AI systems are now rapidly producing major mathematical results, including a Lean-checked forced Navier–Stokes blowup AI Has Solved One of Math's $1 Million Millennium Prize ..., an ETH refutation, and Yang–Mills/QCD existence proofs OpenAI Says Its Models Produced Three Proofs in Math .... However, these are pure mathematical achievements that do not independently satisfy the criteria. The binding constraint is that an AI-produced theorem, construction, or counterexample must be explicitly cited as the key step in a separate empirical result—a measurement, working design, or data-confirmed prediction. As of October 2026, no such empirical application has been reported The Future of AI-assisted Physics: Discovering, Verifying ..., and experts note that even landmark singularity proofs currently lack immediate practical consequences for physics or engineering AI Has Solved One of Math's $1 Million Millennium Prize ...AI Has Solved One of Math's $1 Million Millennium Prize ....

The most plausible route to an early qualifying result is through applied constructions rather than famous pure-mathematics theorems. Quantum computing and error-correcting codes are particularly promising. For example, Claude Mythos 5 recently designed quantum circuits that beat published records, deriving and Lean-verifying new algebraic identities alongside Harvard's Lukin and Yelin Designing more efficient quantum circuits with Claudehttps://www.anthropic.com/research/designing-efficient-quantum-circuitshttps://arxiv.org/pdf/2610.22367. Similarly, LLM-guided searches have identified hundreds of candidate error-correcting codes Can LLMs discover quantum error correction codes?. Currently, these advances remain at the level of simulations and resource estimates Anthropic's Claude Designs Leaner Quantum Circuits .... A breakthrough will likely occur when such an AI-proved construction is implemented in a physical hardware demonstration or when a theoretical prediction is tested against existing data.

Beyond experimental success, the resulting empirical paper must explicitly state its dependence on the AI theorem, and the discovery must be ranked among the year's top scientific breakthroughs, such as Science's Breakthrough of the Year or Physics World's Top 10 Top 10 Breakthroughs of the Year in physics for 2025 ...Science's 2025 Breakthrough of the YearTop 10 Breakthroughs of the Year in physics for 2025 .... Because the empirical applications do not yet exist, the December 2026 lists will almost certainly not feature a qualifying result. While Science already recognized "LLMs in science" as a 2025 runner-up Science's 2025 Breakthrough of the Year—proving editors are open to AI contributions—the strict requirement for explicit dependency is a high bar. Furthermore, these lists historically tend to credit physical hardware demonstrations over their theoretical underpinnings.

Timelines for this milestone are pulled by competing forces. Historically, decades-old theorems take years to feed into applied work, and physical experiments have inherent delays in hardware construction, execution, and peer review. Set against the timelines for pure mathematics breakthroughs , this milestone should therefore occur somewhat later: a theorem must first be integrated into empirical hardware, and the resulting achievement must then receive official year-end recognition. As Terence Tao has observed, proofs are currently arriving faster than researchers can absorb or integrate them Can maths survive AI? Fields medallist Terence Tao calls .... Conversely, theorem-to-application lags could compress significantly: AI labs have strong incentives to partner directly with top experimental groups, AI systems can now guide both the theorem generation and the downstream applied analysis, and computational engineering designs can iterate far faster than traditional wet-lab cycles.

Balancing these dynamics, the 2027 list cycle represents the first realistic opportunity, most likely via an AI-designed quantum circuit or error-correcting code that is rapidly deployed in a hardware demonstration and receives immediate annual recognition. I assign a 10% probability of a qualifying result by June 1, 2027, representing an unusually fast path through application and recognition, and 25% by January 1, 2028, after the 2027 list cycle. However, given typical 1–3 year hardware implementation cycles and the rigorous bar for independent, top-of-year standing, my median date is January 1, 2029, following the 2028 year-end lists. I put the cumulative probability at 75% by June 1, 2031, and 90% by January 1, 2036. The long upper tail reflects the persistent risk that AI's theoretical output remains difficult to translate into physical breakthroughs, or that clear attribution and explicit dependency fail to align cleanly with top-list recognition criteria.