# Reasoning effort (effort tiers and their cost)

> VI: https://5ac.vn/wiki/reasoning-effort/ · EN: https://5ac.vn/en/wiki/reasoning-effort/ · lang: en · spec_as_of 2026-10-21 (spec no formal spec) · cluster C8

The tier a caller sets to decide how many thinking tokens a model may spend before answering; higher effort buys latency and cost, not correctness, on short tasks.

---
## Definition

Reasoning effort is the tier a caller sets on a reasoning model to decide how many thinking tokens the model may spend before it writes the answer. It is richer than a binary thinking on-off switch: each vendor publishes its own accepted set, some accept a no-thinking tier, and requests outside a vendor's set bounce back as HTTP 400. One distinction worth keeping: the [sub-agent](/en/wiki/sub-agent/) entry uses the phrase Effort control for scaling delegation between agents; this entry is about thinking cost inside a single model.

## Mechanism

The tiers are real money. Artificial Analysis estimates one DeepSeek agent task emits roughly 89,000 output tokens, about 63,000 of them reasoning, billed at output rates. 5ac's own probes on the Chinese flash-model line: GLM 5.3 Flash keeps thinking on at all times, accepts only low, high or max, and its JSON pass took 54.4 seconds per call. Qwen 3.8 Flash accepts all seven tiers including none, so extraction with thinking off is the cheapest job in the fleet. MiMo V2.6 Flash burned reasoning tokens even on a three-word reply. The rule learned the expensive way: on short tasks, reasoning effort buys latency, not correctness — GLM passed 4/4 at every tier ran, while the max tier cost 2.9x the low tier's wall time.

## Application context

For teams shipping agents, this variable sits on the cost boundary. Structured extraction and long prose want the low tier; only genuinely multi-step reasoning, math or hard code justifies max. Pin one tier inside the app config and measure pass-rate before raising it. The tier set for the main model is not automatically valid when a request falls through to a fallback model, so the fallback card needs its own effort value. Anything that does not need real-time answers can ride the batch lanes, where vendors discount.

## Sources fetched

- [Chinese Flash Models for Agent Harnesses: Benchmarks and Prices Compared](/en/blog/chinese-flash-models-agent-harness-comparison/) — 89,000/63,000 token anatomy, GLM 54.4s and 4/4 at every tier, 2.9x max-tier wall time — accessed 2026-10-21

## Related

- [Sub-agent](/en/wiki/sub-agent/)
- [AI token cost](/en/wiki/ai-token-cost/)
- [Model routing](/en/wiki/model-routing/)