How GPT-5.6 Sol's 90-Minute Math Proof Rewrote the Benchmark for AI Reasoning

The Day's Thesis

▶

Signal of the Day: GPT-5.6 Sol disproved a 30-year-old statistics conjecture in 90 minutes — a task GPT-5.5 failed after 20 hours — setting a concrete capability threshold that no prior AI benchmark has captured.

The 30-Second Read:

• GPT-5.6 Sol cracked a Benjamini-Hochberg conjecture in 90 minutes; GPT-5.5 could not in 20 hours
• IBM shares fell 25% in a single session to ~$217, the steepest decline in the stock's modern history
• Anthropic is scheduling investor meetings for a potential IPO as early as October 2026
• ASML raised its full-year sales forecast for a second time in 2026 on sustained AI chip demand

Today's news forces a single, uncomfortable question: if AI models can now disprove mathematical conjectures faster than credentialed researchers, what does that imply for the valuation of companies still selling yesterday's compute or productivity tools? IBM's 25% collapse and Anthropic's IPO sprint are two answers to the same question, arriving simultaneously.

---

AI & Research Frontier

GPT-5.6 Sol's 90-minute disproof of a 30-year-old statistics conjecture is the first documented case of a frontier model resolving an open academic problem that a prior-generation model explicitly failed to solve.

The predecessor, GPT-5.5, logged 20 hours against the Benjamini-Hochberg (a multiple-testing correction procedure used in clinical trials and genomics) conjecture without resolution; GPT-5.6 Sol Pro closed it in 90 minutes by combining known methods in a novel configuration. The episode does not settle whether the model produced genuinely new knowledge or sophisticated recombination — but that philosophical distinction is increasingly irrelevant to research teams measuring time-to-solution.

Separately, OpenAI built GPT-Red, an internal adversarial LLM (a model trained specifically to probe and exploit another model's weaknesses) used as a red-team sparring partner; the company credits GPT-Red's pressure as a primary driver of GPT-5.6's security hardening. On the edge side, PrismML compressed a 27-billion-parameter model — Bonsai 27B — to under 4 GB, retaining 90% of benchmark performance and enabling on-device inference on an iPhone, with Apple reportedly evaluating the compression stack.