Billing Examples

This page walks through real billing calculations for a few typical models, one per pricing style. The mechanics behind each style are explained in How Models Are Billed.

LLM tokens — zai-org/glm-5.2

LLM models are billed per million tokens, with separate rates for input (prompt), cached input, and output (completion) tokens (see LLM models):

$Cost = \frac{Fresh\ Input}{1M} \times \$1.40 + \frac{Cached\ Input}{1M} \times \$0.26 + \frac{Output}{1M} \times \$4.40$

You are charged after the request completes, based on the token counts reported in the response's usage field — streaming and non-streaming requests are billed identically, and failed requests are not billed.

Example — a chat request with 12,000 prompt tokens (10,000 of them served from the prompt cache) and 1,500 output tokens:

1. Fresh input: 2,000 × $1.40 ÷ 1M =$0.0028
2. Cached input: 10,000 × $0.26 ÷ 1M =$0.0026
3. Output: 1,500 × $4.40 ÷ 1M =$0.0066
4. Total: $0.0120

Without prompt caching, the same request would bill all 12,000 input tokens at the full input rate — $0.0168 +$0.0066 = $0.0234, nearly twice as much. Keeping a stable system prompt across requests lets the cache rate apply to the repeated portion.

Token-billed video — bytedance/seedance-2.0/reference-to-video

Seedance 2.0 reference-to-video is billed by output video tokens when the task completes (see token-billed video models). Two token rates apply, depending on your inputs:

Resolution multipliers: 480p / 720p / 1080p × 1.0, 4k × 0.57, 720p-sr / 1080p-sr × 1.8, 1440p-sr × 3.2. Billed tokens are measured from the generated output (and, for requests with reference videos, include the input-video contribution), so use /calculate for an exact quote on the 4K and SR tiers.

Tokens scale with resolution, frame rate, and duration:

$Video\ Tokens \approx \frac{Width \times Height \times FPS \times Duration(seconds)}{1024}$

At 1080p (1920 × 1080, 24 fps) this works out to 48,600 tokens per second of video.

Example A — reference images only, 10s at 1080p:

1. Output tokens: 48,600 × 10s = 486,000 tokens
2. List price: 486,000 ÷ 1,000,000 × $11.20 × 1.0 = **$5.44**

Example B — with an 8s reference video, 10s output at 1080p:

1. Billed tokens count the input video plus the output: (10s + 8s) × 48,600 = 874,800 tokens
2. The lower with-video rate applies: 874,800 ÷ 1,000,000 × $6.88 × 1.0 = **$6.02**

A small temporary hold is placed at submission and released once the task is billed; failed tasks are not charged.

Per-second video — alibaba/wan-2.7/image-to-video

Wan 2.7 image-to-video uses classic per-second pricing: a base rate of $0.10 per second, scaled by the output resolution, with a 5-second minimum:

$Cost = \$0.10 \times max(5,\ Duration) \times Resolution\ Multiplier$

Example — 8s at 1080p:

$0.10 × 8 × 1.5 = **$1.20**, reserved from your balance at submission and captured when the video is delivered.

Example — 3s at 720p:

Duration is below the minimum, so 5 seconds are billed: $0.10 × 5 × 1.0 = **$0.50**.

Per-image with options — google/nano-banana-pro/edit

Nano Banana Pro (edit variant) is priced per output image, with a resolution tier and an optional web-search surcharge:

$Cost = Images \times Tier\ Price\ (+\ \$0.014\ if\ web\ search)$

Example — 2 images at 4K with web search enabled:

1. Images: 2 × $0.24 =$0.48
2. Web search: + $0.014 (added once per request)
3. List total: $0.494

The amount is reserved at submission and released automatically if the task fails.

Token-billed image — openai/gpt-image-2/edit

GPT Image 2 is billed like an LLM — by tokens — but the token counts are computed from your request parameters at submission:

Output token counts for a 1024×1024 image: 196 (low), 1,756 (medium), 7,024 (high). Larger sizes produce proportionally more tokens; input images always use the medium token base regardless of quality.

Example — edit one 1024×1024 image, quality: medium, size: 1024x1024, n: 1:

1. Input text: 1,000 tokens × $5/1M =$0.0050
2. Input image: 1,756 tokens × $8/1M =$0.0140
3. Output image: 1,756 tokens × $30/1M =$0.0527
4. Total: ≈ $0.072 per image

Requesting n images multiplies the whole amount by n.

Character-billed audio (TTS) — xai/tts-v1 and bytedance/seed-audio-1.0

Both text-to-speech models are priced by the length of the input text, at $0.015 per 1,000 characters:

$Cost = \frac{Characters(text)}{1{,}000} \times \$0.015$

Characters are counted as Unicode characters, so a CJK character counts as one character, the same as a Latin letter.

Example — a 1,200-character script:

1,200 ÷ 1,000 × $0.015 = **$0.018**, charged at submission.

Example — a short 300-character prompt:

300 ÷ 1,000 × $0.015 = **$0.0045**.

Check any model yourself

Every example above can be reproduced with the cost estimation API — send your exact request body to /api/v1/model/calculate and it returns the list price, your effective price, and the applied discount without creating a task.