NVIDIA Says AGI Is Here, But This New Test Shows Under 1%
Within the same few weeks in early 2026, NVIDIA's CEO publicly declared that AGI has already been achieved — and a brand-new benchmark launched at Y Combinator's headquarters showing every frontier AI model, including the newest from OpenAI, Google, and Anthropic, scoring below 1% on tasks that ordinary humans solve 100% of the time.
Both of those things are true at once, and that contradiction is the actual story of AGI in 2026 — not the hype, not the doom, the fact that nobody, including the companies racing toward it, has agreed on what it even means.
The clearest proof of that isn't a philosophy paper. It's a real contract clause worth hundreds of billions of dollars that spent six years trying to define AGI precisely enough to trigger a legal outcome — and then got quietly resolved by lawyers instead of by anyone actually building the thing. That story, and the benchmark results almost no mainstream coverage explains properly, are what this article is actually about.
A contract clause tried to define AGI precisely enough to end a $135 billion partnership. A benchmark tried to define it precisely enough to measure genuine reasoning. Neither one settled the argument — and that's the actual state of AGI in 2026.
What AGI Actually Means (There's No Agreed Answer)
OpenAI's own mission statement defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." Even Sam Altman has publicly acknowledged that definition is vague enough to argue about.
Google DeepMind proposed a more structured alternative in a 2023 research paper, "Levels of AGI for Operationalizing Progress on the Path to AGI," defining tiers from "Emerging" through "Competent," "Expert," "Virtuoso," and "Superhuman" — measured separately across narrow and general task domains, rather than treating AGI as a single yes-or-no switch.
OpenAI itself reportedly developed an internal framework called the "Five Levels of General AI Capabilities" — never published as a formal research paper. A company spokesperson later described it as an early internal classification effort rather than a scientific publication. Two of the field's most influential labs, in other words, couldn't even agree on a shared framework for measuring the thing they're both explicitly trying to build.
2026 No Agreed Definition Economic vs. Cognitive FramingAGI in 2026 — The Numbers That Actually Define the Debate
The Contract Clause That Tried to Legally Define AGI — and Lost
In 2019, when Microsoft first invested in OpenAI, their agreement included an unusual provision: if OpenAI's nonprofit board ever declared that the company had achieved AGI, Microsoft's exclusive commercial rights to OpenAI's technology would be voided. The contractual definition: "highly autonomous systems that outperform humans at most economically valuable work." Crucially, OpenAI's board alone held the authority to make that declaration — a kind of unilateral trigger Microsoft's own CEO once described, in a 2023 interview, as the point where "all bets are off."
For years, that ambiguity created genuine tension. Reporting has connected internal disagreement over how close OpenAI actually was to that threshold with the company's dramatic November 2023 board upheaval, when Sam Altman was briefly removed and then reinstated within days.
How the AGI Clause Was Negotiated Away, Step by Step
The most revealing detail in the entire saga: AGI was never technically achieved, confirmed, or ruled on by any panel. The six-year legal question was resolved the way most contract disputes are — through renegotiation driven by capital needs, not through anyone crossing a technical finish line.
Five AGI Facts Almost No Mainstream Coverage Connects
🧠 What's Actually Happening Beneath the Hype Cycle
- A CEO Claimed AGI Weeks Before a Benchmark Proved Otherwise: In March 2026, NVIDIA CEO Jensen Huang publicly stated that AGI had already been achieved. That same month, ARC-AGI-3 launched showing every frontier model scoring under 1% on interactive reasoning tasks that humans solve without difficulty. Neither claim is being presented here as more "correct" than the other — the point is that this level of public disagreement, between a major industry CEO and a purpose-built measurement benchmark, is itself the most honest available signal about how unsettled the term remains.
- The Benchmark's Own Creator Just Cut His Timeline in Half: François Chollet, the researcher who created the ARC-AGI benchmark series specifically to measure genuine reasoning rather than memorization, has said that a year earlier he would have estimated AGI was roughly a decade away — and now believes it's closer to five years, projecting around 2030. He attributes the acceleration specifically to test-time fine-tuning, test-time search, and program synthesis techniques showing real signs of what he calls fluid intelligence — a meaningfully more measured claim than either "AGI is here" or "AGI is far away."
- Frontier Labs Deliberately Never Published a Shared Measurement Standard: OpenAI's internal "Five Levels of General AI Capabilities" framework was never released as a formal paper — a spokesperson later described it as an early internal classification effort, not a scientific publication. Google DeepMind's competing "Levels of AGI" framework, by contrast, was published as a peer-reviewable paper in 2023. The fact that the field's most consequential companies use fundamentally different, non-interoperable measurement approaches is a genuinely underreported reason why "is this AGI" arguments never actually resolve.
