Last week, two of the most powerful people in AI declared a new era of intelligence. And the interesting part was that neither of them agreed on what they were declaring.
On September 3, 2026, OpenAI shipped the GPT-6 Astra. The release page never uses the phrase “artificial general intelligence.” Then Greg Brockman, OpenAI’s president, posted welcome to the AGI era, and immediately tried running it as his “personal judgment” on a term he called “gray and ambiguous.”
Three days later, Nvidia’s CEO Jensen Huang skipped the hedging entirely; AGI has arrived. Congratulations to the OpenAI team. Sam Altman, weeks earlier, had called AGI a poorly defined term bordering on marketing jargon.
But none of them, including OpenAI, can point to a single agreed test that settles the artificial general intelligence debate. So, in this piece, we’ll get into what AGI is and some examples.
Key Takeaways
- AGI means a system that matches human-level performance across most cognitive tasks instead of one narrow skill.
- OpenAI never officially claimed AGI on GPT-6 Astra’s release page. The “AGI era” line came from Brockman’s personal post.
- GPT-6 Astra scored 99.9% on OpenAI’s own ARC-AGI-3 harness, but the ARC Prize Foundation puts the real number at 62.7% under standard conditions.
- OpenAI, DeepMind, and Anthropic don’t currently share a testable definition of AGI.
What Is Artificial General Intelligence?
Artificial general intelligence is a system that can learn, reason, and perform a wide range of cognitive tasks at or above human level, without being retrained for each new one.
IBM frames artificial general intelligence as a hypothetical stage in the development of machine learning in which an AI system can match or exceed the cognitive abilities of human beings across any task. The word hypothetical is very important in this case. It means this is still a category nobody has confirmed entering.
Three separate definitions are used interchangeably:
- Cognitive definition: Matches the versatility of a well-educated adult across roughly ten cognitive domains.
- Economic definition: A highly autonomous system that outperforms humans at most economically valuable work.
- Efficiency definition: A system that can acquire any new skill about as efficiently as a human can.
The efficiency definition is the perfect way to understand where the “is this AGI” argument is. A model can ace a hundred benchmarks and still need a thousand times more data than a curious 10-year-old to learn something genuinely new. That’s where AGI is right now.
Every current AI system, including GPT-6 Astra, is technically narrow AI. It’s just narrow AI with an unusually wide net. AlphaFold folds proteins, AlphaGo plays Go, and GPT-6 Astra writes code, drafts contracts, and holds a conversation, but each is still bound by what it was built and trained to do.
Artificial General Intelligence vs. AI: What Separates General From Narrow Intelligence
Every AI tool you use daily, your maps app, your spam filter, your coding assistant, is narrow. It’s excellent in its job, but useless outside it. For example, a chess engine that beats every grandmaster can’t get you a coffee. It’s built to fulfill a specific objective, nothing more than that.
Artificial general intelligence wouldn’t have a lane. It would move between domains the way a person moves. Like cooking dinner, reading a contract, comforting a friend, etc., using the same underlying intelligence.
Here’s the comparison:
| Dimension | Narrow AI (today) | AGI (hypothetical) |
| Scope | One task or a fixed cluster of tasks | Any cognitive task, unrestricted |
| Learning | Needs retraining or fine-tuning per task | Transfers knowledge across domains on its own |
| Autonomy | Executes within a defined script | Sets and pursues its own sub-goals |
| Current status | Exists everywhere, from spam filters to GPT-6 Astra | Not yet demonstrated under standardized testing |
| Real-world example | AlphaFold, autocomplete, GPT-6 Astra | None confirmed; disputed claims only |
GPT-6 Astra sits in a disputed middle ground. It goes well beyond single-task narrow AI, but still needs retraining, scaffolding, and a favorable testing harness to hit its best scores. That undefined middle ground is exactly why the “artificial general intelligence era” framing took off.
What OpenAI Means When It Says “Welcome to the AGI Era”
It implies way less than what the actual phrase means.
Here’s the sequence:
- September 3, 2026: OpenAI releases GPT-6 Astra, describing it as the most intelligent and most aligned model in the world. No mention of AGI appears on the release page.
- Same week: Greg Brockman posts “welcome to the AGI era,” but qualifies AGI as a “gray, fuzzy thing” and a “mission concept or spiritual concept,” not a contractual or technical threshold.
- September 6, 2026: Jensen Huang posts “AGI has arrived. Congratulations to the OpenAI team, in a longer note referencing 100,000 GPUs used in training and 400,000 more coming online.
Interestingly, the company that built the model hasn’t put “AGI achieved” in writing anywhere officially.
There’s a harder problem underneath the language, too. OpenAI reported a 99.9% score for GPT-6 Astra on ARC-AGI-3 using its own testing harness. The ARC Prize Foundation, which built and owns that benchmark, puts the real number under standardized conditions at 62.7%. That’s a huge gap.
The Capabilities That Would Define True Artificial General Intelligence
Here’s what separates a genuinely general system from a very good narrow one.
Reasoning and abstraction
A general system generalizes from a handful of examples the way humans do, instead of pattern-matching against billions of training examples it has already seen something similar to. Current models are remarkable at recombination, but they still struggle with problems structured in a truly novel way instead of as a variation on familiar training data.
