Pricing and budget ยท 2026-09-22
Qwen-Plus-2025-04-28: pinning to a dated snapshot and the cost of thinking
Covers why you'd pin to a dated Qwen-Plus snapshot, and how enabling thinking mode roughly triples the output price.
What the 'Snapshot' label guarantees
Qwen-Plus-2025-04-28 is listed on its catalog page with a 'Snapshot' status label; that means it's a version of the model frozen at a specific date. The catalog description says this directly: as a balanced general-purpose model it offers strong reasoning, long context, and tool use, and being a dated snapshot makes it better suited to integrations that want stable behavior across releases.
Pinning to a dated version, rather than using an unnamed, continuously updated 'qwen-plus' identifier, prevents behavior from silently changing when the provider updates the model in the background; the version your tests pass on stays the same version running in production.
Thinking mode roughly triples the output price
This model's pricing card carries two separate modes: with thinking off (NON_THINKING), input and output are priced separately; with thinking on (THINKING), the input price stays the same but the output price rises to more than three times as much. That difference means leaving thinking on by default on every request can add unnecessary cost on simple tasks.
On tasks requiring complex reasoning, the quality gain thinking brings can justify that cost; but leaving thinking on for short, direct questions unnecessarily raises both latency and output cost.
- Output price is low in NON_THINKING mode and noticeably higher in THINKING mode.
- Input price is the same in both modes; the difference is only on the output side.
- Keeping thinking off on simple, direct questions lowers cost.
Choose the mode per task
Rather than using the same mode for every task in a single request flow, toggling thinking based on task complexity preserves quality while optimizing cost. Building a request-routing layer (thinking off for simple questions, thinking on for complex ones) automates that decision.
Frequently asked questions
Could I just use the latest Qwen-Plus instead of pinning to a dated version?
You could, but behavior can silently change when the provider updates the model in the background. If stable behavior across releases matters, pinning to a dated snapshot is a more predictable choice.
Does turning thinking on also raise the input price?
No, on this model the input price stays the same in both modes; the difference is only in the output price.