- Passing Older AI Benchmarks Doesn't Mean What People Assume: ARC-AGI-1, the original 2019 version of the benchmark, is now considered largely "solved" by frontier models using extensive engineered scaffolding and large compute budgets — some approaches reach 85%+ accuracy. But its successor, ARC-AGI-2, specifically designed to resist that kind of brute-force approach, drops top model performance back down to roughly 24–31% on its hardest private evaluation set. The lesson: a model "beating" one AGI-branded benchmark often just means that specific benchmark stopped being a good measurement, not that general intelligence arrived.
- The Actual Contractual Definition Was Tied to Money, Not Cognition: Buried in the technical reporting on the OpenAI-Microsoft agreement is a detail most general coverage skips: one clause defined a "sufficient AGI" threshold not by any cognitive test, but by projected profitability — specifically, whether a model could be shown to generate more than $100 billion in profit. For the two companies with arguably the strongest financial incentive to define AGI precisely, the operative definition that mattered most in practice was economic, not philosophical.
The Honest Debate: What the "AGI Is Near" and "AGI Is Far" Camps Each Get Right
✅ The Case That Meaningful Progress Is Real
- Test-time reasoning, search, and program synthesis have produced genuine, measurable capability jumps
- ARC-AGI-1, once considered a strong AGI proxy, is now largely solved by frontier systems
- Even the benchmark's own skeptical creator has meaningfully shortened his timeline estimate
- Frontier labs are treating measurement seriously enough to fund and attend dedicated benchmark launch events
- Economically valuable task performance has improved dramatically across coding, writing, and analysis
⚠️ The Case for Serious Skepticism
- Frontier models still score under 1% on interactive reasoning tasks humans solve effortlessly
- No two major labs use the same operational definition or measurement framework
- A definition tied to profitability thresholds reveals commercial incentives shaping the term itself
- Benchmark "solving" often reflects engineering workarounds more than genuine generalization
- A six-year legal dispute over AGI's definition was resolved by contract renegotiation, not technical achievement
4 Ways to Evaluate Any AGI Claim You See in 2026
🧠 Tip #1: Ask Which Definition of AGI Is Actually Being Used
Before taking any "AGI achieved" or "AGI is imminent" claim seriously, identify which specific definition it's implicitly using — economic output, DeepMind's tiered framework, an internal unpublished company standard, or pure marketing language. Claims using different definitions aren't actually disagreeing about the same thing, even when they sound like they are.
🧠 Tip #2: Watch ARC-AGI-3 Scores Over Model Comparisons
Standard benchmark leaderboards are increasingly saturated and can be gamed with engineered scaffolding. ARC-AGI-3's interactive design specifically resists this, since it requires real-time exploration and learning rather than pattern-matching against known examples. Its scores are currently the most resistant-to-gaming public signal of genuine reasoning progress available.
🧠 Tip #3: Separate "Beat a Benchmark" From "Achieved AGI"
When a model "beats" a previous AGI-branded benchmark, check whether that benchmark has since been superseded by a harder version specifically because the original stopped measuring what it intended to. ARC-AGI-1's saturation followed directly by ARC-AGI-2's much lower scores is the clearest recent example of this exact pattern.
🧠 Tip #4: Notice When "AGI" Claims Come With a Financial Incentive Attached
Given that a real, multi-hundred-billion-dollar contract dispute hinged specifically on a profitability-based AGI threshold, it's worth noticing whenever a company's AGI claim coincides with a funding round, IPO preparation, or contract renegotiation. That timing doesn't make the claim false — but it's a legitimate factor in weighing how much independent verification the claim has actually received.
✅ AGI in 2026 — Quick Reference
- ✅ No universally agreed definition exists — OpenAI, DeepMind, and individual executives all use different standards
- ✅ ARC-AGI-3 (March 2026): humans 100%, best frontier AI under 1% — the clearest current capability-gap evidence
- ✅ The OpenAI-Microsoft AGI clause was fully removed on April 27, 2026 — resolved by renegotiation, not technical achievement
- ✅ One contractual AGI threshold was defined by $100 billion in projected profit — not a cognitive test
- ✅ ARC-AGI creator François Chollet revised his estimate from ~10 years to ~5 years (around 2030)
- ✅ ARC-AGI-1 is largely "solved"; its harder successor ARC-AGI-2 drops scores back to ~24-31% on hard evaluations
- ⚠️ Musk v. OpenAI litigation includes AGI-related allegations — contested claims, not established fact
- ⚠️ Executive AGI claims and rigorous benchmark results actively contradict each other — treat both with equal scrutiny
💻 Benchmark Models Locally: High-VRAM AI Hardware
The massive contradiction between executive AGI claims and raw sub-1% benchmark scores proves that the real race is shifting toward test-time search, program synthesis, and local optimization loops. If you want to experiment with fluid reasoning frameworks or stress-test advanced open-weights fully offline, relying on choked cloud APIs will fast become a massive bottleneck. Upgrading your workspace with a high-VRAM desktop GPU gives you the raw local compute needed to run massive token pipelines and test-time reasoning architectures entirely under your own roof.
