GPT-5.6 Luna (max)

openai/gpt-5.6-luna

Every ranked configuration of this model. Rank is dollars per unit of score on that job, and the unit is the job's own — the `per` column names it, because a rating point and a solved task are not the same thing to divide by. Displayed values are rounded; exact values in site.json and the dumps.
job	benchmark	#	score	task $	ratio $	per	released
Code	ALE-Bench	41	1667.4	$0.359171	$0.0002154	rating point	2026-07-09
Reason	ProofBench	3	0.54	$0.449325	$0.8321	correct answer	2026-07-09
Agent	DeepSWE	2	0.442	$0.777901	$1.758	solved task	2026-07-09
Agent	DeepSWE	4	0.113	$0.21631	$1.917	solved task	2026-07-09
Agent	DeepSWE	7	0.569	$1.53562	$2.701	solved task	2026-07-09
Agent	DeepSWE	13	0.672	$3.02812	$4.507	solved task	2026-07-09
Agent	DeepSWE	14	0.0155	$0.0724062	$4.675	solved task	2026-07-09

These are the numbers on #frontier. How a fixed capability's cost falls over time is on /methodology.