Evaluating Recent Claims of AGI by NVIDIA CEO Jensen Huang and OpenAI President Greg Brockman

Jump to the comparison tablePress enquiries

In September 2026, following the release of GPT-6 Astra, NVIDIA CEO Jensen Huang stated that “AGI has arrived.” See Huang's statement ↗

His statement followed OpenAI President Greg Brockman's description of the release as the beginning of “the AGI era.” Brockman said that people may eventually look back on this period as when AGI was created, potentially with GPT-6 Astra, adding: “For me personally, I do think we're there.” See Brockman's remarks ↗ Neither statement specified a particular definition, test, or evidentiary threshold for AGI.

We therefore compare GPT-6 Astra's demonstrated capabilities against tests and frameworks that researchers have proposed for evaluating general intelligence.

Our current evaluation is that the published evidence does not yet support concluding that GPT-6 Astra represents the achievement Artificial General Intelligence.

Evidence available for GPT-6 Astra

ARC Prize independently evaluated Astra on ARC-AGI-3 ↗, an interactive benchmark in which agents explore unfamiliar environments, infer goals, construct models of how those environments work and plan actions without natural-language instructions.

Astra scored 62.7% using ARC Prize's standard provider-neutral harness and 99.9% using a provider adapter that preserves OpenAI's reasoning state between requests. ARC Prize also found that Astra used fewer actions than its median human baseline on 96% of levels.

ARC Prize describes its objective as measuring progress toward systems capable of acquiring new skills efficiently. It explicitly cautions that saturation of ARC-AGI-3 should not itself be interpreted as proof of AGI.

More broadly, GPT-6 Astra has been evaluated across reasoning, computer use, browsing, software engineering, cybersecurity, science and professional work.

OpenAI reports 97.6% on FrontierMath Tier 4, 72.6% on OSWorld 2.0, 41.4% on AutomationBench, 57.9% on Terminal-Bench 4.0 and 59.3% on Agents' Last Exam, among other results.

97.6%FrontierMath Tier 4 ↗
72.6%OSWorld 2.0 ↗
62.7 / 99.9%ARC-AGI-3 ↗
57.2%Humanity's Last Exam (with tools) ↗

These results provide evidence across several dimensions that have been proposed as components or indicators of general intelligence. Other proposed dimensions have not been directly evaluated on GPT-6 Astra.

OpenAI also disclosed a proposed solution to the Navier–Stokes existence and smoothness problem within days of these statements. According to OpenAI, the proposed proof was generated by an internal model more capable than GPT-6 Astra, while Astra was subsequently used to formalize and verify the result. The proof remains subject to independent scrutiny. Its proximity to the AGI claims makes it relevant evidence of the broader capabilities being developed at OpenAI, while its production by a different system illustrates an important distinction in evaluating the specific claim: capabilities demonstrated across multiple AI systems are not necessarily capabilities demonstrated by GPT-6 Astra.

Proposed tests of general intelligence

There is no generally accepted empirical test for AGI.

This problem predates current frontier models. A survey presented at the 13th International Conference on Artificial General Intelligence (AGI-20) reviewed proposals including the Imitation Game, Lovelace Test, psychometric testing, Piaget-MacGyver Room, Wozniak Coffee Test, Robot Student Test and Nilsson Employment Test.

Other approaches attempt to construct more general measures. Hernández-Orallo and Dowe's Anytime Universal Intelligence Test evaluates adaptive performance across procedurally generated interactive environments and is designed to accommodate agents operating at different levels and speeds of intelligence.

More recent evaluations include ARC-AGI, real-world agent benchmarks such as GAIA, measures of autonomous task duration, and frameworks such as DeepMind's Levels of AGI, which separates the breadth of a system's capabilities from its level of performance.

Longer-standing benchmarks also exist for domains outside text: general game learning through the Arcade Learning Environment and deep reinforcement learning on Atari, open-world robotic autonomy through Open X-Embodiment and the RT-X models, autonomous scientific discovery through work such as the Robot Scientist, and autonomous driving through Waymo's published safety research, including its crash-rate comparisons against human benchmarks.

The Marcus-Brundage 10-task test is also included. Gary Marcus referenced these criteria directly in his response to Huang's statement. Marcus and Miles Brundage proposed ten substantially different tasks and specified successful completion of at least eight as strong evidence for the “G” in AGI.

The table below compares GPT-6 Astra with this broader history of proposed tests and frameworks proposed so far for what would be considered “AGI.”

Reading the comparison

These tests and frameworks examine different aspects of intelligence. By evaluating major AGI claims consistently against proposed tests and the available evidence, AGI Society aims to identify where different measures converge, where they conflict, and what additional evidence is needed.

