Billing and concurrency

How do key billing modes differ?

Billing basis: Text models are billed separately for input and output tokens. Per-request billing for media models depends on the available models in the Model Catalog (Models in the console).

Does the API have concurrency limits?

Yes. Rate limits are set per model and route. Models have independent concurrency pools, so reaching one model's limit does not affect others. See each route's information in the Model Catalog (Models in the top navigation) for RPM/TPM limits.

Recommendations by model type

• Text models: Suitable for frequent conversations, batch translation, data processing and other high-concurrency uses.
• Image models: Generation takes longer and involves more data. Use an asynchronous task queue to smooth peaks.
• Video and large-file tasks: Processing takes longer. Use asynchronous processing and allow sufficient timeout.

What should I do about 429 Too Many Requests?

The immediate request volume exceeds the current route's protection threshold. Recommended steps:

• Implement retries with exponential backoff.
• Control the client request rate and stagger batch tasks.
• Put large batches into a queue and submit them steadily.

How do I increase concurrency?

For business workloads requiring higher concurrency quotas, contact support for an assessment.

How do I choose the right model?

Basic principles

• Prefer newer generations: Newer models in the same family usually improve performance, cost and speed.
• Choose for the use case: There is no universally best model, only the model that fits your needs best.

Recommendations by use case

• General chat and content creation: First choice: claude-sonnet-5 (balanced speed and intelligence for production); alternatives: deepseek-v4-flash (very fast, lightweight and cost-effective), glm-5.2 (a stable, reliable Chinese model).
• Complex reasoning and data analysis: First choice: claude-opus-4-8 (top reasoning for long documents and complex agents); alternatives: gpt-5.5 (a mainstay for deep thinking), claude-sonnet-5 (1M context, balanced and efficient).
• Code generation and technical development: First choice: gpt-5.6-sol (general-purpose flagship for coding and multiple agents); alternatives: gpt-5.6-terra (specialized coding with strong value), deepseek-v4-pro (deep reasoning with strong value).
• Batch processing and high concurrency: deepseek-v4-flash, glm-5.2, kimi-k3.
• Chinese-language tasks and tool calling: glm-5.2, deepseek-v4-pro.
• Image and video generation: See the Model Catalog (Models in the console) for available media models, and the Media generation section for details.