OpenAI GPT-5.6 Sol Ultra solves 50-year math proof in under 1 hour

OpenAI's GPT-5.6 Sol Ultra solved a 50-year-old unsolved mathematical conjecture autonomously in less than 60 minutes, demonstrating reasoning capabilities that could accelerate scientific discovery and challenge the boundaries of automated research.

"This represents a step change in what AI can achieve in formal reasoning," a company spokesperson said, confirming the model produced a verifiable proof without human intervention beyond the initial prompt.

The achievement came on GPT-5.6 Sol, the flagship tier of OpenAI's three-model family released this week alongside Terra and Luna. Sol uses Ultra Mode, a new feature that coordinates four AI agents in parallel for complex workflows. The model scored 59 on the Artificial Analysis Intelligence Index, within one point of Anthropic's Claude Fable 5 at 59.9, while costing roughly one-third as much per task at $5 per million input tokens and $30 per million output tokens. It also leads the Coding Agent Index at 80 points, outperforming Fable 5 on Terminal-Bench 2.1 with a 91.9% score.

The proof — a conjecture that had resisted resolution since the 1970s — positions OpenAI's technology as a potential tool for fields where human researchers spend years on single problems. If AI can routinely generate verifiable proofs in hours rather than decades, the implications extend beyond mathematics to drug discovery, materials science, and cryptography, where combinatorial search spaces routinely exceed human capacity. The breakthrough also bolsters OpenAI's competitive position against Anthropic, Google DeepMind, and other labs racing toward artificial general intelligence.

The Reasoning Frontier

GPT-5.6's architecture supports a 1.5 million token context window, allowing the model to maintain coherence across long chains of logical deduction — a prerequisite for mathematical proof generation. The Sol tier's Ultra Mode distributes sub-problems across four parallel agent instances, each handling distinct verification steps before synthesizing results. This multi-agent approach mirrors how human mathematicians decompose complex proofs into lemmas, though at machine speed.

The model's performance on formal reasoning benchmarks suggests the technique generalizes beyond this single result. On the Coding Agent Index, Sol's 80-point score reflects its ability to generate and verify functional software, including complex applications such as macOS programs. The same underlying reasoning engine that produced a 50-year proof can also debug code, verify cryptographic protocols, and validate scientific hypotheses.

Investment Implications

The breakthrough could accelerate OpenAI's valuation trajectory as the company competes for enterprise contracts and research partnerships. Microsoft, which has invested more than $13 billion in OpenAI, stands to benefit from deeper integration of GPT-5.6 into Azure and productivity tools. Nvidia, whose H100 and B200 GPUs power the training infrastructure for models of this scale, remains the primary hardware beneficiary as AI labs race to deploy increasingly capable systems.

OpenAI's pricing strategy — undercutting Anthropic by roughly 66% on a per-token basis — pressures competitors to match cost efficiency while maintaining quality. For investors, the key question is whether performance gains continue to outpace the exponential training costs required to achieve them. GPT-5.6's training run, estimated at thousands of H100-equivalent GPU-hours, represents a bet that reasoning breakthroughs will unlock commercial applications far beyond the current chatbot paradigm.

This article is for informational purposes only and does not constitute investment advice.