Atomic Canyon, the California company behind the Neutron software for the nuclear sector, announced on November 5, 2026 the launch of NuclearBench. It is the first open source benchmark designed to evaluate how large language models answer questions about nuclear energy. The work was developed together with Idaho National Laboratory (INL).
The test works like an open-book exam. Founder and CEO Trey Lauderdale explained it to Heatmap, which broke the news. The exam consists of 280 questions, prepared with INL and put to AI models. Responses are graded on two criteria: accuracy and adherence to the nuclear source material supplied to the model. Anyone can use it, whether a utility, a regulator or a software vendor.
NuclearBench builds on FermiBench, the benchmark Atomic Canyon has already published to measure how well models retrieve the right information from large archives of nuclear documents. FermiBench includes 156 queries across 4,558 documents from the Nuclear Regulatory Commission’s ADAMS archive. The new test shifts the evaluation from document retrieval to answer generation.
The project starts from a gap. Interest in AI at nuclear plants is growing, but there were no standard metrics for comparing product quality. Generic benchmarks do not account for the specialized language, regulatory framework and safety context typical of nuclear operations. The agreement with INL was announced in September 2025. In parallel, Atomic Canyon is working with Oak Ridge National Laboratory, where it trained the FERMI models on the Frontier supercomputer using more than 53 million pages of NRC documents.
Jess Gehin, INL’s associate laboratory director for nuclear science and technology, points to benchmarks as the tool for understanding where AI delivers real value and adopting it accordingly. Lauderdale admits the test is not yet on par with a medical licensing exam. That comparison remains the end goal: in healthcare, new models are measured against standard tests, while in nuclear they have not been until now.
A common yardstick helps those who have to make purchasing choices. NuclearBench sits alongside the Nuclear Industry Virtual Assistant (NIVA), developed with INPO, EPRI and NEI, which already uses Atomic Canyon’s retrieval technology to deliver authoritative technical information to professionals. Operators will now be able to compare models using the same criteria.
With public data, definitions and tools, the next version can be expanded by the community. If utilities and regulators adopt the test, its results could factor into software procurement decisions.
