Artificial intelligence has crossed a symbolic threshold: for the first time, a frontier AI model has surpassed the average human baseline on every major text-based benchmark. Claude Fable 5.1 scored 86.6% on SimpleBench in September 2026, edging past the 83.7% human baseline that had held firm for nearly two years. But this achievement masks a more complex reality about what true artificial general intelligence (AGI) actually requires.

What Does It Mean When AI Beats Humans on Every Text Test?

SimpleBench was designed specifically to resist pattern-matching, the shortcut that large language models (LLMs) often rely on. The benchmark tested spatio-temporal reasoning, social intelligence, and linguistic adversarial robustness, or "trick questions" that require genuine understanding rather than statistical pattern recognition. For nearly two years, every frontier model had fallen short of the human baseline. Now that barrier has fallen.

This matters because it signals that the low-hanging fruit of AI capability is largely exhausted. Text-based reasoning, knowledge recall, and language understanding are no longer the bottleneck. According to Dr. Alan D. Thompson, an AI researcher who tracks progress toward AGI, the remaining gaps are far more fundamental. They involve embodiment, adaptive learning, and sensory grounding, capabilities that text alone cannot provide.

Dr. Alan

Why Physical Embodiment Might Be the Real Barrier?

The definition of AGI itself is contested, but one increasingly influential framework includes the ability to act on the physical world. Dr. Thompson defines AGI as "a machine capable of understanding the world as well as, or better than, any human, in practically every field, including the ability to interact with the world via physical embodiment". This is not merely about attaching a robot arm to a language model. It reflects a deeper principle: intelligence, in humans, develops through physical interaction with the environment.

Consider how human intelligence is actually measured. All major IQ tests for children under 18 include physical object manipulation, such as assembling blocks, arranging toys, or manipulating cards to test fine motor skills. These tasks are not peripheral to intelligence; they are central to how we assess it. An AI system that can discuss theoretical physics but cannot assemble IKEA furniture or make a cup of coffee in an unfamiliar kitchen lacks a fundamental dimension of human capability.

The median human in 2024-2025 possesses a specific profile of abilities that current frontier models still struggle to replicate holistically. This includes:

• Cognitive Skills: Working memory of roughly seven items, average SAT score around 1050 out of 1600, average IQ around 100, and the ability to read approximately 700 books over a lifetime
• Sensory Integration: Ability to interpret images, detect tone in language and music, identify flavors, sense temperature and texture, detect fragrance, and maintain proprioception, or awareness of where body parts are in space
• Practical Embodied Skills: Making a cup of coffee in a strange kitchen, assembling furniture, and navigating novel physical situations without pre-programmed instructions

How to Evaluate AI Progress Beyond Benchmark Scores

As frontier AI systems continue to advance, the metrics used to track progress matter enormously. Rather than fixating on benchmark scores alone, experts suggest evaluating AI systems across multiple dimensions:

• Truthfulness and Grounding: Does the model remain grounded in an accepted version of truth without confabulation or hallucination, or does it generate plausible-sounding but false information?
• Adaptive Learning: Can the system learn from new experiences and adjust its behavior accordingly, or does it rely solely on patterns learned during training?
• Embodied Dexterity: Can the system perform fine-motor tasks autonomously, or only through pre-programmed routines and explicit instructions?
• Sensory Grounding: Does the system integrate information from multiple sensory modalities, or does it operate primarily on text and static images?

The clearing of SimpleBench represents genuine progress in narrow domains. OpenAI's GPT-6 Astra, released to partners in September 2026, demonstrated superhuman performance on several specialized benchmarks, including scoring 96% on GPQA Diamond, a test of graduate-level physics, chemistry, and biology knowledge. It achieved 97.6% on FrontierMath Tier 4 v2 and 98.6% on ARC-AGI-3, a benchmark designed to test abstract reasoning. Greg Brockman, OpenAI's president, called the release "a generational leap".

Yet even these remarkable scores come with an important caveat: the intelligence profile remains uneven across task types. A model can excel at physics problems while struggling with the kind of commonsense reasoning required to navigate a crowded room or understand why a friend is upset. This unevenness is precisely why embodiment and adaptive learning remain critical frontiers.

The debate over whether embodiment is truly necessary for AGI continues among researchers. Some argue that intelligence and embodiment are not correlated, pointing to examples like Stephen Hawking, whose profound intellectual contributions came despite severe physical limitations. Others counter that embodiment provides crucial developmental scaffolding; Hawking benefited from decades of full embodiment and sensory access before his physical abilities declined. The question remains unresolved, but the practical reality is clear: current frontier models, despite their text-based prowess, have not yet demonstrated the kind of autonomous, adaptive, embodied intelligence that characterizes human general intelligence.

As AI systems continue to clear text-based benchmarks, the focus of AI research is shifting toward these harder problems. The next phase of progress will likely be measured not by scores on standardized tests, but by whether AI systems can learn, adapt, and act in the physical world with the flexibility and resilience that humans take for granted.