How BenchAlign rankings work
TokenBench republishes BenchLM's BenchAlign results without recalculating them. Read BenchLM's methodology for the source method.
What each view represents
Overall, Agentic, and Coding are validated BenchAlign views. Reasoning, Multimodal, and Knowledge are BenchLM-published category evidence lenses, not additional TokenBench rankings.
Supported rows are source-published results eligible for their exact view. Reviewed estimated rows stay visibly estimated and appear after supported evidence where a route allows them; they are never silently promoted into a validated ranking. Missing measurements remain Unavailable, never zero.
Metrics and runtime
Weighted metrics affect the relevant BenchAlign method only. Display-only metrics add context without changing the published order. Runtime is a separate signal, not a substitute for capability evidence or a hidden ranking weight.
Method and refresh status
Published method version: Unavailable.
BenchLM refreshes its source output on its own schedule. TokenBench checks that source once daily within its broader Worker, which runs twice daily; a successful TokenBench check does not claim that BenchLM published a new method or result.