:Billing for models

Model calling fees are automatically deducted. Bills are generated hourly. For more information, see Introduction to payment methods.

Subscription method:

Model inference (call): Click here to purchase LLM inference savings plan.

After March 15, 2024, Model Studio upgraded its commercial services. All subscription services were changed to pay-as-you-go services. Therefore, you do not need to manually renew your services. The pay-as-you-go billing method is used automatically.

• Model inference and model training

  You will no longer incur fees after you stop using the related features. For model inference, you can delete the API key (Singapore or Beijing) to prevent further charges from accidental calls.

  

You can set a monthly spending alert. Set the alert threshold to a low value. Alibaba Cloud will notify you when unexpected charges occur to help you avoid further losses.

You can view the number of calls and token consumption for a specific model on the Model Observation (Singapore or Beijing) page.

Tokens are the basic units that a model uses to represent text. You can think of them as characters or words.

• In Chinese, one token is usually one character or word. For example, the text "你好，我是通义千问" (Hello, I am Qwen) is converted to ['你好', '，', '我是', '通', '义', '千', '问'].

• For English text, one token usually corresponds to three to four letters or one word. For example, "Nice to meet you." is converted to ['Nice', ' to', ' meet', ' you', '.'].

Different LLMs may use different methods to chunk tokens. You can use a SDK to view the token data chunked by a Qwen model on your local machine.

View the token data chunked by a Qwen model:

# Make sure that the DashScope Python SDK is installed.
from dashscope import get_tokenizer

# Get the tokenizer object. Currently, only Qwen series models are supported.
tokenizer = get_tokenizer('qwen-turbo')

input_str = 'Qwen has powerful capabilities.'

# Chunk the string into tokens and convert them to token IDs.
tokens = tokenizer.encode(input_str)
print(f"The token IDs after chunking are: {tokens}.")
print(f"There are {len(tokens)} tokens after chunking.")

# Convert the token IDs to strings and print them.
for i in range(len(tokens)):
    print(f"The string corresponding to token ID {tokens[i]} is: {tokenizer.decode(tokens[i])}")

// Copyright (c) Alibaba, Inc. and its affiliates.
// dashscope SDK version >= 2.13.0
import java.util.List;
import com.alibaba.dashscope.exception.NoSpecialTokenExists;
import com.alibaba.dashscope.exception.UnSupportedSpecialTokenMode;
import com.alibaba.dashscope.tokenizers.Tokenizer;
import com.alibaba.dashscope.tokenizers.TokenizerFactory;

public class Main {
  public static void testEncodeOrdinary(){
    Tokenizer tokenizer = TokenizerFactory.qwen();
    String prompt ="If you had to walk a very long distance, how long would it take to arrive? ";
    // encode string with no special tokens
    List<Integer> ids = tokenizer.encodeOrdinary(prompt);
    System.out.println(ids);
    String decodedString = tokenizer.decode(ids);
    assert decodedString == prompt;
  }

  public static void testEncode() throws NoSpecialTokenExists, UnSupportedSpecialTokenMode{
    Tokenizer tokenizer = TokenizerFactory.qwen();
    String prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nSan Francisco is a<|im_end|>\n<|im_start|>assistant\n";
    // encode string with special tokens <|im_start|> and <|im_end|>
    List<Integer> ids = tokenizer.encode(prompt, "all");
    // 24 tokens [151644, 8948, 198, 7771, 525, 264, 10950, 17847, 13, 151645, 198, 151644, 872, 198, 23729, 80328, 9464, 374, 264, 151645, 198, 151644, 77091, 198]
    String decodedString = tokenizer.decode(ids);
    System.out.println(ids);
    assert decodedString == prompt;

  }

  public static void main(String[] args) {
      try {
        testEncodeOrdinary();
        testEncode();
      } catch (NoSpecialTokenExists | UnSupportedSpecialTokenMode e) {
        e.printStackTrace();
      }
  }
}

The local tokenizer helps estimate the number of tokens in your text. However, the result is for reference only and may not match the server-side count exactly. For more information about the Qwen tokenizer, see the tokenizer reference.

Refer to the Error messages document for the corresponding solution.

The free quota data is updated hourly. During peak hours, there may be a delay of up to one hour. Therefore, you need to view the remaining quota one hour after the model call is complete.

In a multi-turn conversation, the input and output from historical conversations are billed as input tokens for the new turn.

You are not charged for creating an application. However, if you call the application for a Q&A pair, you are charged a model calling fee based on the model that is called.

If the free quota is not used up, no bill is generated and no fee is incurred. In this case, the savings plan is not used for deduction. The savings plan is used for deduction after the free quota is used up and a bill is generated.

