Automated reasoning costs plummet across complex mathematical and visual puzzles

Plunging compute expenses make high-level logical problem solving widely accessible, allowing small teams to explore formal mathematical proofs and adaptive planning tasks once restricted to supercomputing clusters.

For decades, solving an entirely unfamiliar logic puzzle required the flexibility of human thought. Computers could store vast libraries of chess openings or calculate arithmetic at blinding speeds, but confronting a novel game with unstated rules routinely exposed the limits of machine learning. When artificial intelligence systems attempted to reason through unfamiliar scenarios, the computational power required to explore every plausible hypothesis quickly made experiments prohibitively expensive. That economic barrier has begun to collapse.

In September 2026, research released by OpenAI and the benchmark organization ARC Prize revealed that new reasoning architectures can solve advanced logic and mathematics problems at a fraction of earlier operating expenses.¹² OpenAI research scientist Noam Brown announced that the company's GPT-6 Astra model now matches and exceeds difficult benchmark thresholds for the price of an ordinary lunch.¹ At the same time, OpenAI reported deploying an internal system of coordinating agents to solve a case of the Navier-Stokes Millennium Prize problem, demonstrating how automated reasoning scales to open mathematical questions.³¹

Why does novel problem solving demand so much computation?

Solving an unfamiliar problem forces an artificial intelligence system to explore many alternative paths before finding a valid sequence of steps. In an unfamiliar environment, a model cannot rely on memorized training data because the rules must be inferred from direct interaction. The system must observe an initial state, generate candidate hypotheses about underlying mechanics, simulate potential actions, and test whether its predictions match reality. Because every prospective move branches into multiple future possibilities, unguided trial and error creates an exponential explosion of options that consumes immense computing power. A reasoning system becomes efficient only when it can distill observations into compact symbolic rules, retain internal state across steps, and discard flawed strategies early in the search.

What did GPT-6 Astra achieve on logic benchmarks?

OpenAI's GPT-6 Astra achieved state-of-the-art scores across multiple reasoning benchmarks while drastically reducing the financial cost of computing the answers. According to a public announcement by Brown, when OpenAI originally released its o3 model, achieving an 87.5 percent score on the ARC-AGI-1 benchmark required approximately $500,000 in computational expenditure.¹ Brown stated that Astra scores higher on the same benchmark for roughly $20 in compute.¹ Brown also noted that while achieving gold-medal performance on the International Mathematical Olympiad in 2025 required an enormous expenditure of compute by OpenAI and Google DeepMind, by 2026 any user with a $20 per month ChatGPT subscription could run a system capable of the same standard.¹

The benchmark evaluator ARC Prize confirmed similar efficiency gains on ARC-AGI-3, a test measuring how agents infer hidden rules within novel, turn-based environments.² In a blog post written by Greg Kamradt, ARC Prize reported that Astra achieved a 62.7 percent score on the semi-private evaluation set for $26,098 using the organization's Standard harness, which requires models to manage their own explicit notes.²⁴ When evaluated with a Provider Adapter harness that preserves internal reasoning state between requests and uses conversation compaction, Astra achieved a 99.9 percent score on the same semi-private set for $18,817.²⁴ ARC Prize also measured action efficiency against a baseline of approximately 500 members of the general public, who were paid $115 per 90-minute session plus $5 per completed game. In the Provider Adapter harness, Astra used fewer moves than the median human player across 96.0 percent of levels, averaging 51.7 percent fewer actions per level.²

How did multiagent systems resolve the Navier-Stokes problem?

OpenAI deployed an internal next-generation model across groups of communicating agents to construct a mathematical proof showing that fluid dynamics equations can develop infinite speeds in finite time. The Navier-Stokes equations treat fluids as continuous media to describe the motion of liquids and gases, but whether smooth solutions can break down into mathematical singularities has remained an open question for roughly 90 years. According to OpenAI's research report, the company organized concurrent agents with access to code execution tools and web caches into collaborative clusters, with the Navier-Stokes effort involving on the order of 10,000 concurrent agents.³ The agents produced an analytical proof showing that an initially smooth fluid at rest can develop a singularity in finite time while maintaining finite total energy under a smooth external force.³ OpenAI reported that the agents resolved the problem approximately 88 hours after launch, sending 2.7 million messages and consuming roughly 130 billion output tokens, before GPT-6 Astra completed formal verification in the Lean proof assistant in 17 hours.

What limits remain for automated mathematical discovery?

The high benchmark scores and formal mathematical proofs remain confined to tightly bounded environments and specific formal problem statements. The ARC-AGI-3 test beds feature closed, deterministic mechanics calibrated to human cognitive primitives, which do not reflect the ambiguous goals or unconstrained physics of real-world environments.²⁵ Similarly, OpenAI's Navier-Stokes proof resolves statements C and D of the Millennium Prize formulation by applying a smooth external force, but it does not establish whether unforced fluid flow can break down in physical nature.³ OpenAI emphasized that its internal multiagent run required strict containment and millions of dollars in compute, framing the demonstration as an interim snapshot of capability progress rather than an all-purpose scientific discovery engine.¹

This piece was prepared from research publications and public records; the authors have not been interviewed.

References

This article is based on 7 sources, listed in the order they are cited.

1. 1 H https://x.com/polynoamial announcement · 8 Sep 2026 Astra Demonstrates Lower-Cost AI Reasoning on ARC-AGI and IMO Tasks See the source
2. 2 A arcprize.org OpenAI's GPT-6 Astra on ARC-AGI-3 | ARC Prize See the source
3. 3 O openai.com On the Navier–Stokes Millennium Prize Problem See the source
4. 4 A arcprize.org GPT-6 Astra - ARC-AGI Results See the source
5. 5 A arcprize.org ARC Prize - What is ARC-AGI? See the source
6. 6 A arcprize.org ARC Prize Blog See the source
7. 7 A arcprize.org ARC Prize See the source