rcoh/angle-grinder | ProgramBench

Slice and dice logs on the command line

1,128

Generated Behavioral Tests

94.1%

Best Score

Claude Opus 5 (xhigh)

Results by Model

Hover for details · Click a model to open it · Drag to zoom the x-axis · Shift-drag to pan · The line marks the Pareto frontier (best score per cost)

25 runs

#	Model	Score	Cost	Calls
1	Claude Opus 5 (xhigh)	94.1%	$76.50	385	trace →
2	Muse Spark 1.3 (max)	85.3%	$3.40	250	trace →
3	Claude Opus 4.8 (xhigh)	77.7%	$19.74	148	trace →
4	GPT 5.5 (xhigh)	73.4%	$10.07	111	trace →
5	GLM-5.2	70.7%	$39.36	231	trace →
6	Muse Spark 1.3 (xhigh)	70.0%	$1.36	214	trace →
7	GPT 5.5 (high)	64.9%	$4.46	44	trace →
8	GPT-5.6 Sol (xhigh)	56.3%	$3.72	28	trace →
9	Muse Spark 1.1 (xhigh)	52.9%	$0.15	145	trace →
10	Claude Opus 4.7 (xhigh)	50.5%	$11.18	211	trace →
11	Muse Spark 1.2 (xhigh)	48.7%	$3.45	271	trace →
12	GPT-5.6 Sol	48.3%	$1.02	18	trace →
13	Gemini 3.5 Flash	46.8%	$8.07	253	trace →
14	GPT 5.5	44.3%	$1.23	20	trace →
15	Claude Opus 4.7	38.0%	$3.55	84	trace →
16	Gemini 3.6 Flash	37.6%	$4.27	192	trace →
17	GPT 5.4	22.0%	$0.49	15	trace →
18	Gemini 3 Flash	20.2%	$0.25	73	trace →
19	Gemini 3.1 Pro	18.8%	$1.46	114	trace →
20	Claude Haiku 4.5	13.1%	$1.18	147	trace →
21	GPT 5 mini	6.9%	$0.03	17	trace →
22	GPT 5.4 mini	6.8%	$0.07	12	trace →
23	Gemini 3.7 Flash	0.4%	$1.47	133	trace →
24	Claude Sonnet 4.6	0.4%	$37.01	660	trace →
25	Claude Opus 4.6	0.0%	$3.83	303	trace →

Click a row to replay how that model rebuilt this program, or the model name to open its full run