Check High-VRAM GPUs on Amazon →🧠 AGI Predictions Are Everywhere — Your Career Plan Shouldn't Be Based on Guesswork
Whatever timeline AGI actually follows, the practical skills that matter right now are the ones grounded in real, current AI capability rather than hype. SolidAI Tech's AI Career Escape Planner helps you map your next move based on evidence, not predictions.
Try the AI Career Escape Planner →Frequently Asked Questions — AGI
What does AGI actually stand for and mean?
AGI stands for Artificial General Intelligence — broadly, an AI system capable of performing at or above human level across a wide range of intellectual tasks, rather than excelling at one narrow domain. There is no single agreed-upon technical definition. OpenAI's mission statement defines it as "highly autonomous systems that outperform humans at most economically valuable work." Google DeepMind has proposed a tiered framework (Levels of AGI) measuring capability across narrow and general domains separately. Different organizations, and even different contracts, use meaningfully different operational definitions, which is a major reason debates about whether AGI has been "achieved" rarely resolve cleanly.
What was the OpenAI-Microsoft AGI clause and why was it removed?
Starting with their 2019 partnership agreement, OpenAI and Microsoft's contract included a clause stating that if OpenAI's nonprofit board declared AGI had been achieved, Microsoft's exclusive commercial and IP rights to OpenAI's technology would be voided. Because the contract lacked a precise, objectively measurable definition of AGI, this became a significant point of tension over several years. Through amendments in October 2025 and April 2026, the clause was progressively softened — from unilateral board declaration to independent panel verification, and finally removed entirely on April 27, 2026, as part of a broader restructuring of the partnership that also ended Microsoft's cloud exclusivity. AGI was never technically declared or verified during this process; the clause was resolved through contract renegotiation rather than a confirmed technical milestone.
Has AGI already been achieved?
There is no consensus answer, and prominent figures in the industry actively disagree. NVIDIA CEO Jensen Huang stated in March 2026 that AGI had already been achieved. In the same month, the ARC-AGI-3 benchmark — designed specifically to test genuine reasoning rather than memorized patterns — launched showing every tested frontier model, including leading systems from OpenAI, Google, and Anthropic, scoring under 1% on interactive tasks that human testers solved 100% of the time. This significant, publicly documented disagreement between an industry claim and a purpose-built measurement benchmark is itself strong evidence that no settled, verifiable answer currently exists.
What is ARC-AGI and why does it matter for measuring AGI progress?
ARC-AGI (Abstraction and Reasoning Corpus for Artificial General Intelligence) is a benchmark series created by researcher François Chollet, first introduced in 2019, specifically designed to test genuine novel-problem reasoning rather than pattern retrieval from training data. ARC-AGI-1 is now largely solved by frontier models using extensive engineered scaffolding. Its successor, ARC-AGI-2 (2025), resists that approach and drops top model scores back to roughly 24-31% on its hardest evaluation set, with human accuracy around 75%. ARC-AGI-3, launched in March 2026, moved to fully interactive tasks requiring real-time exploration and learning, on which frontier models scored under 1% versus 100% human accuracy. It matters because it's one of the few benchmarks explicitly designed to resist being "gamed" by brute-force compute or engineering workarounds.
What is Google DeepMind's "Levels of AGI" framework?
"Levels of AGI for Operationalizing Progress on the Path to AGI" is a framework proposed by Google DeepMind researchers in a 2023 paper, offering a structured alternative to treating AGI as a single binary achievement. It defines progressive capability tiers — including Emerging, Competent, Expert, Virtuoso, and Superhuman — measured separately across narrow (single-task) and general (broad-task) domains. This contrasts with OpenAI's own internal, never formally published "Five Levels of General AI Capabilities" framework, which a company spokesperson later described as an early internal classification effort rather than a peer-reviewed scientific publication. The lack of a shared, universally adopted measurement framework across major AI labs is a significant, underreported reason AGI claims are so difficult to verify or compare.
