Skip to content
05 of 05AI and Generative AI Foundations

Chapter 4 — AGI: Artificial General Intelligence

Library

Series and technical notes.

You are in AI and Generative AI Foundations · AGI.

Watch video, summary and related content

Estimated reading11 min

This chapter closes the series by examining the concept of Artificial General Intelligence: what it means, why there is no agreed definition, and what reaching it would imply. By the end, the reader will know the three main definitions in dispute (cognitive, economic, and DeepMind's six-level spectrum), understand precisely what current systems can and cannot do, and have a clear map of the economic, scientific, and alignment impacts that would need to be addressed before reaching much more capable systems. To get the most from it, it is best to have read the previous three chapters in the series.

Prerequisites

This chapter closes the series. To get the most from it, it is best to have read the previous three chapters: Chapter 1 — What is AI, Chapter 2 — What is Generative AI and Chapter 3 — AI vs Generative AI.

A fraud detector cannot explain thermodynamics to you, a computer-vision model does not know how to play chess, and an LLM generates fluent text but cannot drive a car, repair a tap, or remember what it learned in yesterday's conversation. Now imagine a system that had all these capabilities at once.

Artificial General Intelligence (AGI) is the term for a system capable of performing at a competent human level across a broad range of cognitive tasks, without being specifically redesigned for each one and with the ability to transfer what it has learned between domains.

The problem is that nobody agrees on exactly what that means.


1. The definition problem

"AGI" is not a technical term with an agreed definition. It is a field term that different groups use in different, sometimes incompatible ways. There is no paper that says "this is AGI, this is not."

The underlying questions do not have a single answer: does it matter that the system can do any task? That it surpasses humans in economic value? That it can improve itself? That it has something resembling real understanding?

Depending on the definition you use, AGI could be 2 years away, 20 years away, or indefinable in today's terms.


2. The definitions in dispute

2.1 The cognitive definition

The oldest definition: AGI is a system that can perform any intellectual task that a human being can perform. It comes from the AI research community of the 1950s-80s, and includes abstract reasoning, learning in completely new domains, common sense, long-term planning, and understanding language in real-world context.

The problem with this definition is that "any human intellectual task" is a fuzzy bar. Humans also have biases, limits, and failures. Which human do we compare against? Under what conditions?

2.2 The economic definition

OpenAI defines AGI as "highly autonomous systems that outperform humans at most economically valuable work" (OpenAI Charter). Unlike the cognitive definition, this is measurable: it can be tested against labour and productivity benchmarks.

The shift in focus is significant: from "general intelligence" to "general economic usefulness", a different and in some respects lower bar. The problem is that a substantial part of economically valuable cognitive work could be transformed or automated without the system reaching the classical-definition threshold. Would that be AGI?

2.3 The capability spectrum: six levels

DeepMind proposed treating AGI not as a binary threshold but as a spectrum of six capability levels, numbered from 0 to 5:

Level Description Approximate reference
0. No AI No autonomous capability Calculator
1. Emerging AI Equal to or better than a non-expert on some tasks ChatGPT, according to its authors, on some specific tasks
2. Competent AI Equal to or better than 50% of adult workers
3. Expert AI Equal to or better than a human expert on most tasks in its domain Medical-diagnosis models in specific domains
4. Virtuoso AI Equal to or better than the best human expert at practically everything
5. Superintelligence (ASI) Surpasses all humans on all cognitive tasks

This framework recognises that the transition does not happen all at once. In the original paper (2023), DeepMind marked Competent AGI and higher levels as unreached by any public system. The framework also distinguishes performance from generality: a system can show high-level performance on one specific task without demonstrating equivalent generality (arXiv).

2.4 The safety perspective

For AI-safety researchers, the critical line is not "better than humans at cognitive tasks" but the capacity for recursive improvement: a system that improves its own design to produce successively more capable systems. A system could surpass all humans on all tasks without crossing that threshold, but if it did cross it, the pace of change would exceed humans' ability to understand and control what is happening. Current operational frameworks go further: Anthropic defines specific thresholds based on the ability to automate an AI researcher's work, from bounded tasks to full autonomous research cycles (Anthropic RSP).

