3.4 Seconds or 246: Same Model, Different Task
3.4 Seconds or 246: Same Model, Different Task
How long an LLM call takes in these findings depends first on the kind of work, and only after that on which model runs it. Qwen 3.8 Max answers language identification in 3.4 seconds [^1] but takes 246 seconds on topic sequence optimization [^2]. A latency decision therefore has to be made for each model-and-task pairing, not for a model alone.
One model, many speeds
Every figure here is a median: the time by which half of a model’s calls on a task have returned. Measured that way, the same model can move a long way between tasks. Qwen 3.8 Max also takes 187 seconds on thematic topic discovery [^3]. Grok 4.7 runs from 2.3 seconds on language identification [^1] to 138 seconds on reference-preserving analytical writing [^4].
The pattern is not limited to those two models. deepseek-v4-pro goes from 3.0 seconds on language detection [^5] to 65 seconds classifying claim continuity, a judgement over pairs of claims [^6]. Nothing in the findings supports calling a model fast or slow in general.
The kind of work sets the range
Short labelling work comes back in about a second to a few seconds, and there the choice of model has little to do with speed. On language detection, gpt-5.4-mini has a median of 722 ms, gpt-5.4-nano 860 ms and gpt-5.6-terra 913 ms [^5]. At this end of the distribution, latency gives little reason to prefer one option over another.
Long matching and writing work sits in a different range. Matching content to a topic taxonomy takes Qwen 3.8 Max 400 seconds, Tencent Hy4 Preview 314 seconds, Qwen 3.8 Flash 220 seconds and GLM-5.3 Flash 167 seconds [^7]. For taxonomy matching, topic matching, long analytical writing and structured extraction, medians run from tens of seconds to hundreds even for the faster models.
Within a task, the model order is not predictable
Model choice still matters once the task is fixed. On repairing Markdown newlines, gpt-5.6-luna has a median of 16 seconds and gpt-5.6-terra 19 seconds, while Grok 4.7 takes 206 seconds and Qwen 3.7 Plus 215 seconds over 46 calls [^8]. No spread is reported for these medians, so close pairs such as luna and terra are ties, not rankings.
What does not carry over from one task to the next is the order. Smaller or newer variants are usually quicker, but inside a family the ranking often reverses with the task:
| Models (A vs B) | Task where A is faster | Task where B is faster |
|---|---|---|
| Grok 4.7 vs grok-4.6 | Topic cluster labelling: 31 s vs 57 s [^9] | Topic taxonomy matching: 269 s vs 238 s [^7] |
| GLM-5.3 Flash vs GLM-5.3 | Topic taxonomy matching: 167 s vs 439 s [^7] | Persona policy eligibility checks: 42 s vs 35 s [^10] |
| Qwen 3.8 Flash vs Qwen 3.8 Max | Content-to-image prompts: 36 s vs 58 s [^11] | Section prompt generation: 117 s vs 76 s [^12] |
A rule such as “the Flash tier is the fast one” holds on some tasks and fails on others. The only reliable unit is the model-and-task pairing.
What to do about it
Classify the work before choosing the model. For short work where a person is waiting, such as language detection, latency will rarely separate the candidates, so let quality or cost decide.
For long work, look up the exact pairing you plan to run instead of borrowing a model’s speed from some other task. Where that work is interactive, build the wait around the pairing’s measured 90th percentile, the time by which nine in ten calls return, rather than its median.
Taxonomy matching, topic discovery, long analytical writing and structured extraction belong behind a queued or batch path. Keep the synchronous lane for work whose value decays while it waits.
The live figures are at https://fronset.ai/benchmark/, where you can set the quality bar your work actually requires.