Surprising results: Gemini 4 Neon vs Claude Fable 5.2, head to head on our hardware.
Dead heat on every test with a right answer. Neon won the open-ended one, Connect Four, 68.5/98. Fable used 1/5 the tokens.
Not the same model. Both frontier. Google is back at the frontier.
Method: same sandbox, briefs and time limits for both. Every score from a program, not a judge: generated reasoning puzzles, real bug fixes graded by the real commits' tests, a regex engine from scratch, 181K-token context, code review, each model setting exams for the other. #AI
2
90
This path is not GPU-bound: the cards drew 175 to 380 W of their 600 W. Tencent reports 150+ tok/s on its own stack with megakernels. The official vLLM image does not start on these cards yet, because its sparse attention kernel is built for data-centre GPUs.
1
29