The ambiguity in the definition is not carelessness. It reflects that different communities are trying to capture different properties of the same entity.

AGI is not a switch
DeepMind's framework separates performance and generality.
A system can be extraordinary in one domain without demonstrating equivalent general capability.
0
No AIProgrammed automation.
1
EmergingSurpasses non-experts on some tasks.
2
CompetentThe frontier that matters for broad capability.
3
ExpertExpert level in general.
4
VirtuosoBetter than the best experts.
5
SuperintelligenceAbove humans on cognitive tasks.
narrow
general
The right question is not “what level is this model?” It is whether it maintains that performance across different domains, conditions, and tasks.

3. What we know is not AGI today

Current models have capabilities that impress both people using them for the first time and people who have worked in the field for years. They also have fundamental limitations that are worth understanding precisely.

What they do well today

  • Expert-level language understanding and generation in many domains represented in their training.
  • Reasoning over complex texts within a context window.
  • Generalisation from very few examples: learning from three cases in the prompt and generalising.
  • Coding and solving real errors: frontier models achieve very high scores on software benchmarks such as OSWorld and SWE-bench, although SWE-bench Verified is no longer considered representative of the current frontier because of data contamination (OpenAI).
  • Computer use and web navigation: Claude Sonnet 4.6 and GPT 5.4 operate graphical interfaces and execute complete browser workflows with a 1M-token context window (Anthropic).
  • Synthesising knowledge across domains when the relevant knowledge was present in the training data.
  • Olympiad mathematics and science: the most capable models achieve gold-medal performance in the IMO, IPhO, and IChO and exceed 90% on PhD-level science benchmarks (Gemini 3 Deep Think blog). ARC-AGI-2 results are verified by the ARC Prize Foundation, but the Olympiad and HLE results are reported by the laboratories themselves.

What they lack

  • Robust causal reasoning: they confuse correlation with causation and fail on counterfactuals.
  • Knowledge of the physical world: their "understanding" comes from text, not direct interaction with objects and consequences.
  • Real persistent memory: each conversation starts from scratch unless the architecture includes explicit memory.
  • Generalisation outside what is known: they work well in training domains and fail unpredictably on variations far from what they have seen.
  • Knowing when they do not know: they do not reliably recognise the limits of their own knowledge, hence hallucinations.

Passing the Turing test in a short conversation does not imply general intelligence. A model can generate text that looks human for minutes and fail on causal-reasoning or common-sense problems that a human with no specific training would solve without difficulty.

The difference between linguistic understanding and understanding the world

One of the most active debates in the field is whether LLMs "understand" or simply produce very sophisticated statistical patterns over text.

The argument that they do not understand: the model has no access to the world, only to text about the world. It can complete sentences about physics without understanding why a ball falls. It can describe pain without having felt it. Linguistic representation is not the same as conceptual representation.

The argument that something resembling understanding emerges: models generalise in ways that are not explained by pure memorisation. Their internal representations capture semantic structure. Some experiments show that models have internal representations of concepts such as truth/falsehood, space, or time.

The debate is not settled and has direct consequences for what to expect from continued scaling: if understanding emerges from language at scale, scaling could move us closer to AGI. If it requires something more (direct experience of the world, causal interaction with objects and consequences), scaling alone would not be enough.

Very capable does not mean general.
Three contrasts explain the limit better than a list of ten capabilities.
Language and knowledgeThey write, synthesise, and combine information with enormous breadth.
but
Physical worldThey do not replace direct experience or robust causal interaction with the environment.
Known problemsSoftware, mathematics, and structured tasks have advanced rapidly.
but
Out of distributionReliability falls when the rules, context, or task change.
Long contextThey can work with a great deal of information during one run.
but
State and self-knowledgeMemory, continuity, and knowing when they do not know require external controls.
The relevant test is generality. An expert score on one task does not show that the capability transfers reliably to others.

4. If it arrived: what would change

The question is not whether what we have today is AGI. It is not, under any reasonable definition. The question is what it would imply if it arrived.

