Benchmark
Nerf Bench
Track whether a deployed model still completes the same work as efficiently as it did at launch.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
BridgeMind retests each model on a private task suite and compares it with its own launch baseline. The Claude Opus 5.5 page is one model’s history within the same benchmark.
Methodology
The index weights task performance at 60%, token use at 20% and cost at 20%. BridgeBench looks for two consecutive readings outside a 10% band. A decline can reflect cost or token use as well as answer quality; it does not establish a model change by the provider.