GLM-5.3 Flash vs DeepSeek V4.1 Flash on a 174k-row Excel agent job
Deliberately boring test: Austin Animal Center's outcomes export as one xlsx (173,775 rows), plus six questions on counts, 2024 outcome shares, ages, top breeds and the busiest month. Three runs each of GLM-5.3 Flash and DeepSeek V4.1 Flash, provider pinned on every request: GLM on Z.AI, DeepSeek on DeepInfra.
Each run cost 1.6 to 3.1 cents. GLM took 4:19 to 7:21 and 10 to 21 requests, DeepSeek 6:10 to 10:58 and 20 to 33.
Checked every answer against my own pandas. DeepSeek got all six headline answers right in every run, GLM in two of three. In run 1 GLM dropped 33 rows (birth date after outcome date) from the whole table, though its notes said age answers only, so its 2024 dog and cat counts were 1 low (percentages still matched). One DeepSeek run stopped on an API error when it opened a chart image it had made, after writing both files.
SenseNova-Skills supplied both of the agent's skills, Excel analysis and its large-file helper, and every run read the sheet in chunks rather than all at once, as those skills require above 100k rows.
How quick is GLM-5.3 Flash for you on longer agent jobs?