Imagine a scientist using an advanced AI model beside a national-laboratory supercomputer to test ideas that once took months to evaluate. OpenAI has outlined a broader commitment to American science, working with the US government, National Laboratories, universities and researchers. The aim is to connect frontier AI with scientific data, simulations, experimental facilities and expert teams so researchers can explore more possibilities and move from an idea to a validated result faster.
The programme includes practical resources rather than a single demonstration. As part of the US Department of Energy’s Genesis Mission, OpenAI says it will provide $4 million in Codex access to about 2,000 researchers and $3 million in API support for two large scientific campaigns. Participating researchers may also receive expanded API usage, while selected laboratory teams can gain access to specialised bioscience capabilities and advanced defensive cybersecurity tools.
Two proposed campaigns show the scale of the ambition. One would combine AI, simulation, materials expertise and physical experiments in the search for high-temperature superconductors, which could influence energy, transport, medicine and advanced computing. Another would map the so-called machine-accessible frontier: scientific problems that AI may be able to advance using knowledge, data and computation already available, while identifying questions that still require evidence from the physical world.
OpenAI is not presenting the technology as an autonomous scientist. Researchers remain responsible for choosing questions, challenging model outputs and validating results. That distinction is essential in fields where a plausible answer can still be wrong or unsafe. OpenAI and Los Alamos have already developed evaluations to study how multimodal AI can be used safely in realistic laboratory settings, including the risks created when increasingly capable systems enter consequential scientific work.
For the United States, this is an attempt to turn AI leadership into scientific and economic capacity rather than treating models only as office assistants. If the approach works, laboratories could examine more hypotheses, write and review code faster and direct expensive experiments towards stronger candidates. The broader lesson for organisations is that the next wave of AI value may come from combining models with specialist people, trusted data and existing infrastructure. Breakthroughs will still require evidence, but the path to finding them could become considerably shorter.