Astra Cracks Ten Unsolved Proofs. Washington Saw It First. The Regulatory Theater Outruns The Mathematics.
OpenAI is constructing a model family called "Astra" that orchestrates multiple agents to work on complex problems for hours or even days. CEO Sam Altman has already demoed it to policymakers in Washington. The model may ship as GPT-6 or a GPT-5 variant, and is expected to be the first subjected to the Trump administration's new AI framework, which requires models to be submitted to the federal government before public release.
The mechanism here is multi-agent orchestration. When you decompose a problem into subtasks, assign each to a specialized agent, and let them iterate collaboratively over extended time horizons, you produce emergent reasoning that no single inference pass can replicate. The ten unsolved proofs are not a parlor trick. They demonstrate that coordination architecture, not merely parameter count, drives breakthroughs.
OpenAI, under CEO Sam Altman, demonstrated Astra to Washington policymakers and released ten previously unsolved mathematical proofs as evidence of capability. The company also argued that claiming human authorship for AI-generated proofs would misrepresent genuine intellectual work, citing the Leiden Declaration on AI and Mathematics.
Step 1: Open ChatGPT and give it a complex, multi-constraint problem, such as planning a trip with a fixed budget, specific destinations, and time limits. Step 2: Ask it to decompose the problem into subtasks (routing, budgeting, scheduling), solve each independently, then identify contradictions between the solutions. Step 3: Ask it to reconcile the contradictions and produce a unified plan. Observe how the structured, multi-pass approach yields a more coherent result than a single-shot request.