What is AGI in AI? Definitions, levels and timelines in 2026

AGI stands for artificial general intelligence: an AI system that can match a capable adult across most thinking tasks and learn new ones as flexibly as a person does. As of late October 2026, no lab has publicly declared AGI and no independent body has verified one, partly because labs and researchers still define it differently. Honestly, I think that disagreement matters more than any single headline.

So the useful question is probably "which definition, measured how?" This explainer covers the main definitions, including Google DeepMind's levels of AGI, how AGI would be measured, where AI stands in 2026 and what named people and surveys say about timing, based on research papers and reporting checked in late October 2026. For a side-by-side comparison of narrow AI, AGI and superintelligence, see our companion piece on AI vs AGI.

The short answer

AGI is best seen as a threshold on two axes at once: breadth (how many kinds of task a system can handle) and depth (how well it handles them compared with people). Today's chatbots score high on breadth. They can write, code, summarise and analyse images. Depth is uneven: expert-level on some tasks, below an average person on others.

AGI would be the point where depth is good enough across most of that breadth, and where the system can pick up unfamiliar skills without being retrained for each one. Everything else in the debate is about where to draw that line, and I'd guess nobody will agree on it soon.

The term and where it came from

AGI means artificial general intelligence. "General" separates it from narrow systems such as a chess engine or spam filter, which each do one job. The phrase spread among researchers in the early 2000s as a label for the field's original goal: a machine with broad, human-like intelligence. Large language models blurred that old split; our explainer what are LLMs in AI shows how they work, and our guide to the types of AI sets out the wider categories.

Four working definitions of AGI

There's no single agreed AGI meaning in AI. Four definitions do most of the work in 2026. Each sets a different finish line.

1. OpenAI: economically valuable work

OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work". The definition has contract weight. Under the October 2025 Microsoft–OpenAI agreement, Microsoft says any AGI declaration by OpenAI "will now be verified by an independent expert panel".

2. Google DeepMind: levels of AGI

In November 2023, a Google DeepMind team led by Meredith Ringel Morris published "Levels of AGI for Operationalizing Progress on the Path to AGI", last revised in September 2025. Instead of a yes-or-no line, it borrows the idea of self-driving levels. Systems are rated on performance (how good) and generality (narrow or general).

The authors place OpenAI's labour-replacement threshold closer to "Exceptional AGI", and say "Competent AGI" is "probably the best catch-all for many existing definitions". Level 4 was originally called "Virtuoso"; the 2025 revision renamed it "Exceptional". It also adds a separate autonomy scale, from AI as a tool to AI as an independent agent.

3. Hendrycks et al.: a cognitive scorecard

The October 2025 paper "A Definition of AGI", from 33 authors including Dan Hendrycks, Yoshua Bengio and Gary Marcus, defines AGI as "an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult". It splits cognition into ten domains weighted at 10% each and scored GPT-4 at 27% and GPT-5 at 57%.

4. ARC Prize: learning efficiency

The ARC Prize Foundation, built around François Chollet's benchmark, defines AGI as "a system that can match the learning efficiency of humans". It argues that skill alone can be bought with more training data.

Measuring AGI

A definition is only useful if you can test against it. That's the weak point the DeepMind levels paper itself admits: it says a system that passes most of a future AGI benchmark at a given level can be treated as general, but it hesitates to name a pass mark. It expects "a very high percentage" that "will probably not be 100%".

Google DeepMind's 2026 cognitive framework

In 2026 DeepMind went a step further, which I think is a sensible move. Its paper "Measuring Progress Toward AGI: A Cognitive Framework" breaks general intelligence into ten cognitive faculties and proposes testing systems on "held-out" tasks to build a "cognitive profile". SingularityHub reports that the ten include perception, generation, learning, memory, reasoning, attention, metacognition and executive functions, plus problem solving and social cognition. Human baselines would come from a representative sample of adults with at least a high school education.

The catch, per the same report: there aren't reliable tests yet for metacognition, attention, learning and social cognition, and many public benchmarks may already have leaked into training data. DeepMind says it's working with academics on non-public tests.

ARC-AGI-3: learning on the fly

The ARC Prize launched ARC-AGI-3 on 24 March 2026. It drops AI agents into game-like environments with no instructions; they must explore, work out the goal and plan. AI News reported that at launch every frontier model scored below 1%, while humans solved 100% of the environments. Scores are weighted by how efficiently the agent acts compared with people, not just whether it finishes.

A worked example: scoring "most tasks"

Here's an illustrative profile we built to show how the levels framework works. It isn't real data for any model. Take a system tested on ten task families, with its percentile against skilled adults:

This system clears the bar on 6 of 10 families. Is that "most tasks"? Arithmetically, yes. Under the DeepMind authors' "very high percentage" expectation, no. And on a learning-efficiency definition like ARC's, the four failures are probably the ones that count most. Same system, three verdicts. To me that's the AGI debate in one table.

