Atomic Canyon Launches NuclearBench AI Benchmark

The open-source tool tests large language models on nuclear-energy questions, developed with Idaho National Laboratory.

Updated on Nov. 5, 2026 in Nuclear

Atomic Canyon Launches NuclearBench AI Benchmark

Atomic Canyon has launched NuclearBench, an open-source benchmark for evaluating large language models on nuclear-energy questions. The benchmark scores model responses on answer accuracy and faithfulness to provided nuclear source material.

Why it matters

Nuclear operators and regulators increasingly consider AI tools, and NuclearBench gives organizations a public way to test how accurately models handle nuclear power operations. Publicly available testing of this kind can help the nuclear sector assess AI reliability before deployment.

NuclearBench was built around 280 questions and scores models on accuracy and faithfulness to nuclear source material. The extraction does not report how specific models performed on the benchmark.

The players

Atomic Canyon

Atomic Canyon is a company that develops AI tools for the nuclear energy industry, including its flagship product Neutron Enterprise.

Idaho National Laboratory

Idaho National Laboratory is a U.S. Department of Energy national laboratory focused on nuclear energy research and collaborated on the benchmark's development.

The details

NuclearBench was developed in collaboration with Idaho National Laboratory, which contributed subject matter, datasets, and evaluation frameworks. It builds on Atomic Canyon's earlier FermiBench, which tests AI embedding models on retrieving nuclear records, and it is designed to be updated as new AI models become available.

November 5, 2026: Atomic Canyon announced the launch of NuclearBench.

Big Picture

NuclearBench extends Atomic Canyon's FermiBench, which benchmarks AI embedding models in retrieving nuclear records, by shifting the focus to whether language models answer nuclear-energy questions accurately and faithfully. The Nuclear Industry Virtual Assistant, developed with INPO, EPRI and NEI, shows the same company's effort to bring AI into nuclear operations.

For people working in or following the nuclear power industry, NuclearBench offers a public test for judging whether an AI model's answers on operations questions can be trusted. Organizations can run the open-source benchmark themselves rather than relying only on vendor claims.

The takeaway

Before relying on an AI tool for nuclear operations questions, organizations can test its accuracy and faithfulness to source material using a public benchmark. Checking answers against original documentation remains essential even when a model performs well on such tests.