Do reasoning effort levels actually do anything?
On GLM, yes, measurably. On DeepSeek and Kimi the parameter is accepted and the output does not change.
What we found
An API that accepts a parameter is not the same as an API that honours it, and the difference is invisible unless you go looking. We sent the same prompt at each available effort level across all five models and compared what came back.
The GLM models grade. Raising the effort produces more reasoning and a different answer, consistently enough to rely on.
DeepSeek and Kimi accept the parameter and return output we could not tell apart between levels. That is not a criticism of those models, and it may change on any deployment upstream. It does mean that tuning effort on them currently spends your attention on a knob that is not connected to anything.
This is why our published performance figures for those models come from the lab baseline set rather than from a sweep across effort levels. A sweep would produce five numbers that are really one number, printed five times.
How this was measured
An unattributed measurement is a marketing claim, so here is the receipt.
- Probe
- probe_effort_distinctness.py
- Measured
- 2026-09-08
We found this by running the probe against the models we serve, which is also why it is worth trusting: it is a reading rather than a restatement of somebody's documentation. The full catalog is on /models, and what a request costs is on /pricing.
Upstream behaviour changes without notice, and a measurement is only as good as its date. If you find this is no longer true, tell us on the support page and we will re-run the probe.