Skip to main content
Example pricing is shown below for reference; exact, up-to-date rates are always available from the API. deAPI uses a pay-as-you-go pricing model, where costs are calculated dynamically per task based on the resource usage, such as resolution, steps, duration, or number of output characters. The pricing examples shown on the public site (https://deapi.ai/#pricing) serve as reference points; final cost is determined by the API at runtime depending on the selected model and parameters.

​
Pricing Overview

Up-to-date prices are always available via the /price endpoint for the selected model (v2) or /price (v1).

​
Partner (external) model pricing

Some models in the catalogue run on our partners’ infrastructure rather than on deAPI’s own GPU network — see Partner (external) models. You can recognise them by the experimental tag and status: test_model in the model list. They are billed differently from native fleet models.
For partner models, /price returns an estimate, not a fixed rate. The response carries is_estimated: true, and the amount you are finally charged is settled after the job completes.
Native fleet models are priced from a formula over your request parameters, so the quote is exact. Partner models are priced by the partner, so deAPI quotes an estimate up front and reconciles it once the partner reports the actual cost. What this means in practice:
  • Read is_estimated on every /price response. When it is true, treat the number as indicative and budget some headroom.
  • The final amount lands on the job, not on the quote. Poll GET /api/v2/jobs/{request_id} — while the job runs, price is null; once it finishes, price.amount first appears with is_estimated: true, and switches to is_estimated: false once settled. Your balance is charged at that point.
  • Only the settled figure is authoritative. Do not reconcile your own accounting against the estimate or against the unsettled amount on a freshly finished job.
  • Estimate accuracy is best for the model’s default resolution and lower for unusual sizes or quality tiers.
Failed jobs are not charged: they report price.amount: 0 and your balance is untouched.

