Open Benchmark Tests AI on Nuclear Questions
The free tool, built with Idaho National Laboratory, gives operators a common test for AI accuracy in nuclear-energy work.
Updated on Nov. 5, 2026 in Nuclear

Atomic Canyon has launched NuclearBench, an open-source benchmark for testing large language models on nuclear-energy questions, built with Idaho National Laboratory. The benchmark is a research-stage evaluation tool, publicly available for any organization to use, not a deployed product.
Why it matters
As utilities and regulators weigh AI tools for nuclear plant operations, a shared yardstick for model accuracy on nuclear questions gives buyers a way to compare claims. NuclearBench extends Atomic Canyon's earlier FermiBench, which tests embedding models on retrieving nuclear records.
NuclearBench is built from 280 questions and scores models on answer accuracy and faithfulness to provided nuclear source material. The source does not report scores for any specific model, so the benchmark's discriminating power has not been demonstrated publicly.
The players
Atomic Canyon
A San Luis Obispo, California company building AI tools for the nuclear industry, whose flagship product is Neutron Enterprise and whose prior FermiBench benchmark tests embedding models on nuclear records retrieval.
Idaho National Laboratory
A U.S. Department of Energy national laboratory in Idaho Falls, Idaho, that contributed subject-matter expertise, datasets and evaluation frameworks to NuclearBench.
The details
NuclearBench tests a large language model, a text-prediction system trained on large volumes of writing, by asking it nuclear-energy questions and grading two things: whether the answer is correct and whether it stays faithful to the nuclear reference material supplied with the question. Idaho National Laboratory contributed subject-matter expertise, datasets and evaluation frameworks to the design. Because it is open source, any organization can run the same questions against its own models and compare results directly.
Timeline
• November 5, 2026: Atomic Canyon announces the launch of NuclearBench.
The Tech Race
NuclearBench extends Atomic Canyon's FermiBench, which evaluates AI embedding models on retrieving nuclear records, from retrieval to question answering. The source does not name competing benchmarks in nuclear AI, so the wider competitive position is not established.
NuclearBench is publicly available now, so any organization can test language models against nuclear-energy questions. The source gives no pricing or timeline for broader deployment of tools built on it, so the practical reach for everyday users is not yet clear.
The takeaway
Watch for published model scores on NuclearBench, which would show how current systems actually perform on nuclear questions rather than claimed accuracy. Benchmark updates as new models appear will show whether the tool keeps pace with the field.
Further reading
For broader context on how the country is approaching reactor technology, see Nuclear coverage across United States Nuclear.
Source note: This article includes information reported by The Manila times.