Table 1

Proposed tests of general intelligence vs. GPT-6 Astra

GPT-6 Astra tested directly
Earlier GPT/OpenAI model evidence
Estimate or interpretation
No comparable published evaluation

Proposed tests of general intelligence compared with GPT-6 Astra results
Test	What it evaluates	GPT-6 Astra results
Imitation Game1950Turing	Human-equivalent conversational behavior	Earlier GPT/OpenAI model evidenceEarlier GPT-family models have produced relevant conversational evidence; this is not a recorded test of Astra.
Lovelace Test2001Bringsjord, Bello & Ferrucci	Origination of artifacts beyond what the system's designers can account for through its design	Earlier GPT/OpenAI model evidenceEarlier GPT-family models have produced relevant evidence; this is not a recorded test of Astra.
Psychometric AI tests2003Bringsjord & Schimanski	Standardized human cognitive and intelligence tests	Earlier GPT/OpenAI model evidenceEarlier GPT-family models have been assessed on psychometric tests. Astra has related benchmark results, but no published assessment against a single defined psychometric AGI test was identified.
Autonomous driving2004DARPA Grand Challenge / Waymo safety research	Sustained real-time perception, prediction, planning and physical action in an open human environment	No comparable published evaluationNo published Astra demonstration comparable to sustained autonomous driving was identified.
Employment Test2005Nilsson	Perform economically valuable occupations ordinarily performed by humans	Estimate or interpretationEarlier GPT evidence and Astra results are relevant to parts of the test, but no published demonstration of Astra performing a complete occupation was identified.
Autonomous scientific discovery2009King et al., Robot Scientist	Generation of genuinely novel scientific or mathematical knowledge	Estimate or interpretationAstra was used to formalize and verify a proposed proof, but the reported discovery was produced by a different OpenAI system.
Anytime Universal Intelligence Test2010Hernández-Orallo & Dowe	Adaptive performance across procedurally generated unfamiliar environments	No comparable published evaluationNo published evaluation of Astra using this test was identified.
Wozniak Coffee Testc. 2010	Navigate an unfamiliar home, identify and manipulate the necessary objects and independently make coffee	No comparable published evaluationNo published Astra demonstration of independently completing this physical task was identified.
Robot Student / AGI Preschool tests2010Goertzel	Learn and function across ordinary educational environments requiring multiple abilities	No comparable published evaluationNo published Astra demonstration covering the complete test was identified.
Piaget-MacGyver Room2012	Physical problem solving using unfamiliar objects and affordances	No comparable published evaluationNo published Astra demonstration comparable to this physical test was identified.
General video-game learning2013Arcade Learning Environment (Atari)	Rapid acquisition of previously unfamiliar interactive tasks through experience	Estimate or interpretationAstra's ARC-AGI results provide related interactive-learning evidence, but no broad video-game learning evaluation was identified.
BEHAVIOR-1K2021Stanford	1,000 everyday household activities in simulation, requiring full-body navigation and manipulation grounded in real human needs	No comparable published evaluationNo published evaluation of Astra on BEHAVIOR-1K was identified.
GAIA2023	Real-world reasoning combining multimodality, browsing and tool use	Estimate or interpretationPublished Astra results show related multimodal, browsing, and tool-use capabilities, but no explicit Astra evaluation on GAIA was identified.
DeepMind Levels of AGI2023	Breadth of capabilities together with performance relative to humans	Estimate or interpretationAvailable Astra evidence is relevant to the framework, but no published assessment assigning Astra a level was identified.
Marcus-Brundage 10-task test2023	Breadth across ten tasks spanning media comprehension, games, coding, mathematics, science and creative work	Estimate or interpretationApproximately 2 of 10 tasks, estimated from available evidence; this was not a formal test administration.
Open-world robotic autonomy2023Open X-Embodiment / RT-X	Physical perception, navigation, manipulation, planning and adaptation	No comparable published evaluationNo published Astra demonstration comparable to this embodied robotics test was identified.
OSWorld 2.02024	Autonomous completion of tasks across real computer operating systems and applications	GPT-6 Astra tested directly72.6% reported for GPT-6 Astra.
Humanity's Last Exam2025	Expert-level academic knowledge and reasoning across disciplines	GPT-6 Astra tested directly57.2% with tools reported for GPT-6 Astra.
METR Task-Completion Time Horizon2025	Duration of autonomous tasks completed at specified reliability relative to human task duration	Earlier GPT/OpenAI model evidenceEarlier GPT-family models have been evaluated using this measure; no published Astra result was identified.
ARC-AGI-32026	Skill acquisition through exploration, modeling, goal-setting and planning in unfamiliar environments	GPT-6 Astra tested directly62.7% using the standard harness; 99.9% using the provider adapter.

Evidence from earlier GPT or OpenAI models is included for context and must not be read as direct testing of GPT-6 Astra. Estimates and interpretations are AGI Society assessments based on the cited, publicly available evidence.

In most formal mathematical models of AGI, the concept of AGI is a spectrum, with different systems displaying different levels of generality, so that a human might have more general intelligence than a dog, which in turn has more general intelligence than an insect, and so on. However, in popular parlance “AGI” is now often used to mean roughly “human-level AGI”, and when responding to a claim of having “achieved AGI” we will here use the term “AGI” in the same implicitly human-level-oriented sense as the claim itself. Our definition of AGI is set out in the Constitution of the AGI Society.

Editorial note

Update — September 14, 2026: Following discussion and feedback on this assessment, we have clarified AGI Society's current position and detailed how evidence types in the comparison should be interpreted.

AGI Society's Ambition is to be the community people rely on to understand the future of AI. This evaluation serves that ambition, as part of our commitment to Scientific Inquiry and fulfills our goals to Measure and Inform the world about AGI. We measure claims about AGI and AI systems against evidence, then inform the public about what that evidence shows.

We strive to be independent and unbiased and do not speak for any company, government, or research lab, working hard to keep the most up-to-date information about AGI claims and AI system capabilities available.

We are a small team and may miss things. If any data used above is incorrect, please contact us at outreach@agi-society.org for corrections or updates.

Sources

Evidence available for GPT-6 Astra

Proposed tests of general intelligence

Proposed tests of general intelligence vs. GPT-6 Astra

References