Measurement 4 of 4 · timed on video
Speed & temperament
Four terminals raced side by side on one recorded screen, and this time there's a stopwatch. The twist: the cheapest model finished last, and the two slowest were the only ones that tested their own work before calling it done. Fast, cheap, careful — nobody got all three.
Time to done
▶ Watch the race on X ↗Wall-clock from prompt to done, timed on the side-by-side race recording (published with the X post above). Sorted fastest first. ✓ = the model tested its own work before finishing.
Times are from the race recording; token traits are from each CLI's own logs.
What the logs can quantify
| Model | Observable trait | Evidence |
|---|
Why speed made no difference to quality
Grok finished in 4:41 without testing; DeepSeek took 10:17 and spent 32,120 tokens thinking plus a self-check before shipping. A 2.2× time spread, four opposite strategies — and the accuracy table can't tell them apart. Note the inversion against cost: the $0.045 model was the slowest, the priciest was second-slowest. You pay in dollars or in minutes, not neither.
What it does affect
Workflow feel. A sprinter is a joy for tight iteration loops; a deliberator inspires trust on tricky specs; a heavy writer produces code you enjoy reading. Since quality converged, this — along with the bill — is a legitimate reason to pick one model over another. Taste is allowed to be the tiebreaker when correctness ties.