Is GPT-6 Astra really AGI or are we just seeing corporate marketing?
Nvidia says AGI is here with GPT-6 Astra, but behind the scenes, their $100B deal with OpenAI remains unsigned as chip buying slows. Read the real story.
Key Takeaways
What: GPT-6 Astra’s launch sparked corporate claims that AGI has arrived [1, 2].
Why: While Nvidia markets infinite hardware scale, its tentative $100 billion deal with OpenAI remains unsigned [3].
How: OpenAI is scaling back chip procurement while deploying automated AI interns to absorb high-cost, routine engineering tasks [3-5].
The Unsigned Alliance: Inside the Nvidia-OpenAI Hardware Friction
When GPT-6 Astra debuted, Nvidia CEO Jensen Huang immediately took to social media to proclaim that the race to artificial general intelligence (AGI) had ended. He pointed out that Astra was trained on Nvidia hardware and announced that another 400,000 GPUs would soon ship to OpenAI. Underneath the celebratory posts, however, lies a very different transactional reality.
The massive $100 billion partnership announced by Nvidia and OpenAI in 2025 was never actually signed. In fact, reports from June 2026 indicate that OpenAI has quietly scaled back its purchases of Nvidia chips. While Huang relies on OpenAI’s scale to validate Nvidia’s massive hardware investments, OpenAI appears to be quietly easing its reliance on Nvidia hardware. This friction contradicts the industry assumption of an inseparable, ever-expanding partnership. Huang’s earnings calls assert that “compute is revenue,” yet his biggest customer is quietly pulling back on the procurement reins.
Real-World Utility: GPT-6 Astra vs. Fable 5.1
Outside of hardware disputes, the performance of the software itself raises questions about whether AGI is truly here. Huang claimed Astra crossed a historic threshold, but artificial intelligence experts and everyday users are not seeing an undisputed leap. For instance, practitioners using GPT-6 Astra alongside competing systems like Fable 5.1 report that the new model is not meaningfully better for their daily workloads. Instead of a singular leap, users are finding the most value in running both systems side-by-side as complementary tools.
Astra boasts impressive scores on specific tests, such as 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3. But these benchmark victories do not translate directly to human-level adaptability. François Chollet, the creator of the ARC benchmark, noted that satisfying ARC-AGI-3 is not proof of AGI. He emphasized that a high score on a test is merely a benchmark result, not a guarantee that a machine can reason through unfamiliar, real-world problems.
The Economics of the Automated AI Research Intern
While the macro-level debate continues, OpenAI has integrated a new agentic system into its operations. In September 2026, the company deployed its “Automated AI Research Intern,” fulfilling a timeline proposed by Sam Altman a year prior. This system is designed to independently execute labor-intensive, tedious research tasks once human researchers provide a direction.
The deployment of these interns has transformed the daily rhythm of OpenAI’s research division. By mid-August 2026, the total running time of these agents surpassed human working hours. For every human researcher at a workstation, the company now runs 3.1 AI working days simultaneously. This is driven by concurrency, as over 70% of researchers now manage four or more agents at the same time.
However, running these systems is incredibly expensive. The median daily inference cost for a single OpenAI researcher using these agents reached $600 by mid-August. For the top 10% of researchers, the daily cost exceeds $7,000 per person in inference fees alone. While AI agents accelerate output, they consume capital at an extraordinary rate.
The Technical Boundary: Where Automation Cedes to Human Control
The 3.1-fold increase in daily workloads is not evenly distributed across research activities. OpenAI analyzed the specific tokens generated by its agents across six phases of research: decision-making, design, construction, operation, analysis, and communication. The analysis showed that agents are highly effective at structured, labor-intensive tasks. Tokens consumed for writing research code, technical support, code review, and launching training tasks skyrocketed, with researchers consuming over 130,000 additional tokens daily in these categories.
The impact on human coordination is clear. OpenAI’s active internal help channel, which averaged around 20 new posts a day in 2025, dropped to single digits by August 2026 because researchers now ask agents to debug their systems. Some team-based office hours were even disbanded because attendance dropped so low.
Yet, the agents hit a hard ceiling when it comes to high-level strategy. In categories like deciding resource allocation, reviewing external research, and announcing decisions, agent participation remains extremely low, averaging fewer than 1,000 additional tokens daily. The AI intern can manage the infrastructure, but humans still make all strategic decisions.
The Definition Deficit: Why Corporate Declarations Fail to Create Consensus
The fundamental issue with claiming AGI has arrived is the complete lack of scientific consensus. Gary Marcus, an NYU professor and researcher, criticized Huang’s declaration as an attempt to take over a scientific question by corporate fiat. Marcus pointed out that Astra still falls short under conventional frameworks and warned that declaring a victory without clear, agreed-upon definitions only muddies the scientific waters.
OpenAI’s own leadership remains hesitant to join Huang in his declaration. While Nvidia’s business model benefits when the market believes AGI has arrived, OpenAI President Greg Brockman stopped short of making the claim, stating only that the company is moving into the AGI era. Real scientific progress requires rigorous, standardized, and open metrics, not corporate marketing campaigns. Until a machine can truly reason across unfamiliar situations without a human guide, corporate declarations of AGI remain premature.