If your account has an overdue payment, you cannot make model calls even if you have a free quota (Singapore only) or resource plans. You can go to the Recharge page to top up your account.

1. Service not activated

Use your Alibaba Cloud account to go to the Model Studio console (Singapore or Beijing) and activate the model service of Model Studio.

2. Insufficient account balance

• Check balance: Log on to the Expenses and Costs page to check whether your balance is sufficient.

• Top up: Click Top-up & Remittance, enter the required amount, and complete the payment.

3. Set a spending alert to prevent repeated errors

• Configure alert: For more information, see How to set an alert for high spending? to set a spending threshold. The system sends a reminder when the amount reaches the threshold.

Possible reasons are:

• The billing system is updated on an hourly basis. During peak hours, there may be a delay of up to one hour. For example, charges that are incurred between 16:00 and 17:00 may not be billed until 19:30:00.

• Free models and model inference within the free quota (Singapore only) do not generate bills. Only usage that exceeds the free quota generates bills.

On the Cost Analysis page, set Cost Type to Pretax Amount, set Time Unit to Month, select a time range, and set Product Name to Alibaba Cloud Model Studio. You can then view the costs of Model Studio within the selected time range.

On the Cost Analysis page, set Cost Type to Pretax Amount, set Time Granularity to Month, select a time range, and set Product Detail to Model Studio Foundation Model Inference. You can then view the total cost of model inference within the selected time range.

Take qwen-max as an example. On the Bill Details page, select a Billing Month. Set Commodity Name to Model Studio Foundation Model Inference and click Search.

In the Instance ID column, find all instances related to qwen-max. Sum the pretax amounts for these instances to get the total inference fee for the qwen-max model in the selected billing cycle.

Bills for model inference, deployment, and training that are generated after September 7, 2024 can be reconciled based on the ApiKeyID, workspace ID, model name, input/output type, calling channel, and tags of instances.

On the Bill Details page, select a Billing Month. Set Commodity Name to Model Studio Foundation Model Inference and click Search. Download the search results to your local machine and reconcile the bills based on the content in the Instance ID column.

A complete Asset/Resource Instance ID, such as 12xxx;llm-xxx;qwen-max;output_token;app, represents ApiKeyID;Workspace ID;Model Name;Input/Output Type;Calling Channel respectively. If your Asset/Resource Instance ID does not contain an ApiKeyID, the charge item was generated by a call from the console.

A complete Instance ID, such as text_token;llm-xxx;qwen-max;output_token;app, represents Billing Type;Workspace ID;Model Name;Input/Output Type;Calling Channel respectively.

A complete instance tag, such as key:test value:test, represents Tag Key (key) and Tag Value (value) respectively. If an instance has two or more tags, the tag key-value pairs are listed sequentially and separated by semicolons, such as key:test1 value:test1; key:test2 value:test2.

Go to the Model Studio API Key Management page and confirm the API key that corresponds to the ApiKeyID to reconcile bills based on the API key.

Go to the Workspace Management (Singapore or Beijing) page and confirm the workspace that corresponds to the workspace ID to reconcile bills based on the workspace.

Calling channels include app, bmp, and assistant-api. app indicates that the model is called through an application. bmp indicates that the model is called through the Playground (Singapore or Beijing). assistant-api indicates that the model is called through the assistant API.

Pay-as-you-go cloud resource bills are Not settled in real time. Instead, the system first freezes the amount that is consumed but not yet settled from the account's available credit. At the beginning of the next month, after the final monthly bill is issued, the bill for the previous month is actually deducted.

You can set a monthly spending alert in the Expenses and Costs center.

• Stop charges after your free quota is used up

  To avoid extra costs, Model Studio provides a Free quota only feature.

• Limit the number of model calls or tokens consumed per unit of time

  Set rate limits for a sub-workspace. Go to the Workspaces page, find the target sub-workspace, and click Authorization & Throttling Settings. Adjust the Request Number Limit and Token Limit for each model.

• Set an alert for token consumption

  Set an alert rule for model overhead. For more information, see Usage and performance observation.

  • If the Advanced monitoring service has not been activated, the Alibaba Cloud account must first switch to the target workspace and then manually enable or disable it on the Model Observation page. To use a RAM user, the Alibaba Cloud account must first grant the necessary permissions to the RAM user.

  • Go to the Model Alerting page and follow the instructions to activate the CloudMonitor service.

  • Click Create Alert Rule to configure the rule. When a specified metric becomes abnormal, the system notifies you or your O&M team.

  Model alerts only trigger notifications and do not stop model calls.