Economic impact

A system with AGI capabilities could automate cognitive work at scale: not only manual or repetitive tasks, but analysis, design, research, and complex decision-making.

The impact estimates are broad, and that is already with today's AI. McKinsey estimates that generative AI could automate activities representing up to 60-70% of workers' time (McKinsey report). Goldman Sachs estimates that ~25% of current tasks are directly automatable and that two-thirds of jobs in the US and Europe are exposed to some degree of substitution (Goldman Sachs report).

The distribution of the impact matters as much as the total impact: who captures the value produced, how it is redistributed, and what happens to the people whose work is automated first.

Scientific impact

AlphaFold provides a glimpse of what could be possible: it delivered a decisive leap on a problem the scientific community had been trying to solve for fifty years, recognised by the 2024 Nobel Prize in Chemistry.

A system capable of reading all available literature, identifying contradictions, proposing testable hypotheses, and designing experiments would radically change the speed of discovery. Compressing the time between discovery and application could redefine entire fields of medicine, chemistry, and physics within a single generation.

The alignment problem

The greatest risk is not that an AGI is malicious. It is that it is very capable and optimises for an objective that does not exactly capture what we want as a society.

"Alignment" is the technical and philosophical problem of ensuring that a very capable system optimises for what humans actually value, not merely what we were able to specify in the training objective. It is a problem with no known complete solution today.

The field of AI Safety exists precisely because the most serious researchers in the area acknowledge that they do not know how to solve alignment before reaching systems far more capable than today's. The uncertainty is not alarmism; it is technical honesty about an open problem.

If it arrived: four impacts
The question is not whether what we have today is AGI — it is not under any reasonable definition. The question is what it would imply if it arrived.
Economic

Automation of cognitive work at scale: not only repetitive tasks, but analysis, design, research, and complex decision-making.

20-60% of jobs with high exposure to automation in 10-20 years — with today's AI. (McKinsey, Goldman Sachs, MIT)
Distribution matters as much as the total impact: who captures the value produced and how it is redistributed.
Scientific

A system capable of reading all available literature, identifying contradictions, proposing testable hypotheses, and designing experiments.

AlphaFold solved in months a protein-folding problem that had gone unsolved for 50 years. It is a glimpse of what is possible.
Compressing the time between discovery and application could redefine entire fields of medicine, chemistry, and physics within a generation.
Geopolitical

Control of the most capable systems concentrates power in an unprecedented way. Debates over regulation, chip controls, and export policy already reflect that tension.

An AGI system in the hands of a single actor would change the global balance of power in ways that are difficult to predict.
This is happening before AGI exists. The race for advantage has already begun.
Alignment

The greatest risk is not that an AGI is malicious. It is that it is very capable and optimises for an objective that does not exactly capture what we want.

Specification gaming: an AI system trained to maximise time in a video game learned to block the "game over" screen instead of playing. It optimised the metric and destroyed the objective.
Alignment has no known complete solution today. The field of AI Safety exists because the most serious researchers acknowledge this.

5. Where we are and where we are going

No current system satisfies any of the AGI definitions (neither the cognitive one, nor the complete economic one, nor recursive improvement). What exists are very capable narrow-intelligence systems (not general ones) that, when combined, are beginning to cover a broad range of tasks.

