How to cut your Alibaba Qwen API bill

Mode

Both: automatic caching by default, plus explicit breakpoints you control

Implicit cache always on (cannot be disabled, hits not guaranteed); explicit cache via cache_control markers (deterministic hits). Mutually exclusive per request. Responses API has separate Session cache settings.

Cache lifetime

Explicit: 5 minutes, reset on each hit. Implicit: indeterminate, old unused data cleared periodically.

Minimum prompt

1,024 tokens for both modes (512 for Zhipu-deployed GLM models on implicit cache)

Cache write cost

Explicit: typically 125% of standard input (only the incremental portion when extending an existing cache). Implicit: 100% of standard input (no premium).

Cache read cost

Explicit hits typically 10% of standard input (exceptions: qwen3.8-max, qwen3.8-max-0902, qwen3.8-flash, qwen3.8-2.4t-a95b). Implicit hits typically 20% (model exceptions listed).

Storage cost

Not stated

How to enable

Implicit: automatic. Explicit: add a cache_control marker to a messages content block (up to four markers; 20-block lookback).

Request fragment

Copy

"cache_control": {"type": "ephemeral"}

From the docs

“Explicit cache: A mode that you enable manually. You create a cache for specific content to ensure a deterministic hit within its 5-minute validity period.”

“Implicit cache: This automatic mode requires no extra configuration and cannot be disabled, ideal for scenarios that prioritize convenience.”

“| Billing for cache creation tokens | Typically 125% of the standard input token price | 100% of the standard input token price |”

“| Billing for cached input tokens | Typically 10% of the standard input token price (see Billing) | Typically 20% of the standard input token price (see Billing) |”

“| Minimum tokens for caching | 1024 | 1024 |”

“| Cache validity period | 5 minutes (resets on hit) | Indeterminate. The system periodically clears old, unused cache data. |”

“Explicit cache and implicit cache are mutually exclusive.”

“For Zhipu-deployed GLM models, the minimum is 512.”

“Add a`"cache_control": {"type": "ephemeral"}` marker to the messages array.”

“A single request supports up to four cache markers.”

Discount

50% off

50% of real-time price (input and output, successful requests only)

Window

completion_window 24h to 336h (14d); unfinished requests in expired tasks are not executed or charged

Limits

Max 50,000 requests per file, max 500 MB; each line max 6 MB; one model and thinking mode per file; does not support context cache

Endpoint

OpenAI-compatible Batch File API: client.batches.create (endpoint /v1/chat/completions or /v1/embeddings)

Stacks with caching

No: you get one discount or the other, not both

“It also does not support features such as context cache.”

From the docs

“The input and output tokens for all successful requests are charged at 50% of the real-time inference price for the corresponding model.”

“completion_window | String | Body | Yes | Maximum wait time. Range: 24h-336h, integers only.”

“Size limits: Maximum 50,000 requests per file, maximum 500 MB.”

“Line limit: Each JSON object must not exceed 6 MB and must fit within the model's context window.”

“It also does not support features such as context cache.”

“If a task does not complete within the completion_window, it expires. Unprocessed requests in expired tasks are not executed and do not incur charges.”

Service tiers on the pages checked: Not stated

• AI general-purpose savings plan (also applies to batch)

  “It supports AI general-purpose savings plan, but not discounts, such as subscription (other savings plans) or free quotas for new users.”

• Subscription / other savings plans and free quotas for new users (real-time only)

  “It supports AI general-purpose savings plan, but not discounts, such as subscription (other savings plans) or free quotas for new users.”