Google Deepmind's AlphaProof Nexus Solves Decades-Old Math Problems for a Few Hundred Dollars

Why It Matters

By grounding LLM reasoning in formal verification, AlphaProof Nexus shows that cost‑effective AI can crack decades‑old math questions, shifting the focus from headline‑grabbing raw language‑model feats to practical research tools.

Key Takeaways

• AlphaProof Nexus solved 9 of 353 Erdős problems, two 56‑year‑old
• Inference cost per problem stayed under a few hundred US dollars
• Simple Agent A matched full Agent D on solved problems, just costlier
• System proved 44 of 492 OEIS conjectures and improved convex‑optimization bound
• Formal Lean feedback grounds LLM reasoning, turning failed attempts into insight

Pulse Analysis

The race to embed artificial intelligence in pure mathematics has accelerated in recent years, with OpenAI’s GPT‑series and DeepMind’s AlphaTensor paving the way. While earlier breakthroughs relied on raw language‑model reasoning, they often struggled with logical consistency and required massive compute budgets. AlphaProof Nexus changes the equation by pairing a Gemini 3.1 Pro model with the Lean theorem prover, creating a feedback loop where each generated proof step is instantly validated. This hybrid architecture reduces hallucination risk and translates abstract model output into verifiable mathematical statements, a crucial advancement for rigorous research.

AlphaProof Nexus operates through four progressively sophisticated agents. Agent A runs a simple loop: the LLM proposes a Lean step, the compiler checks it, and error messages guide the next attempt. Agents B and C enrich this loop with AlphaProof’s reinforcement‑learning queries and an evolutionary population of proof sketches, respectively, while Agent D combines all capabilities. Despite Agent D’s edge on the toughest Erdős challenges, post‑hoc analysis revealed that Agent A alone solved every problem the system ultimately cracked, though it required higher computational spend. The cost per problem—only a few hundred dollars—demonstrates a dramatic efficiency gain compared to earlier AI‑only approaches that demanded orders of magnitude more resources.

The broader implication is a shift toward AI tools that augment, rather than replace, human mathematicians. Formal proof sketches generated by AlphaProof Nexus can surface hidden sub‑goals, allowing researchers to focus on the most promising avenues. Early adopters in quantum optics and graph theory report that even failed attempts sharpen their intuition. As language models continue to improve, the gap between simple agentic loops and more complex systems is expected to narrow, positioning frameworks like AlphaProof Nexus as the backbone of future AI‑driven scientific discovery.