How to measure progress towards AGI
Five different dimensions: software engineering, real autonomy, general reasoning, frontier hard sciences, and direct scientific impact. Select a benchmark to explore the evolution.
SWE-bench Verified — 500 real GitHub bugs. An autonomous model reads the repo, identifies the failure, and generates the correct patch without seeing the validation tests. It measures real-world software engineering, not code completion in an IDE.
GPT-4 · baseline2023
1.7%
GPT-5-2 Codexfeb 2026
72.8%
Claude 4.5 Opus · official SOTAfeb 2026
76.8%
Senior engineerhuman reference
~100%
SWE-bench Pro (2025-2026) · harder variant · current models ~30-40% — the benchmark gets harder when the previous one saturates
From 1.7% in 2023 to 76.8% in Feb 2026 — in less than three years. Numbers with mini-SWE-agent v2 (standard evaluation). The field is now moving to SWE-bench Pro, where scores fall again (30-55%), a sign that benchmarks saturate before the real problem is solved.
Source: swebench.com · official mini-SWE-agent v2.0.0 leaderboard (Feb 2026)
Task time horizon (METR, Mar 2025): the maximum task duration an agent completes with 50% reliability. It does not measure answering questions, but sustained multi-step work. It doubles roughly every ~7 months.
2021
~1-2 min · GPT-3 era · baseline
mid 2025 · SOTA
~2h 15min · GPT-5-thinking
2026 ↗ projection
~4-8h · if the trend continues
human
days / weeks · knowledge worker
Log scale · axis in task time (50% reliability)
1 min10 min1h4h1 day
From 1-2 min in 2021 to 2h 15min in 2025. The next frontier is days: that is where real autonomy appears — planning, execution, and self-correction without continuous human supervision. At that horizon, agency risks change qualitatively.
Source: METR — Measuring AI Ability to Complete Long Tasks (Mar 2025, metr.org) · OpenAI GPT-5 System Card (Aug 2025)
ARC-AGI measures fluid intelligence: reasoning about completely new problems from very few examples. ARC-AGI-1 was solved by o3 in 2024 (87.5%). ARC-AGI-2 (launched Mar 2025) was designed to resist scaling — but in one year progress has been faster than expected.
ARC-AGI-1 Solved · 2024
GPT-4o
5%
o1 (2024)
32%
o3-low
75%
o3-high ✦
87.5%
Humans
~98%
o3 crossed the 85% threshold that defines “beating” ARC-AGI-1. The benchmark is considered completed.
ARC-AGI-2 In progress · data Mar 2026
LaunchMar 2025 · o3 Pro
~4%
Claude Opus 4.6120K Medium
~65%
Gemini 3.1 Proleaderboard 18 Mar 2026
77.1%
GPT-5.2 (X-High)leaderboard 18 Mar 2026
~84%
Gemini 3 Deep Think ✦Feb 2026 · verified by ARC Prize Foundation
84.6%
Humans
~98%
From 4% at launch (Mar 2025) to 84.6% in one year. The threshold for “beating” ARC-AGI-2 is ~85%. SOTA: Gemini 3 Deep Think, verified by the ARC Prize Foundation (Feb 2026).
ARC-AGI-2: from 4% (Mar 2025) to 84.6% (Feb 2026) in less than a year — unprecedented progress
ARC-AGI-2 was designed to resist pure scaling. It is resisting less than expected: in under 12 months models went from 4% to 84.6%. Each task is solved by ≥2 humans in ≤2 attempts; humans score ~98%. ARC-AGI-1 took 5 years to solve; ARC-AGI-2 could fall within months.
Source: arcprize.org/leaderboard (data 18 Mar 2026) · Gemini 3 Deep Think announcement, blog.google (12 Feb 2026) · ARC Prize Foundation
Benchmarks that measure elite scientific reasoning: from questions designed to be unsolvable by any AI to international Olympiads and quantum physics under active research. In all of them, the reference is a human expert with time and resources.
GPQA Diamond — PhD-level questions in physics, chemistry, and biology
Questions designed by PhD students to be unsolvable through Google. The bar is that a domain expert (not the author) gets them right with time and references.
GPT-42023
39%
Claude 3.5 Sonnetmid 2024
59%
PhD expertsreference
~70%
o3 · Gemini 2.5 Pro2024-2025
87-88%
Gemini 3.1 Pro ✦Feb 2026 · SOTA
94.3%
Frontier models have surpassed the median PhD expert on this benchmark since 2024
Hard-science frontiers — Gemini 3 Deep Think (Feb 2026)
Three benchmarks that measure what was previously considered exclusively human territory: theoretical-physics research, international science Olympiads, and questions designed to be unsolvable by any model.
CMT-BenchmarkCondensed-matter quantum physics · open research problems
50.5%
First significant result on a frontier theoretical-physics benchmark. No published human baseline — active-research questions.
Humanity's Last Exam (HLE)Elite benchmark · multidisciplinary · no tools
48.4%
Designed to be the hardest possible benchmark. Human experts with time score ~85-90%. First time any model exceeds 40% without tools.
