Pooled recall is the size of a union. Three passes over the same 32 cases produce three sets of findings, and the score counts every case that turns up in at least one set [10][3]. That is why repetition reorders the table. DeepSeek V4 Pro 0813 returned 17 on its first pass and 28 pooled, while Grok 4.6 went from 24 to 26 because its runs kept covering the same ground, according to Aikido [6][7][14].
Price that union and the shape of the spend shows up. Three Pro passes cost about $295, so one pass is roughly $98 [4][1]. The first pass bought 17 findings at about $5.80 each [2]. The next two bought 11 more at about $17.90 each [3]. Repetition is worth buying, and it is worth buying at a rising unit price, which is what sampling a partly explored search space looks like when you write the arithmetic down.
Flash is the line worth arguing about. Three DeepSeek V4 Flash 0731 passes cost $108 and reached 24 [5]. That is 0.22 findings per dollar against Pro's 0.095, a 2.3x advantage on recall per dollar [4]. The same $295 would fund eight Flash passes instead of three Pro passes [5], and nothing in the published data says where the Flash curve flattens.
Aikido also says the Flash triple matched Grok's best individual pass for under a quarter of the cost [5]. Read strictly, one Grok pass then sits above $432 [6]. Read as a quarter of three Grok passes, it sits above $144 [6]. Per-model pass prices are not published, so the defensible form of that claim is the ratio.
The dataset caps how much of the ranking is signal. Seventeen of the 32 cases were found by every model at least once, and one case defeated all ten [13]. Fifteen cases therefore separate the field, and only 14 of those were winnable by anyone [7]. DeepSeek Pro's 28 against Grok's pooled 26 is a two-case gap inside that band [8].
What the dollar figures leave out is triage. Aikido reports that the open models caught up on recall at much cheaper rates while also producing the most false leads for the pipeline to reject [9]. A union pools findings and it pools noise, so whoever reads the queue pays for pass two and pass three a second time, in hours. The benchmark measures recall against an answer key and defines precision as the share of reported findings accepted [15]; production has no key, which means no stopping rule. Pass four costs the same $98 whether it adds a real bug or 30 turns of confident restatement.
Three settings decide whether any of this transfers. The agent got 30 turns and no internet access, inside a bounded version of the harness behind Aikido's own AI Code Analysis product [11]. The vendor sells the harness, which is a reason to read the setup section before the TL;DR. The 32 vulns were pulled from recently disclosed CVEs specifically to reduce the chance the models had seen the write-up or patch in training, with cases, prompts, tools and evaluation policy frozen across runs [12]. That is the right control for measuring reasoning rather than memorisation, and it also means your older code may not sit in the same distribution as freshly disclosed upstream bugs. The 11.7 billion tokens spread over 960 case-runs, about 12.2 million each [9], so the binding constraint in that harness was the turn cap, not the price list.
