs · your key". With your own key (Models → Your keys → ElevenLabs; the key is checked by listing your voice library when you save it, nothing is synthesized) ElevenLabs bills you and Tarkovsky charges 0 credits.
  • On credits the quote reserves a ceiling of $0.30 per 1,000 characters (priced ×1.5 into credits, so about 45 credits per 1,000 characters). After the take is saved the charge settles down to the characters ElevenLabs reports in its character-cost response header (or the request's own count if the header is missing) and the rest of the reservation comes back at once. A provider error releases the whole reservation; a storage failure after the provider did its work holds the reservation for review rather than refunding blindly.
  • The casting list (GET /api/voices?provider=elevenlabs) shows name, category, labels and a preview link for every voice on the key; provider=kokoro lists the seven Kokoro voices. Synthesis goes to POST /api/projects/:id/voice with provider, text, voice (the ElevenLabs voice id or a Kokoro voice), payment and the maxCredits from POST /api/projects/:id/voice/quote. Editors and owners can synthesize; viewers can quote only.
  • Model: eleven_multilingual_v2, output mp3_44100_128. The MP3 is saved to the project like any other take ("AI voice · ElevenLabs · {voice}.mp3").

Models: makers that may train on what you send (0.22)

  • By each maker's own API terms, three may use what is sent to improve their models: MiniMax (Hailuo), Kuaishou (Kling) and Black Forest Labs (FLUX, FLUX 3 Video). Google, OpenAI, Anthropic, xAI, Alibaba and ByteDance say they don't by default.
  • Every model menu marks the three: "This model's maker may use what you send to improve its models", beside the model's name, before you choose it. Tarkovsky Director never picks one for you and never offers one as a switch; it is used only when you pick it.
  • Everything else the studio sends through OpenRouter asks for providers that don't collect data: text and speech carry OpenRouter's no-data-collection setting; stills leave out the providers of the makers that may train (the images endpoint has no such setting); a video take goes to the model's own maker (OpenRouter's video endpoint has no routing setting).
  • A model you pick from the three is sent without that setting, so it can run.
  • Admin → Video models shows the flag per model in "Maker may train on what we send", with the source.

Models: the default video model and what each model takes (0.42)