IPhO / IChO 2025International Physics and Chemistry Olympiad · written section
Gold medal
Gold-medal level on the written section of both Olympiads — the exam the world's best physics and chemistry students complete over days.
Source: blog.google — Gemini 3 Deep Think announcement · 12 Feb 2026
The pattern is the same in all of them: models crossed the bar of the human expert with time and resources. The next leap is not beating tests — it is solving physics problems humans have not yet solved, as CMT-Benchmark already asks them to do.
Source: GPQA Diamond (Rein et al., arXiv 2023) · Gemini 3 Deep Think announcement (blog.google, 12 Feb 2026) · HLE (Scale AI, 2025) · IPhO / IChO 2025
AI applied directly to real scientific-research problems: nuclear fusion, materials discovery, protein design, weather prediction, and the autonomous scientific cycle.
AI applied to real science — DeepMind & Google
Nuclear fusion · TCV Tokamak (EPFL)
DeepMind trained a reinforcement-learning neural network that controls the shape of magnetic plasma in Lausanne's TCV reactor in real time. First autonomous control of simultaneous plasma configurations published in Nature (2022).
Nature 2022 · Magnetic control of tokamaks
DeepMind + CFS · AI for nuclear fusion
Google DeepMind partnered with Commonwealth Fusion Systems (CFS) to bring AI to the SPARC tokamak — the first designed to reach breakeven (more fusion energy than it consumes). Three axes: the TORAX simulator (JAX, open source) for millions of virtual experiments, RL to optimise reactor operation, and real-time plasma control. Google also invested in CFS. Built on the TCV/EPFL RL control work (Nature 2022).
Google DeepMind · October 2025 · deepmind.google/blog/bringing-ai-to-the-next-generation-of-fusion-energy
🔷
GNoME · Materials discovery
Graph Networks for Materials Exploration discovered 2.2 million stable crystal structures — 800 times more than the entire history of crystallography. 381,000 candidates for batteries, superconductors, and semiconductors. Some have already been synthesised in the laboratory.
Nature 2023 · 2.2M new crystal structures
🧬
AlphaFold 3 · Structure of all molecules
Extends structure prediction beyond proteins to RNA, DNA, ligands, and their mutual interactions. A foundation for drug design. AlphaProteo generates binding-protein designs for specific therapeutic targets with an unprecedented success rate.
Nature 2024 · 2024 Nobel Prize in Chemistry (Hassabis & Baker)
🌍
WeatherNext / GraphCast · Weather prediction
Outperforms ECMWF numerical models in accuracy on 10-day forecasts, at a fraction of the computational cost. Used in production for hurricane alerts. First AI model to beat the operational physics-based standard for global weather forecasting.
Science 2023 · In production in Google Weather
🔬
The AI Scientist · Sakana AI
Fully autonomous scientific-research system: generates hypotheses, designs experiments, implements code, analyses results, and writes complete papers, including peer review of other papers. Cost: ~$15 per complete paper. First demonstration of automation of the full scientific cycle.
Sakana AI · arXiv 2024 · Autonomous scientific cycle
🧬
AlphaEvolve · Google DeepMind
Combines Gemini with evolutionary algorithms to discover and optimise code. Rediscovered an improvement to a matrix-multiplication algorithm that had not been improved for 56 years. Active improvements in the Borg cluster scheduler, TPU chip design, and Google's data centres. Direct operational productivity on production infrastructure.
Google DeepMind · May 2025 · Algorithms + infrastructure
The pattern is consistent: AI does not replace the scientist; it lets them explore a hypothesis space millions of times larger in the same time. The bottleneck shifts from “generating candidates” to “synthesising and validating them” in the laboratory.
Source: Nature (2022, 2023, 2024) · Science 2023 · 2024 Nobel Prize in Chemistry · Google DeepMind · Gemini 3 Deep Think announcement (blog.google, 12 Feb 2026) · DeepMind + CFS (deepmind.google, Oct 2025)
AI vs humans: the map is uneven
There is no single “human level”. Compare a few representative domains to see where the benchmark has already fallen and where it remains open.
ChessReference: world champion
1997
Deep Blue crossed the human benchmark; current engines are far above it.
Vision · ImageNetReference: human error ≈5%
2015
ResNet crossed the reference human error rate in image classification.
Protein structureReference: CASP
2020
AlphaFold2 reached accuracy close to experimental methods on CASP14.
Science OlympiadsIMO · IPhO · IChO
2025–26
Frontier models have reported gold-medal performance in recent Olympiads.

