Stata Translation Workshop · Post-Workshop Survey
9 respondents. One bar per model, height = mean time.
Opus's average (109 min) is pulled up by a single 180-minute
attempt — its four attempts actually ranged 60–180 (see the table
below). Haiku and Haiku 4.5 were the fastest (5–10 min), but each is a
single data point. Fable's 30 minutes isn't a completed translation —
that run stopped when the respondent ran out of tokens/credits, so it
understates how long Fable would actually take.
With 1–4 attempts per model and effort/complexity varying by task,
this is far too little data to conclude one model is reliably faster
than another — read it as what happened in these 9 attempts, not a
benchmark.
| Model | Mean (min) | Range | n |
|---|---|---|---|
| Opus | 109 | 60–180 | 4 |
| Sonnet | 75 | 60–90 | 2 |
| Fable | 30* | — | 1 |
| Haiku 4.5 | 5 | — | 1 |
| Haiku | 10 | — | 1 |
| Model | Time (min) | Note |
|---|---|---|
| Fable | 30 | stopped – ran out of credits |
| Opus | 60 | |
| Opus | 75 | "with lot of debugging" |
| Opus | 120 | |
| Opus | 180 | |
| Sonnet | 60 | |
| Sonnet | 90 | |
| Haiku 4.5 | 5 | |
| Haiku | 10 |