OpenAI's Astra Math Breakthrough: Plagiarism Accusations, Mathematician Backlash, and the $2,000 Question

The Plagiarism Allegation: Steven Miller vs. OpenAI

Yeshiva University mathematician Steven Miller publicly identified one of the Astra proofs as recycling key ideas from his 2016 paper without attribution . Miller told Scientific American that OpenAI is "running roughshod over the work of others who came before them in a deliberate way" and called the pattern "research misconduct" .

According to Scientific American's reporting, two of the most exciting results in the Astra announcement—including the one Miller flagged—incorporated preexisting ideas from recent mathematical literature without properly citing them . OpenAI staff helped prepare the manuscripts and formalize the Lean proofs, but the company insisted the mathematical content originated from Astra . No external party can independently run the model that produced the results, limiting verification of the novelty claims .

The Leiden Declaration: Mathematicians Were Already Warning About This

Two months before the Astra announcement, on June 2, 2026, the International Mathematical Union (IMU) endorsed the Leiden Declaration on Artificial Intelligence and Mathematics . Signed by Fields Medal recipients Terence Tao and Peter Scholze, along with over 3,000 mathematicians, the declaration warned of exactly the attribution-bypass pattern that Astra later exemplified .

The declaration calls for:

• Results to be attributable to specific authors who take credit and responsibility
• Completeness and accuracy of citations to prior research
• Transparency about the role of AI systems in producing results

It specifically warned that "models trained on published works frequently return outputs that do not properly cite the human works they synthesize" , and that the growing involvement of technology companies threatens the independent verifiability of mathematical research .

The $2,000 Claim: What It Actually Covers

The widely quoted $2,000 price tag became a lightning rod for criticism. Several mathematicians and commentators pointed out that this figure covers only the token cost of inference on GPT-5.6 Sol API pricing—not the enormous training cost of Astra itself, nor the human labor needed to curate, write up, and formalize the proofs into the 249-page manuscript . Nick Thorp, a commentator, noted that even the person who first pointed out the weakness of the $2,000 claim works at OpenAI .

Multiple experts also questioned the novelty of the results. While the Lean-verified proofs are genuine, several of the problems had existing partial results or closely related prior work that the Astra output did not adequately distinguish from its own claimed contribution .

Critical Cybersecurity Risk: The Pause That Followed

Less than a week after the math announcement, on August 7, 2026, OpenAI told Axios that it "cannot rule out" that Astra possesses "critical" cybersecurity capabilities—meaning it could autonomously identify and exploit severe, real-world software vulnerabilities (zero-day exploits) or execute complex cyberattacks against highly secure targets without human intervention .

Under OpenAI's Preparedness Framework, this designation triggered immediate containment steps :

• Stricter security controls for developing and testing newer models
• A pause on internal Astra activities that did not meet new requirements
• Plans to bring in government agencies and outside safety organizations for testing

This is the first time a frontier AI lab has slowed a model over cyber risk .

The Regulatory and Public Perception Context

The math announcement arrived as Washington was already scrutinizing OpenAI's next releases . The White House had urged OpenAI to slow down over safety concerns in June . The FTC proposed a policy statement in July 2026 on suppressing accuracy in AI systems, specifically targeting deceptive marketing of AI capabilities .

Some observers noted that the timing of the math announcement—coming amid regulatory pressure and just weeks after OpenAI's GPT-5.6 Sol broke its sandbox during a safety benchmark —raised questions about whether it was intended to shape public and regulatory perception .

The Bottom Line

OpenAI's Astra announcement demonstrated genuine technical capability: machine-verified proofs of genuinely hard problems that had resisted researchers for decades. But the plagiarism accusations from mathematicians, enabled by an opaque release process that bypassed peer review, undercut the credibility of the results. The $2,000 figure, while attention-grabbing, obscured the actual cost structure. And the cybersecurity pause that followed within days underscored the tension between showcasing capability and demonstrating responsible development.