Four-model penalty shootout: Jev and Microsoft Decision-1 save all 18
Under the same 18-shot rules, Jev and Microsoft Decision-1 both saved 18 of 18, Haiku 5.5 saved 10, and Gemini 3.8 Flash saved none.
Satya Nadella's Microsoft just saved 18 out of 18 penalties while Sundar Pichai's Gemini let in every single one and billed $300,000 for it We put 4 AI models in goal with the same rules for all: 18 shots, the ball hits the net in 0.45 to 1.1 seconds, real time Every shot = 1 model call: left, center or right Results: Microsoft Decision-1: 18/18 saves, $12,600 per 1 billion decisions TypeSafe Jev: 18/18 saves, the same $12,600 Dario Amodei's Anthropic Haiku 5.5: 10/18 saves, $40,000 per 1 billion decisions Google Gemini 3.8 Flash: 0/18 saves, still thinking when the ball hits the net, $300,000 per 1 billion decisions A 9B model from Microsoft reacts in 0.21 seconds, costs 24x less than Google's fast model, and a startup that came out of stealth last month matched it shot for shot
This post quotes another post: x.com/leopardracer/status/2108479372675977496