OpenAI grants Resolution $10M in API credits
OpenAI has granted Resolution $10M in API credits, bringing our total support to $170M. We're pleased to see OpenAI investing in long-term safety on several fronts: through this grant, its contribution to UK AISI's Alignment Project, and its new RSI preparedness team. We hope to see OpenAI—and every frontier lab—invest much more in long-term safety, both internally and by supporting independent research.
Alongside our engineering work, these credits will be especially helpful for our autoformalization research, where we're scaling the production of reliable, machine-verifiable proofs for alignment.
Why does theory need so many tokens?
AI is becoming strong at mathematics. Recent months saw impressive mathematical contributions from AI: OpenAI’s disproof of the Erdős unit-distance conjecture, AlphaProof Nexus' resolution of nine open Erdős problems (Tsoukalas et al. 2026), a Fable 5-assisted counterexample to the Jacobian conjecture, and OpenAI Astra’s ten advances in mathematics and theoretical computer science.
Stronger AIs can use more tokens (McFadyen et al. 2026). Newer models are better at exploring many different approaches, following long chains of reasoning, critiquing failed attempts, and verifying their own conjectures.
Humans will get better at scaffolding. In particular, autoformalization creates a ground-truth feedback signal to allow models to work productively for longer.
This is already expensive. Claude Mythos Preview found mathematical flaws in two cryptographic algorithms, each at a cost of roughly $100,000.
With stronger models and better scaffolding, the useful scale of these projects could grow rapidly within the next year.
Join us
Turning these credits into alignment progress won’t happen by default. We’re hiring a CTO to lead that conversion and prepare us for more. If you’ve run a large engineering team in an ML research context and have strong views on how to scale research automation, please reach out to jesse@resolution.org. If you can contribute in other ways, apply here.