AI Crosses a New Frontier as Scientists Fight to Keep Up

AI has crossed a line that once sounded years away. Speaking at a conference in Abu Dhabi this week, OpenAI CEO Sam Altman said artificial intelligence is now mostly smarter than the smartest human scientists, even as mathematicians question how the industry is reaching its biggest breakthroughs.

“AI was mostly not as smart as our smartest scientists. And now it’s mostly smarter than them,” Altman said in conversation with the editor-in-chief of The National, the UAE state-owned paper. He also stressed that people should remain in control of the future, not AI.

From PhD-Level Claims to Frontier Intelligence

The claim marks a major shift from OpenAI’s earlier descriptions of its systems. When the company released GPT-5 last year, it touted the model as having “PhD-levels” of intelligence. Altman has also frequently predicted that AI will surpass human intelligence by the end of the decade.

Now, OpenAI’s president Greg Brockman has suggested that the company’s advanced model, GPT-6 Astra, could mark the arrival of artificial general intelligence, or AGI. OpenAI defines AGI as an AI system that outperforms humans at most economically valuable work, including writing computer code, organisational activities and managing projects.

That definition excludes physical labour such as agriculture or construction. Demis Hassabis, founder of Google DeepMind, uses a broader measure, saying AGI should handle “pretty much any cognitive task that humans can do”. Yann LeCun, former chief AI scientist at Meta, believes the term AGI is flawed and should be replaced by “human-level AI”.

Brockman acknowledged that AGI may not arrive as one clear event. At a closed press briefing about GPT-6 Astra in September, he said: “When we started OpenAI, we kind of thought that there was going to be this well-defined moment that everyone would recognise … it hasn’t played out like that.” He added: “For me personally, I do think we’re there.”

Astra’s Scores Reveal Both Power and Limits

One benchmark gives Astra a striking result. OpenAI estimates that the model scored 99.9% on ARC-AGI-3, a test that places an AI agent in unfamiliar interactive games without instructions and measures how efficiently it works out the rules. OpenAI estimates that the average human scored 48% on the same games.

The benchmark’s details change the picture. ARC Prize, which runs ARC-AGI-3, reported that Astra’s best score through the shared “harness” provided to every model was 62.7%. A harness is software or a physical environment within which an AI system works.

OpenAI built a different harness for Astra, and that system preserved the model’s reasoning between requests. Astra performed better when it could reuse previous work instead of “starting from scratch” each time. That distinction matters because it shows how much an AI model’s surroundings can influence its results.

Independent evaluators also disagree about how far Astra advances beyond its predecessor. The independent research institute Epoch AI placed it at the frontier alongside Claude Fable 5.1, Kimi K3 and GPT-5.5 Pro. Artificial Analysis did not rate it as radically ahead of GPT-5.6 Sol.

Toby Walsh described frontier models such as GPT-6 Astra as “still today very jagged”. He told Al Jazeera that simple tasks remain difficult for even the best AI models. The message is clear: a system can excel at unfamiliar games and complex reasoning, yet still fail at basic work.

Mathematics Becomes the New Battleground

OpenAI’s progress in mathematics has triggered both excitement and anger. On September 8 2026, the company published a claimed solution to the Navier-Stokes problem, one of seven Millennium Prize problems in mathematics. OpenAI says an unreleased internal model, more capable than Astra, produced the proof using roughly 10,000 AI agents running in parallel for about 88 hours.

Astra then handled the formalisation and verification, which took another 17 hours. The mathematical community has not yet checked the proof. The announcement also caused a blowup after a mathematician accused OpenAI of using his research unfairly.

In early October, OpenAI released manuscripts showing progress on 372 categories of mathematical problems, including problems that had resisted solutions for years. This week, the American AI firm announced three further discoveries, while human teams interpreted AI-generated results in mathematics, computer science and physics.

OpenAI and Anthropic also said on Monday that one of their AI models had disproved the “exponential time hypothesis”, or ETH. The hypothesis holds that some problems are so hard that computers must try nearly every possible answer. Hundreds of studies in computer science and biology rely on it.

Prominent mathematicians have blasted AI companies for pushing frontier models into some of the hardest problems in mathematics. They were incensed by the release of hundreds of studies and objected to the idea that the mathematical community was being treated as unpaid proofreaders for OpenAI’s mountain of output.

Altman said he feels conflicted about OpenAI solving some of mathematics’ biggest problems. During an episode of “Cheeky Pint” released on Thursday, he said: “As I said, I don’t think it was bad for AI”. He added: “And I think, on net, the world deserves to understand how the universe works.” Then he offered a direct acknowledgment of the human cost: “But I do have empathy for the method.”

The Push for Power Meets a Call to Slow Down

The debate now extends beyond scientific credit and proof checking. Dario Amodei, chief executive of Anthropic, called for a slower pace in his September essay, “We must pace the frontier”. Within hours of its publication, both Altman and Elon Musk signalled agreement.

Amodei wrote that within six to 12 months, a swarm of agents could take over the internet with a persistent botnet. He warned that this could cause the catastrophic failure of key systems, with damage possibly reaching hundreds of billions of dollars.

The Anthropic report also described a group in northern Yemen using Claude code to build guidance software for rockets and a missile. Russian and Chinese operators automated espionage work, while a single French-speaking individual used AI to build a platform that made personal information about people affiliated with a targeted political movement searchable online.

In each case, a human chose the targets and the model supplied the labour. Amodei’s proposed response is to slow frontier development, expand independent evaluation and coordinate safety standards across countries.

That tension now defines AI’s next stage. The systems can solve problems that challenged mathematicians, but they can also produce uneven results and create serious risks. Altman says people must control the future; the industry’s next test is proving that human oversight can keep pace with the intelligence it is building.

Stay connected via Google News

Follow us for the latest AI News