OpenAI’s GPT‑6 Astra scored 99.9% on ARC‑AGI‑3 in OpenAI’s harness but 62.7% on the shared harness, while industry leaders called for slower development. The mixed results and caution show that claims AGI is already here remain contested, even as frontier models tackle complex reasoning tasks.
What Happened
OpenAI’s chief technology officer, Greg Brockman, told a closed press briefing about GPT‑6 Astra in September that “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.”
Within days, Dario Amodei, chief executive of Anthropic, published an essay titled “We must pace the frontier” urging the industry to slow its development pace. Sam Altman and Elon Musk quickly echoed the sentiment, signaling agreement with Amodei’s call for caution.
GPT‑6 Astra’s most visible shift is in user interaction. Earlier models required detailed step‑by‑step instructions; Astra is engineered for longer tasks where a user sets a goal and boundaries, allowing the model to maintain reasoning across requests.
On the ARC‑AGI‑3 benchmark, Astra achieved a score of 99.9 % in a harness built by OpenAI that preserves reasoning between requests. The benchmark normally pits AI agents against unfamiliar interactive games with no instructions, measuring how efficiently they learn the rules.
In contrast, the shared harness used by ARC Prize reported Astra’s best score at 62.7 %, while the average human scored only 48 % on the same games.
Independent evaluations differ. Epoch AI places GPT‑6 Astra at the frontier, alongside Claude Fable 5.1, Kimi K3, and GPT‑5.5 Pro. Artificial Analysis, however, did not rate it as radically ahead of its predecessor, GPT‑5.6 Sol.
Toby Walsh, an AI researcher at the University of New South Wales, described frontier models, including Astra, as “still today very jagged,” noting that even top models struggle with simple tasks.
On September 8 2026, OpenAI announced a claimed solution to the Navier‑Stokes problem, one of the seven Millennium Prize problems.
The proof came from an unreleased internal model, reportedly more capable than Astra, that ran roughly 10,000 AI agents in parallel for about 88 hours. Astra’s role was the formalisation and verification that followed, taking an additional 17 hours.
These figures illustrate the breadth of Astra’s capabilities and the scale of the internal model that produced the Navier‑Stokes proof.
Are we able to agree on what we mean by "AGI". I've been using this definition from OpenAI which I thought was relatively standard and ok:https://t.co/gmY4okWlON
AGI: "a highly autonomous system that outperforms humans at most economically valuable work"
For "most economically… https://t.co/dJj4y05wHn— Andrej Karpathy (@karpathy) September 13, 2024
What This Means For You
For developers building long‑form AI applications, Astra’s ability to retain reasoning across requests means you can design workflows that require fewer user prompts, potentially reducing development time and improving user experience.
If you are a compliance officer, the rapid pace of AGI claims and the industry’s call for slower development suggest you should review your organization’s AI governance frameworks to ensure they can adapt to sudden capability shifts.
Investors eyeing AI companies should note that while GPT‑6 Astra shows impressive benchmark scores, independent analyses are mixed. This uncertainty may affect valuation expectations for OpenAI and its competitors.
Why It Matters
The debate over whether AGI is already present reflects a broader tension between technological optimism and caution. Brockman’s assertion that “we’re there” contrasts with Amodei’s plea to pace progress, highlighting divergent views on safety and responsibility.
Benchmark results such as the 99.9 % score on ARC‑AGI‑3 demonstrate that AI agents can now navigate unfamiliar environments with minimal instruction, a step toward more autonomous systems.
Yet Walsh’s observation that models remain “very jagged” underscores that significant gaps persist, especially in handling simple tasks reliably.
The Navier‑Stokes proof claim, achieved by a model running 10,000 agents for 88 hours, signals that AI can tackle complex scientific problems at scale. However, the reliance on massive parallelism raises questions about energy consumption, hardware requirements, and the environmental footprint of future AGI systems.
Key Takeaway
- GPT‑6 Astra’s 99.9 % ARC‑AGI‑3 score shows unprecedented reasoning continuity across requests.
- Independent reviews diverge, with Epoch AI placing Astra at the frontier but Artificial Analysis seeing no radical leap.
- OpenAI’s claimed Navier‑Stokes proof involved 10,000 agents over 88 hours, highlighting the scale of resources needed for AGI breakthroughs.
- Industry leaders are split: Brockman believes AGI is present, while Amodei, Altman, and Musk call for a slower development pace.
Frequently Asked Questions
What is the ARC‑AGI‑3 benchmark?
The ARC‑AGI‑3 benchmark places an AI agent in unfamiliar interactive games with no instructions, measuring how efficiently it learns the rules.
How many agents were used in the Navier‑Stokes proof?
OpenAI reported that roughly 10,000 AI agents ran in parallel for about 88 hours to produce the proof.
Did OpenAI confirm the Navier‑Stokes solution?
The source states that OpenAI published a claimed solution on September 8 2026, but it does not confirm independent verification.
Leave a Reply