BREAKING
Sakana AI's 0.6B Coordinator Beats Frontier LLMs
0
B
SLM size
0
K
head params
0
K
trainable total
Thinker, Worker, Verifier Roles
1
Thinker: strategy
↓
2
Worker: solving
↓
3
Verifier: check
86.2% pass@1 on LiveCodeBench
LiveCodeBench pass@1
86.2
4-bench error cut
21.9
Beats GPT-5, Gemini 2.5 Pro, Claude-4
Baselines beaten
●
GPT-5
●
Gemini 2.5 Pro
●
Claude-4-Sonnet
Zero-shot OOD wins
●
AIME, BigCodeBench
●
MT-Bench, GPQA-D
Routing Layer as the Next Key Layer
AI NEWS BLITZ
Sakana AI says a tiny coordinator can route bigger models to beat them all.