​
Pricing by Task

Image Generation
  • Users define width, height, steps via the API or UI.
  • Public example: Flux Schnell model is used to estimate cost in the UI; for example, 512x512 at 4 steps gives 0.00136 USD. For the Z-Image-Turbo INT8 model with the same parameters, the price is 0.0034 USD, but the advantage is very realistic images.
  • Higher resolutions and more steps yield better quality but incur higher cost.
  • Important: For models other than Flux Schnell, pricing is model-specific and calculated on the server side.
Text-to-Speech (TTS / Speech Generation)
  • Charged per character in your input (e.g. 1M characters → 0.77 USD).
  • Playback speed modifiers:
    • Standard (1.0×): base cost
    • Fast (2.0×): 0.5× the base cost
    • Slow (0.5×): 2.0× the base cost
  • Useful tip: using faster playback (2×) for drafts can reduce cost by ~50%.
  • All three TTS modes (custom_voice, voice_clone, voice_design) use the same per-character pricing model.
Text-to-Music (Music Generation)
  • Price depends on track duration and the number of inference steps. Duration is model-specific: 10–300 s on the turbo models, 30–300 s on AceStep_1_5_Base.
  • Turbo models use fewer steps (e.g. 8) and are cheaper; base models use more steps (e.g. 32+) for higher quality.
  • Use the /price endpoint with your model, duration, and inference_steps to get exact pricing before generation.
Text-to-Video
  • Price depends on clip length (frames / fps) and resolution. Length is model-specific: 1–4 s on Ltxv_13B_0_9_8_Distilled_FP8 (30–120 frames @ 30 fps), ~2–10 s on Ltx2_3_22B_Dist_INT8 and Ltx2_5_22B_Dist_INT8 (49–241 frames @ 24 fps).
  • Example public rate: 0.001737 USD for a 2-second clip at 256x256.
  • You can scale resolution or duration, but cost increases accordingly.
Image Edit (Image-to-Image)
  • Transforms an existing image based on a new prompt or style.
  • Pricing scales with the number of inference steps and GPU time required — similar to Text-to-Image tasks.
  • Example: a 20-step transformation costs roughly the same as generating a 512×512 image.
  • Fewer steps → faster & cheaper transfers; more steps → better fidelity.
Image-to-Video
  • Transform a static image into a motion clip with interpolation. Clip length follows the same per-model frame/fps limits as Text-to-Video above.
  • Example: 256x256 for 2 seconds costs 0.001737 USD.
  • Use lower resolution or shorter duration to reduce cost.
  • Use motion parameters in prompts to guide movement.
Video-to-Text (Transcription)
  • Billed per hour of video processed.
  • Sample public rate: 0.05 USD per hour.
  • For 5 minutes, cost is estimated at 0.0089 USD.
  • Supports timestamps, multilingual transcription, and batching for better throughput.
Transcription — WhisperLargeV3Ct2 (TikTok length bands) WhisperLargeV3Ct2 adds word-level timestamps, per-segment confidence scores and speaker diarization. It is also the only model that transcribes TikTok — a TikTok URL sent with any other model is rejected with 422.
The price list for TikTok transcription depends on the video length and is as follows:
  • The band is selected from the whole video length, rounded up: a 1:05 clip falls into the up to 2 min band and costs 0.0225 USD.
  • The over 10 min band is flat — a 12-minute and a 30-minute TikTok both cost 0.105 USD.
Every other source — YouTube, X, Twitch, Kick, uploaded files — is billed by duration and runs on either model. Segment timestamps are free. include_ts: true with the default ts_level: "segment" costs the same as a plain transcription. On WhisperLargeV3Ct2 only word-level timestamps (ts_level: "word") and speaker diarization (diarize: true) add +50% to the transcription price — diarization carries that surcharge at any granularity, segment included. WhisperLargeV3 does not charge for timestamps at all. Source metadata add-on (YouTube & TikTok)
  • Setting include_metadata: true adds a flat 0.005 USD per request on top of the transcription price.
  • The surcharge applies only to sources that can actually return metadata. An uploaded file has no source to describe, always returns null, and is never charged for the flag.
Image-to-Text (OCR / VLM)
  • Charged per image, based on input resolution. The number of characters returned does not affect the price — a blank image and a dense document page of the same size cost exactly the same.
  • Baseline rate: 0.00928 USD per image at 1024×1024.
  • Example rates (read back from POST /api/v2/images/ocr/price):
    • 256×256 → 0.00101 USD
    • 512×512 → 0.00266 USD
    • 1024×1024 → 0.00928 USD
    • 1920×1080 / 1080×1920 → 0.01791 USD
    • 4096×4096 → 0.14168 USD
  • Price scales with pixel count: ≈ 0.00045 USD + 0.0000000084 USD per pixel. Downscaling source images before OCR reduces cost proportionally — useful when running OCR across many video frames.
  • language and format do not affect the price.
  • Volume discounts available for bulk processing (100k+ images)—contact sales.
Text-to-Embedding (Vector Representations)
  • Pricing is based on the number of tokens processed.
  • Sample public rate: 0.000084 USD per 1,000 tokens (client-side pricing).
  • Embeddings are typically used for semantic search, retrieval-augmented generation (RAG), clustering, and similarity matching.
  • Costs scale linearly with token count, making it efficient even at large volumes.
  • Important: Different embedding models may have varying dimensionality (e.g. 768 vs. 1024), but pricing is standardized per token for simplicity.
Video Replace (Animate)
  • Replaces a character in a video using a reference image.
  • Price depends on video duration, resolution, and inference steps.
  • Use the /api/v2/videos/replacements/price endpoint with your model, duration (or upload video), width, and height to get exact pricing.
Video Upscale
  • Increases video resolution (e.g. 2×, 4×) using AI super-resolution.
  • Price scales with input resolution, video duration, and the selected scale factor (only for models with configurable scale — fixed-scale models use their built-in factor).
  • Use the /api/v2/videos/upscales/price endpoint with your model, width, height, duration, and optional scale to get exact pricing before uploading the file. Always send duration — the price is proportional to it, and a request without it returns price: 0.

​
Best Practices & Guidance

  • Use the public calculator as a guide, but always rely on the API’s final cost calculation (model + parameters).
  • Avoid hardcoding prices or cost assumptions—always fetch or compute cost based on actual model usage.
  • For reproducibility (e.g. in production or experiments), pin model versions and seeds so results are consistent across runs.
  • Prepare fallback options in your integration: if a model is deprecated or disabled, switch to a sensible alternative automatically.
  • Monitor usage and budget: higher resolution, longer clips, or more steps increase costs proportionally.
Link to live pricing page for reference: https://deapi.ai/pricing