Autonomy and long-horizon planning
This is a wall. Ask a model to execute a multi-step goal over hours or days without a human checking in; you can’t rely on that, right? But a true artificial general intelligence wouldn’t need a person re-prompting it every few steps. It would set its own intermediate goals, notice when it’s off track, and self-correct without a nudge.
Two more capabilities that are important:
- Self-correction under ambiguity: Noticing you’re wrong without being told, not only when explicitly asked to double-check.
- Tool use as instinct: Reaching for a calculator, a search, or an API because it’s the obviously right move, not because a developer built that path in advance.
AGI Examples: What It Would Actually Look Like
People ask for “AGI examples” expecting a product name. There isn’t one yet. What exists instead falls into two very different buckets.
Hypothetical artificial general intelligence capabilities
(None of this exists today)
- A single system that runs a full scientific experiment end to end: forms a hypothesis, designs the test, runs it in a lab, interprets ambiguous results, and revises the hypothesis, without a human scientist directing each step.
- An assistant that manages a small business’s finances, legal filings, and customer service simultaneously, transferring judgment learned in one area to a completely unrelated one.
- A system that learns a brand-new board game from a one-paragraph rulebook and plays competently within a few rounds, like the way a person would.
Near-AGI systems people mistake for the real thing:
- GPT-6 Astra itself handles code, writing, and conversation impressively but still needs task-specific prompting and guardrails.
- AlphaFold, a genuine scientific breakthrough that is, by design, narrow. But it does nothing except protein structure prediction.
- Autonomous coding agents that can complete a full pull request unsupervised but fall apart the moment the task shifts outside software.
How Close Are We to AGI?
The spread of expert predictions is the actual story here, more than any single forecast.
Six recent, named forecasts, lined up side by side, show the definitional problem is bigger than the technical one:
| Predictor | Predicted arrival | When stated |
| Sam Altman (OpenAI) | Internally targeting end of 2026 | August 2026 |
| Dario Amodei (Anthropic) | As early as 2026 | October 2024 |
| Elon Musk (xAI) | End of 2026 | January 2026 |
| Demis Hassabis (DeepMind) | Around 2030, plus or minus a year | June 2026 |
| Geoffrey Hinton | 2028 to 2043 | 2023, restated 2025 |
| Metaculus forecaster community | April 2033 (general AGI) | August 2026 |
Predictions compiled from public statements as tracked by FelloAI’s AGI timeline tracker.
That’s a 15-year spread among people building the actual technology. George Musser’s reporting in Scientific American traces the disagreement to the word itself, not to new data.
Google’s Blaise Agüera y Arcas argues artificial general intelligence already arrived when large language models first showed open-endedness. AGI pioneer Ben Goertzel tested Astra directly and concluded it executes faster but still lacks a creative spark. Joscha Bach goes further, arguing that real general intelligence would look nothing like current systems.
The field keeps conflating two separate measures; generality and intelligence level. It means how many domains a system covers, its intelligence level, and how well it performs within those domains. A system can be more general than any AI before it while still falling well short of human-level reasoning in most of those domains.
A 2024 survey of 2,778 AI researchers put a 50% probability on AGI arriving by 2047.
Why the AGI Debate Matters
Loose “AGI” claims move real money and real policy.
- Investment: Every “AGI is near” headline does work for the next funding round. But the valuations across the sector are partly priced on a threshold nobody has agreed on how to measure.
- Governance: Safety regulation frameworks in the US, EU, and elsewhere use AGI-adjacent language as trigger points for stricter oversight.
- Labor: Artificial general intelligence-achieved headlines can trigger hiring freezes well before knowing if the underlying capability actually works or not.
OpenAI’s choice of “AGI era” over a specific, falsifiable capability claim captures the excitement of a breakthrough while avoiding the accountability of a claim that could be checked against a defined bar. Deciding whether to change how you work, invest, or plan your career around what’s coming depends totally on spotting that difference between evidence and marketing.
Final Thoughts
OpenAI benefits from the “AGI era” line regardless of whether any real threshold was crossed. Nvidia benefits from a claim that justifies buying more GPUs.
GPT-6 Astra remains a genuinely impressive AI model. The announcement around it is the marketing artifact, built on a system that has moved past old-style narrow AI without reaching anything that meets a strict definition of general.
This piece doesn’t hand you a yes or no verdict, because nobody building the technology agrees on one either. A model that can run an unsupervised, multi-day project with no human correction would be convincing. But a press release wouldn’t.
For more info on tech and AI, visit Yaabot.
FAQs
Artificial general intelligence is an AI system that can learn and perform almost any cognitive task at a human level, without being retrained for each new one. No confirmed system has demonstrated this yet.
No. All AGI would be AI, but not all AI is artificial general intelligence. Today’s AI, including GPT-6 Astra, is narrow: strong within trained domains, unreliable outside them.
Not officially. OpenAI’s own release materials for GPT-6 Astra don’t claim AGI. The “AGI era” line came from a personal post by president Greg Brockman, not a company statement.
Estimates range from “already happened” to 2043, depending on the definition used. A 2024 survey of AI researchers put a 50% chance on AGI by 2047.
None exist yet. In the future, a scientific experiment might be conducted in an unsupervised manner, or a new game might be taught from a short rulebook, as examples. These are the narrowest systems available today, such as GPT-6 Astra and AlphaFold.
As it is, there is no agreed definition. Such a wide range of predictions (2026-2043) is due to each expert answering a slightly different question about what ‘general’ and ‘intelligent’ actually mean.