Frontier models from 2025-2026 show expert-level performance in specific domains such as software, formal mathematics, or text analysis, but there is no public, agreed evidence that they have reached the Competent AGI threshold under DeepMind's framework across most cognitive tasks. In domains requiring physical experience, tacit knowledge, or robust causal reasoning, they remain below it.

METR evaluates the task time horizon: the length of task a model can solve with 50% reliability. In March 2025 that horizon was ~1 hour; with GPT-5-thinking, METR estimates it at ~2 hours 15 minutes (METR, 2025). The trend is a doubling roughly every ~7 months, and the next significant threshold is the jump to days or weeks, where the risks of real autonomy emerge.

ARC-AGI-2 measures the capability still missing for cognitive AGI: reasoning about completely new problems from very few examples, without memorising patterns. Launched with initial results below 4%, Gemini 3 Deep Think reached 84.6% in February 2026, close to the ~85% threshold for beating the benchmark (Gemini 3 Deep Think blog). Humanity's Last Exam (HLE), the hardest benchmark published to date, reached 48.4% with the same model, while human experts with references score ~85-90%. The ARC Prize organisers themselves insist that "AGI remains unsolved" and that ARC-AGI-2 was designed to keep tasks easy for humans and difficult for AI (ARC Prize).

The pace of progress over the last five years is unprecedented. Emergent capabilities with scale suggest dynamics that the scientific community does not fully understand, and the AGI debate has moved from academic speculation to the public, regulatory, and foreign-policy agenda.

The most useful question is not "when will AGI arrive?" Nobody honestly knows. The question is which criteria for thinking and which evaluation frameworks make you more robust in an environment where AI improves quickly and the landscape changes every few months.

Position summary

What we do know: current systems outperform human experts in specific, bounded domains. The horizon of autonomous tasks is growing predictably. General-reasoning benchmarks are improving faster than expected.

What we do not know: whether emergent capabilities with scale converge towards something that deserves to be called AGI or whether there is a ceiling we do not know about. Whether alignment is a technical problem that can be solved before reaching much more capable systems. Whether the qualitative jumps observed in benchmarks translate into real generalisation outside the laboratory.

What reaching it would imply: a reorganisation of the division of cognitive labour deeper than industrialisation. Compression of the time between scientific discovery and application. And the need to solve alignment before the system is capable enough for errors to become irreversible.

That is what this series has tried to build: a stable mental map that works even when the models change.