Real examples from 2026

As of late October 2026, the picture looks like this.

• ARC-AGI-3 scores jumped. According to LLMLearner's tracker of the official leaderboard (last result added 23 October 2026), OpenAI's GPT-6 Astra leads ARC Prize's standard setup at 62.7%, ahead of GPT-6.1 Sol at 52.7% and Gemini 4 Argon at 49.8%. Claude Opus 5 is listed at 30.2%. Astra's top run cost about $26,000 for the evaluation. Near-100% figures you may see come from a different setup that ARC Prize rates as having limited comparability.
• The prize is still open. ARC Prize 2026 offers a $700,000 grand prize for the first eligible agent to score 100% under competition rules, which require open-source code and no internet access.
• OpenAI says it's close but hasn't declared. Sam Altman told TIME, in a profile published on 26 August 2026, that OpenAI expects an internal system meeting his definition of AGI by the end of the year, but is "not quite yet" there, according to X-Risk Daily. GIGAZINE's summary links this to Astra, OpenAI's next flagship model.
• No public model is labelled AGI. The Arabian Post notes that OpenAI did not describe its GPT-5.6 family, released in July, or any other public model as AGI. It adds that OpenAI and Microsoft reaffirmed this year that the expert-panel process is unchanged.

On DeepMind's original scale, the 2023 paper placed frontier chatbots at Level 1, "Emerging AGI". Its authors have not published a new official rating for 2026 models, so I'd treat any claim that a model has reached "Competent AGI" as someone's interpretation.

Benefits and limits of AGI definitions

Clear definitions help: they give researchers a shared vocabulary, let regulators tie rules to capability levels and turn promises into testable claims. They also break in predictable places.

• Benchmark leakage: public tests end up in training data, so scores rise without the underlying ability rising.
• Setup and cost effects: the same model can score very differently depending on the testing setup and how much compute it is allowed.
• Jagged profiles: averages hide failures. The Hendrycks team found "critical deficits" in long-term memory storage even where knowledge scores were high.
• The body problem: some definitions, like Steve Wozniak's "coffee test" cited in the DeepMind paper, need a robot. Most current tests are text and screen only, which seems like a real gap.

Timelines from named people

Dates only make sense next to the definition each person uses. Our AI vs AGI page covers these views in more depth.

Academics are more sceptical about method. In the 2025 AAAI survey, 76% of 475 respondents said scaling current approaches was unlikely or very unlikely to produce AGI, The Decoder reported. The 80,000 Hours review concludes the forecasts "neither rule in nor rule out AGI arriving soon".

Reading an AGI claim

You don't need to settle the AGI debate to use AI well. When you see an AGI claim, run this checklist:

1. Which definition? Economic work, human cognitive breadth, or learning efficiency. They aren't interchangeable.
2. Which test, and who ran it? A lab's own evaluation isn't independent verification.
3. What setup and cost? Check the test setup, compute budget and whether the test is public.
4. What fails? Look for the weakest areas, such as memory, long-horizon planning and new-skill learning.
5. Does it change your work this month? Usually the practical question is what today's AI agents can do reliably, not whether they count as AGI.

For career planning, focus on tasks rather than labels. Our guide to jobs AI can't replace looks at where human skills still hold up.

Next steps to watch

Three things to watch into 2027. First, whether OpenAI formally declares AGI around Astra, and how the Microsoft–OpenAI expert panel rules. Second, whether DeepMind publishes cognitive profiles of real models using its non-public tests, which would give the first like-for-like scorecard. Third, whether ARC-AGI-3 scores keep climbing at lower cost, which would be stronger evidence of general learning than raw scores alone.

Until a claim survives independent testing against a stated definition, I'd call AGI a goal, not a product.

FAQ

What is AGI AI?

AGI AI refers to artificial general intelligence: a hypothetical system that matches a capable adult across most cognitive tasks and learns new skills without task-specific retraining. Today's models are broad but uneven. As of late October 2026, no lab has publicly declared AGI and no independent body has verified that any system meets an agreed definition.

What does AGI stand for in AI?

AGI stands for artificial general intelligence. "General" contrasts it with narrow AI, which does one job well, such as playing chess or filtering spam. Definitions vary: OpenAI ties AGI to outperforming humans at most economically valuable work, while the ARC Prize defines it as matching human learning efficiency.

What are the levels of AGI?

Google DeepMind's framework rates AI from Level 0 (no AI) through Emerging, Competent (50th percentile of skilled adults), Expert (90th), Exceptional (99th) and Superhuman. Each level can be narrow or general. In 2023 the authors placed ChatGPT and similar chatbots at Level 1, "Emerging AGI", and said Competent AGI had not been achieved.

Has anyone achieved AGI yet?

Not by any independently verified standard. Sam Altman told TIME in August 2026 that OpenAI expects an internal system meeting his definition by year end, but said it was "not quite yet" there. OpenAI's deal with Microsoft requires an independent expert panel to verify any AGI declaration.