TM GO · model selection

Which model should be making the call?

Every model below ran the same 42 golden email chains through the same production prompt and was graded by the same scorer, so the only variable is the model. Accuracy, cost and latency are recorded in one pass, because a model that is cheaper and worse is not cheaper.

Incomplete comparison: claude-haiku-4.5 could not be reached on this run, so the head-to-head is OpenAI-only. Those rows are blank rather than zero.

gpt-4o costs 17.3× what gpt-4o-mini costs and scores +0.0 points for it.

Accuracy

mean score, 42 cases

gpt-4o-mini100.0%

gpt-4o100.0%

claude-haiku-4.5unavailable

Bars scaled to the best score.

Cost

USD per full golden run

gpt-4o-mini$0.0048

gpt-4o$0.0832

claude-haiku-4.5unavailable

Bars scaled to the most expensive.

Latency

average seconds per call

gpt-4o-mini0.97s

gpt-4o0.86s

claude-haiku-4.5unavailable

Bars scaled to the slowest.

model	mean	passed every run	flaky	latency/call	cost/run
gpt-4o-mini	100.0%	100%	0	0.97s	$0.0048
gpt-4o	100.0%	100%	0	0.86s	$0.0832
claude-haiku-4.5	unavailable this run