The question everyone is actually asking
When people search "is AGI here" in September 2026, they're usually asking one of three things: did OpenAI just cross the finish line? Is the AI I'm using already smarter than me in ways that matter? And should I be worried about what comes next?
All three questions deserve a straight answer. But they're different questions, and mixing them up is how you end up believing contradictory things at once — which is exactly the state of the current public debate.
To cut through it, you need to know which definition of AGI each data point is using. Once you have that, the apparently contradictory signals start to resolve.
What OpenAI actually claimed
On September 3, 2026, OpenAI launched GPT-6 Astra and President Greg Brockman said: "Welcome to the AGI era."
Note what that is and isn't. It's a personal statement, not a corporate declaration. OpenAI made no formal claim that Astra satisfies their AGI threshold. Brockman himself acknowledged: "There's no clearly defined AGI moment" and described the transition as "more gradual than expected." What he announced was an era — a period of transformation — not a threshold crossing.
"I think it's not unreasonable to feel that we are now in the AGI era. I do think we're there."
— Greg Brockman, OpenAI President, September 3, 2026
The framing is deliberate. An "AGI era" can mean AI that is broadly transformative and economically disruptive without requiring any specific technical threshold to have been crossed. OpenAI appears to be announcing the cultural and economic moment — the point where AI starts restructuring industries — without claiming the technical milestone. It's threading a needle.
OpenAI's official definition of AGI is "an automated system that can perform all economically valuable work as well as or better than humans." Whether Astra meets that bar depends entirely on what you count as "all economically valuable work" — and nobody has tested all of it.
What the 2026 benchmarks actually show
GPT-6 Astra's published benchmark scores are remarkable by any prior standard:
| Benchmark | Astra score | Previous SOTA | Human baseline |
|---|---|---|---|
| ARC-AGI-3 | 98.6% (OpenAI setup) / 62.7% (neutral) | ~30% | ~100% |
| FrontierMath Tier 4 v2 | 97.6% | ~38% | Expert-level |
| ExploitBench | 100% | ~50% | ~60% (experts) |
| GPQA Diamond | 96% | ~75% | ~65% |
| DeepSWE v1.1 | 74.1% | ~48% | ~85% |
| OSWorld 2.0 | 72.6% | ~35% | ~95% |
These numbers would have seemed implausible 24 months ago. But they come with caveats that matter. First, the ARC-AGI score gap is enormous: 98.6% under OpenAI's own experimental setup versus 62.7% under neutral testing conditions. That's not a rounding error; it's a 36-point gap that raises legitimate questions about test-time compute, scaffolding, and how reproducible these results are.
Second, and more fundamentally, the ARC Prize Foundation — the organization that runs ARC-AGI — explicitly stated that saturating their benchmark would not count as proof of AGI. Their benchmark tests a specific type of novel reasoning in closed, deterministic, narrow-format environments. It's a hard proxy, not a definition.
Three definitions, three answers
Here's the cleanest way to think about the "is AGI here?" question. There are three main definition frameworks in use, and each gives a different verdict on current systems:
| Definition | Core requirement | Astra verdict | Who uses it |
|---|---|---|---|
| Economic | Performs all economically valuable work ≥ humans | Contested | OpenAI, Jensen Huang |
| Efficiency (ARC) | Solves novel tasks as efficiently as humans | Closer, not there | ARC Prize Foundation |
| Cognitive (learning) | Continues learning after training; builds world models | No | Gary Marcus, Yann LeCun, academic AI |
The cognitive definition is the strictest and the most widely held in academic research. Its core argument: current AI systems, no matter how impressive their benchmark scores, are fixed after training. They cannot update their own weights from new experience. They cannot accumulate knowledge the way a human professional does over a career. Every interaction starts from the same base.
Gary Marcus put it plainly: current systems "simply cannot do the thing that every serious definition requires, which is to keep learning after training ends." Yann LeCun adds that language-only systems also lack grounded models of the physical and social world — they cannot learn from physical interaction, which limits their general intelligence in ways that no benchmark captures.
Astra performs at or above human expert level on math, coding, and cybersecurity benchmarks. It operates computers autonomously at human speed. It discovered two zero-day security vulnerabilities during testing. The economic disruption is already measurable. Greg Brockman, Jensen Huang, and Dario Amodei all indicate we're at or near the threshold.
No current system continues learning after training. Astra's ARC-AGI score drops 36 points under neutral conditions. The ARC Prize authors themselves reject benchmark saturation as AGI proof. Cognitive definitions — the most rigorous — require capabilities no current system has demonstrated. Gary Marcus: the current claims "simply muddy the waters."
What the forecasters say — and why the timeline compressed so fast
The most striking data point isn't where the median sits today. It's how fast it moved.
In 2020, the Metaculus community median for AGI was roughly 2070 — about 50 years away. In 2026, that same community puts the median at January 2033 — about 7 years away. That's a 40-year compression in six years. If you asked someone in 2020 to predict that shift, most would have called it wildly optimistic. It happened anyway.
The howcloseisagi.com forecast tracker aggregates Metaculus, AI Futures Model, Samotsvety, and other sources into a single running median — currently sitting at 2031 as the aggregate central estimate. It updates as predictions move.
The current forecast landscape as of September 2026:
- Metaculus community median: January 2033 (25% by 2029; requires cognitive + robotic capability — a stricter definition than most)
- AI Futures Model aggregate: ~2031 as central estimate, 80% confidence interval spanning 2027–2044
- Samotsvety superforecasters: ~28% chance by 2030 — a ~2-year compression from their 2023 forecast
- Kalshi prediction market: ~55% for OpenAI achieving AGI by 2030 (resolves on OpenAI declaration)
- Polymarket: ~9% for OpenAI AGI by 2027
- Dario Amodei (Anthropic): "AI broadly better than all humans at almost all things" by 2026–2027
- Demis Hassabis (DeepMind): ~50% by 2030
- Geoffrey Hinton: 5–20 years, with 10–20% extinction probability
The wide spread — from "basically now" (Amodei) to "maybe 2044" (Metaculus upper bound) — reflects genuine uncertainty, not confusion. AGI is not a single, agreed threshold. It's a cluster of capabilities that different people weight differently.
The thing that makes this moment genuinely unusual
Here's what's true regardless of which definition you use: the forecast moved from 2060 to 2031 in under a decade. We are in a period where AI systems are demonstrably replacing human judgment in domains — math, coding, legal research, clinical diagnosis — that were considered firmly human territory as recently as 2022.
Whether you call that "AGI" depends on your definition. Whether it changes things for your career, your industry, and the decisions you make in the next 5 years doesn't depend on the label at all.
If your work involves tasks that are procedural, predictable, and digitizable — the benchmarks suggest current AI is already at or past human performance on large portions of it. If your work requires continuous learning in novel physical environments, ongoing relationship-building, or embodied judgment — no current system is close. The profession-by-profession breakdown maps this more precisely.
What "AGI era" means for builders
For anyone building products and businesses right now, the most useful frame isn't "is AGI here?" — it's "what can I do with what exists, and how fast will the floor rise?"
The current floor — GPT-6 Astra, Claude, Gemini Ultra — is already high enough to automate significant portions of software development, content creation, legal research, and financial analysis. That floor is not stable; the trajectory of the last 24 months suggests it rises faster than most planning cycles account for.
The AGI era framing is actually useful here, even if it's imprecise: it marks the moment when AI capability stops being a niche advantage and becomes the baseline assumption for competitive products. That moment, under any reasonable definition, appears to have arrived in 2026. Whether we've technically crossed the AGI threshold is a question for philosophers and forecasters. Whether the game has changed is not.