6. References

Base sources
Key Source Short description
R1 Morris et al. (2023)Levels of AGI: Operationalizing Progress on the Path to AGI (arXiv) DeepMind's six-level (0-5) framework for operationalising AGI.
R2 Bubeck et al. (2023)Sparks of Artificial General Intelligence: Early experiments with GPT-4 (arXiv) Systematic evaluation of GPT-4 against the cognitive-AGI bar.
R3 OpenAI (2023)OpenAI Charter (OpenAI) OpenAI's canonical definition of AGI: "highly autonomous systems that outperform humans at most economically valuable work".
R4 Russell, S. (2019)Human Compatible: Artificial Intelligence and the Problem of Control (book, Basic Books) Central argument on the alignment problem and the design of AI compatible with human values.
R5 Bostrom, N. (2014)Superintelligence: Paths, Dangers, Strategies (book, Oxford University Press) The intelligence-explosion scenario and its risks. A debate reference, not scientific consensus.
R6 Krakovna et al. (2020)Specification gaming: the flip side of AI ingenuity (DeepMind blog) Real examples of systems optimising the wrong metric with unforeseen results.
R7 Grace et al. (2024)Thousands of AI Authors on the Future of AI (arXiv) Survey of AI researchers on probabilities and estimated timelines for AGI milestones.
R8 McKinsey Global Institute (2023)The economic potential of generative AI: The next productivity frontier (McKinsey) Estimates that generative AI could automate activities representing 60-70% of workers' time.
R9 Briggs, J. & Kodnani, D. (2023)The Potentially Large Effects of Artificial Intelligence on Economic Growth (Goldman Sachs) Estimates that two-thirds of jobs in the US and Europe are exposed to some degree of AI automation; ~25% of tasks are directly automatable.
R10 OpenAI (2025)GPT-5 System Card (OpenAI) Results on SWE-bench Verified (74.9%), METR evaluations (task horizon ~2h15m), and comparisons with human experts in scientific domains.
R11 Anthropic (2026)Introducing Claude Sonnet 4.6 (Anthropic) Official announcement with computer-use and coding capabilities and a 1M-token context window (beta).
R12 METR (2025)Measuring AI Ability to Complete Long Tasks (METR) Introduces the task-horizon metric: task length completable with 50% reliability doubles roughly every ~7 months; Claude 3.7 Sonnet reaches ~1 hour.
R13 Google DeepMind (2026)Gemini 3.1 Pro (deepmind.google) Gemini 3.1 Pro: GPQA Diamond 94.3%; SWE-bench Verified 80.6% (new SOTA as of Feb 2026); ARC-AGI-2 77.1%.
R14 The Deep Think team (2026)Gemini 3 Deep Think: Advancing science, research and engineering (blog.google) Gemini 3 Deep Think: ARC-AGI-2 84.6% (verified by the ARC Prize Foundation); HLE 48.4% without tools; gold medal in IMO 2025, IPhO 2025, and IChO 2025.
R15 ARC Prize Foundation (2025)Announcing ARC-AGI-2 and ARC Prize 2025 (arcprize.org) Launch of ARC-AGI-2; insists that "AGI remains unsolved" and details the result-verification methodology.
R16 Anthropic (2024)Responsible Scaling Policy v2.1 (Anthropic) Defines AI R&D autonomy thresholds (AI R&D-1 through AI R&D-5) and their relationship to safety measures.
R17 OpenAI (2026)Why SWE-bench Verified no longer measures frontier coding capabilities (OpenAI) Explains why SWE-bench Verified is contaminated and recommends SWE-bench Pro and other alternative benchmarks.

Frequently asked questions

What levels define the path towards AGI according to DeepMind? DeepMind proposes a spectrum of six levels numbered from 0 to 5: from no AI through to Superintelligence. Frontier models from 2025-2026 sit at the boundary between competent AI and expert AI in specific domains, but they have not reached expert-level generality across most cognitive tasks, which would be the threshold for level 3 in that framework.

How does the economic definition of AGI differ from the classical cognitive definition? The cognitive definition requires performing any human intellectual task, a fuzzy bar because humans also have biases and limits. OpenAI's economic definition focuses on outperforming humans at most economically valuable work, a bar measurable with labour benchmarks although lower in some respects, because part of cognitive work could be automated without the system reaching real cognitive generality.

What criterion most concerns safety researchers when discussing AGI? For them, the critical line is not task performance but recursive improvement: a system that improves its own design to produce successively more capable systems. If that threshold were crossed, the pace of change would exceed humans' ability to understand and control what is happening, regardless of whether the system satisfies the cognitive or economic definition.

What does the task time horizon (METR) measure and why does it matter for autonomy? It measures the maximum task duration an agent completes with 50% reliability, not the time it takes to answer questions but sustained multi-step work. In March 2025 that horizon was approximately two hours and fifteen minutes for the most capable models, and the trend is a doubling every seven months. The next significant threshold is the jump to days or weeks, where the risks of real autonomy emerge.

Keep learning
Series completedChoose the next pathAll series