BridgeBench Releases Nerf Bench to Track AI Model Performance Drift
Nerf Bench scores each model with a single power figure set to 100% at launch, and flags a real change only when results move outside 90-110% twice in a row.
BridgeBench has released Nerf Bench, a benchmark that tracks how an AI model's performance changes after launch, according to gigazine.net. The tool was published on October 13, 2026.
Nerf Bench compresses a model's results into a single "power" number derived from three inputs: performance, tokens used and cost. A model's power on its launch day is fixed at 100%, and every later result is reported as a relative percentage of that baseline.
The benchmark does not treat every fluctuation as a real shift. Power between 90% and 110% of the launch value counts as normal variation, and a genuine change is recorded only when results fall below or above that range twice in a row.
BridgeBench has not disclosed the specific tasks, test contents or test results behind the scores. The company says the omission is intended to protect reliability and to deter benchmark gaming.
Quick answers
What does Nerf Bench measure?
It tracks how an AI model's performance changes after launch, scoring each model with a single power number based on performance, tokens used and cost. Launch-day power is set at 100%, and later results are given as a percentage of it.
When is a change in a model's score counted as real?
Power between 90% and 110% of the launch value is treated as normal variation. A real change is detected only when results fall below or above that range twice in a row.
Are the test tasks and results public?
No. BridgeBench has not disclosed the specific tasks, test contents or test results, saying this protects reliability and deters benchmark gaming.