Realtime Compute for Apache Flink:AI service billing (pay-as-you-go)

Billed by token

For models billed by token, the fee of each request is metered by the following types:

• Standard Input Token: Billed at the base input unit price.

• Standard Output Token: Billed at the base output unit price.

• Implicit Cache Hit Input Token: Except for models that are otherwise specified, billed at 20% of the base input price.

• Explicit Cache Creation Input Token: Except for models that are otherwise specified, billed at 125% of the base input price.

• Explicit Cache Hit Input Token: Except for models that are otherwise specified, billed at 10% of the base input price.

Per-minute request cost = Σ(token usage per type × unit price). The unit price of cache-type tokens = base input unit price × billing coefficient.

Billed by duration

For models billed by duration, the fee of each request is metered by the following types:

• Input duration: billed at the input unit price.

• Output duration: billed at the output unit price.

Per-minute request cost = Σ(billed duration per type × unit price). The billed duration varies by model: some models are billed by input duration only, some by output duration only, and some by both. For details, see the pricing tables.

Billed by number of images

For models billed by number of images, the fee of each request is metered by the following types:

• Number of input images: billed at the input unit price.

• Number of output images: billed at the output unit price.

Per-minute request cost = Σ(number of images per type × unit price). The billed number of images varies by model: some models are billed by input images only, some by output images only, and some by both. For details, see the pricing tables.

Base Pricing

The following tables show the base unit prices of standard input/output tokens, duration, and number of images. Cache-type tokens are billed at the base prices multiplied by the corresponding coefficients. For details, see Cache Token Billing Coefficients.

• Chinese mainland: such as China (Hangzhou) and China (Beijing).

• Outside the Chinese mainland: includes all regions outside the Chinese mainland, such as China (Hong Kong), Singapore, and US.

For the region list, see Supported regions.

Qwen

DeepSeek

GLM

Embedding

Rerank

Qwen-Audio-TTS

Speech synthesis models are billed by input token usage. 1 character = 1.5 tokens.

CosyVoice

Speech synthesis models are billed by input token usage. 1 character = 1.5 tokens.

Qwen-Audio-ASR

Speech recognition models are billed by the duration of the input audio.

Fun-ASR

Speech recognition models are billed by the duration of the input audio.

Qwen-Image

Image generation models are billed by the number of input and output images.

HappyHorse

Video generation models are billed by output duration.

Cache Token Billing Coefficients

Cost of cache tokens = base input unit price of the corresponding model × billing coefficient × token usage. For more information, see Context Cache.

Except for models that are otherwise specified, the cache token billing coefficients are as follows:

Scenario: Using qwen3.6-plus in the China North 2 (Beijing) region (input range 0~256K tokens), with cumulative consumption of 5,000K input tokens and 1,200K output tokens.

Fee Calculation:

• Input fee: 5,000 × 0.00033 = 1.65 USD

• Output fee: 1,200 × 0.00198 = 2.376 USD

• Total: 4.026 USD

Scenario (billed by number of images): Using qwen-image-3.0 in the China North 2 (Beijing) region (output specification 1k), with 20 reference images uploaded and 100 images generated.

Fee Calculation:

• Input image fee: 20 × 0.003301 = 0.06602 USD

• Output image fee: 100 × 0.029705 = 2.9705 USD

• Total: 3.03652 USD

Scenario (billed by duration - video generation): Using happyhorse-1.1-t2v in the China North 2 (Beijing) region (output specification 720P) to generate ten 5-second videos, with a total output duration of 50 seconds.

Fee Calculation:

• Video generation fee: 50 × 0.148523 = 7.42615 USD

• Total: 7.42615 USD

Scenario (billed by duration - speech recognition): Using qwen-audio-3.0-asr-flash in the China North 2 (Beijing) region to recognize audio, with a total recognition duration of 3,600 seconds.

Fee Calculation:

• Speech recognition fee: 3,600 × 0.000036 = 0.1296 USD

• Total: 0.1296 USD

Download Bills

1. Log on to the Bill Details page and select the billing period.

2. Set Product to Realtime Compute for Apache Flink, and set Product Name to Realtime Compute for Apache Flink_AI_PAYG_International, and click Search.

3. Click the export icon in the upper-right corner of the bill list to download the bill.

4. Open the bill file and locate the Instance ID (billing granularity) column. This field is separated by semicolons (;) in the following format:

   1. Models billed by token:

      Region;Workspace ID;Project Name;Flink Usage;Model;Token Type;Input Range

      Segment	Description
      Region	Region of the workspace
      Workspace ID	Unique identifier of the Flink workspace
      Project Name	Currently empty; reserved field
      Flink Usage	Model usage scenario; valid values: ai_function or flink_agents
      Model	Name of the invoked LLM
      Token Type	Input Token or Output Token
      Input Range	Upper bound marker of the input token range, such as 128, 256, 1000, etc.

   2. Models billed by duration or number of images:

      Region;Workspace ID;Project Name;Flink Usage;Model;Output Specification;Usage Type

      Segment	Description
      Region	Region of the workspace
      Workspace ID	Unique identifier of the Flink workspace
      Project Name	Currently empty; reserved field
      Flink Usage	Model usage scenario; valid values: ai_function or flink_agents
      Model	Name of the invoked LLM
      Output Specification	For models billed by number of images: the image resolution, such as 1k, 2k, or 4k. For models billed by duration: the video resolution for video generation models, such as 480P, 720P, or 1080P. Speech recognition (ASR) models do not distinguish output specifications and are marked as all.
      Usage Type	For models billed by number of images: input images and output images. For models billed by duration: input duration and output duration.