Bake AI’s Épi agent reports 120 record results with GPT-6 Astra

Your AI and this story

Key facts

• Agent: persistent research history, hierarchical memory, branching and independent review; experiments run across model sessions and ‘a fixed verifier evaluates their outputs’. Models: ‘We ran Épi over a few days using OpenAI frontier models (GPT-5.6 Sol and GPT-6-Astra)’
• Claimed: 120 record-setting results, 20 numerical bound records across 12 families and 100 construction records across 8 families (packings etc.), compared with literature and registries as of 16 Sept 2026; the results repo says ‘106 improved constructions, 15 mathematical results, and their certificates’
• n-queens constant α (Q(n) = (n e^{−α+o(1)})^n): interval narrowed from Nobel et al. (2023) [1.944000752, 1.944001082] to 1.9440010223327 ≤ α ≤ 1.9440010262694 (width 3.30×10⁻⁷ → 3.94×10⁻⁹, ≈84×), each endpoint with a retained witness
• Cohn–Elkies: new dual lower bounds on the LP optimum in dimensions 9, 10, 11 and 13; in dimensions 10 and 11 they exceed the packing upper bounds of Cohn et al. (2022) by ~0.35% and ~1.92%, so ‘Even exact optimization of this formulation cannot recover the true packing density’ there
• Quartic Feigenbaum attractor dimension: a rigorous interval more than 3.6-million-fold narrower (claimed)
• Tuza’s conjecture (1981, τ△ ≤ 2ν△): computer-assisted proof for split graphs with an eight-vertex clique part and at most three triangle-active neighbourhood types, arbitrary multiplicities (a special case, not the conjecture)
• Épi-turbo: on 30 FrontierOR tasks, a 5.4% higher total score than Codex (with GPT-6 Astra) at 64% of its model cost, same model and reasoning effort
• Status: technical report; certificates public on GitHub (BakeLab/Epi-results, created 7 Sept); no independent check found on 24 Oct