Deep suiteClaude Opus 5.5 had the best average on five of six tasks, two alone and three tied. It also had the highest estimated cost on all six.Two runs per model on each task.

SkillsToken-saving skills hit their claim in 1 of 29 checks.Numbered claims from two skills, checked on six models.

Deep suite: score against estimated cost

Speed-up (C++), one of six tasks, four models. Higher is better, and left is cheaper.

Claude models in Claude Code
GPT models in Codex CLI

Show as a table

Model	Score	Cost per run	Output tokens	Minutes	Frontier
GPT-6 Luna	0.07	$0.014 to $0.016	2,948	1.2	on it
Claude Haiku 5.5	1.01	$0.14 to $0.72	107,036	14.4	on it
GPT-6.1 Sol	1.08	$0.71 to $0.76	27,653	11.0	on it
Claude Opus 5.5	1.66	$4.00	89,104	20.1	on it

Source: Daily Driver Bench at dailydriver.brenhq.com, updated 16 October 2026. Deep suite v3 has six tasks and 48 runs, all at medium effort. Each dot is a model's mean on the chart's scale, and an invalid run counts as 0. 11 of the 48 runs hit the time limit and were graded on the code at that point, nine of them Claude Opus 5.5's. A dot sits at the middle of its estimated cost range.

Copy link Export PNG