  • The default video model is Grok Imagine Video 1.5 · xAI direct, until you choose another in Models → Defaults for new projects → Video. It is the highest-ranked model that takes the most reference images, for its price, among the models whose makers don't train on what you send and that take images of real people: 7 reference images beside the first and last frames. Where the xAI route has no key, the default is Veo 3.1 Fast. A project with no video model of its own, the kickoff's video model question and every take panel start on it.
  • On a Free account (0.51: Grok Imagine Video 1.5 Lite first again, Veo 3.1 Lite second; it was Veo 3.1 Lite first in 0.50 and Grok Lite first in 0.46–0.49): Models → Defaults for new projects → Video shows Grok Imagine Video 1.5 Lite until you choose another (Veo 3.1 Lite when the catalog doesn't list it), and its Recommended list leads with the two, in that order. A new project starts on it, unless you chose a default yourself, and the kickoff's first answer still becomes the film's model (a project made on any earlier Free default too); projects keep the model they have. Every model stays listed and can be chosen. Why: Zach's rule of 28 October, "If neither burns in a caption, Grok Lite goes back first". On 29 October two Grok Lite takes with a spoken line, both opening on the first frame sent, burned in no caption (our 26 October take had shown the line as a caption, but its first frame never reached the model). Grok Lite carries lines: a part with lines runs on it.
  • Each model's line is the same in the Models page, the kickoff's menu and the take panels: sound ("sound", "sound, lines unconfirmed", "sound not known yet" or "silent"), reference images, the last frame, then the price a second with sound and silent where they differ, at the resolution the rate is for: Veo 3.1 Fast · sound · 3 references without a frame (8 s) · last frame · 15 credits/s with sound, 12 silent (720p).
  • The menus checked again (0.50, Zach's model menu check of 28 October, every line against each maker's own pages and OpenRouter): Grok Imagine Video 1.5 Lite says sound and lines, and HeyGen Video 1 sound and lines (always on) (HeyGen's lines are counted on now: every clip has dialogue, ambience and effects, with no off switch). Where MiniMax charges for reference pictures, the line says what each adds: H3 · … · 19.5 credits/s with sound (2K) · + 6 credits a reference; MiniMax H3 · MiniMax direct · … · 12 credits/s with sound (768p) · + 6 credits a reference after the first 5; MiniMax H3 Max (fast) · MiniMax direct · … · 12 credits/s with sound (768p) · + 11.1 credits a reference after the first 2 (nine references add about 78 credits to a take). Wan 3.0 now says 12.75 credits/s at 720p: on 29 October OpenRouter charged a 6 s take $0.51, its 15% discount (0.51; the hold stays at the list price). Wan 2.6 takes no references through OpenRouter: its catalog lists none, so the studio sends none (0.51). Admin › Video models and the facts file carry the rest (xAI's prices by resolution, the Veo sources, FLUX 3 Video's and HeyGen's price notes, Seedance 2.5 "speaks dialogue", Wan 2.6 and 2.7's sound).
  • "N references without a frame": through OpenRouter a take that starts on a first or last frame carries no reference images (the frame wins), so a model's references go only on a take with no frame. Under the menu, the Models page says what our route sends: Google: Veo 3.1 Fast sends up to 3 references on a take with no first or last frame, 8 s long; a take that starts on a frame carries none.
  • Only what our route sends (0.46, Zach: "list only what our site can send"): every menu (the Models page, the kickoff, the take pickers, Admin's carries line) says the references our route sends, and "no references" where it sends none: SpaceXAI: Grok Imagine Video 1.5 · sound · no references · no last frame · 21 credits/s with sound (720p). Where a route of ours sends them for the same model, the note under the Models menu names it, in Zach's words: Grok Imagine Video 1.5 · xAI direct sends up to 7 references beside its first and last frames. (the number is the route's own). What a maker takes beyond our route isn't shown to customers: administrators see it in Admin → Video models → Reference gaps, and every model and route is checked in docs/evidence/0.46/reference-gaps.md, so Zach and Costa can find out why and fix it. (0.44, TAR-437, said the maker's count instead: "takes 14 references, not sent on this route yet".) What a take sends is unchanged.
  • "Sound, lines unconfirmed": the model makes sound by its maker's word or our own takes, but a part with lines still asks for another model before it runs. Kling O1 makes no sound (Kling), whatever OpenRouter's catalog says. "Sound and lines" (0.50) means what "sound" means: the model speaks lines, and a part with lines runs on it.
  • Every fact, its page, a short quote and the date it was checked are in docs/models/model-facts-0.42.md (checked 25 October 2026).

QUESTIONS · MODELS

Questions, answered from the guide.

Do the model makers train on what I send?

By each maker's own API terms, three may use what is sent to improve their models: MiniMax (Hailuo), Kuaishou (Kling) and Black Forest Labs (FLUX, FLUX 3 Video). Google, OpenAI, Anthropic, xAI, Alibaba and ByteDance say they don't by default.

Which video model is the default?

The default video model is Grok Imagine Video 1.5 · xAI direct, until you choose another in Models → Defaults for new projects → Video. It is the highest-ranked model that takes the most reference images, for its price, among the models whose makers don't train on what you send and that take images of real people: 7 reference images beside the first and last frames. Where the xAI route has no key, the default is Veo 3.1 Fast. A project with no video model of its own, the kickoff's video model question and every take panel start on it.

What does each model's line in the menu tell me?

Each model's line is the same in the Models page, the kickoff's menu and the take panels: sound ("sound", "sound, lines unconfirmed", "sound not known yet" or "silent"), reference images, the last frame, then the price a second with sound and silent where they differ, at the resolution the rate is for: Veo 3.1 Fast · sound · 3 references without a frame (8 s) · last frame · 15 credits/s with sound, 12 silent (720p).

Can I use my own ElevenLabs key for voices?

ElevenLabs is for voice-over and voice casting. Choose it in Models → Voices as "ElevenLabs · credits" or "ElevenLabs · your key". With your own key (Models → Your keys → ElevenLabs; the key is checked by listing your voice library when you save it, nothing is synthesized) ElevenLabs bills you and Tarkovsky charges 0 credits.