AGI: Where We Are Now (October 2026)

On 25 March 2026, the ARC Prize Foundation launched a new test called ARC-AGI-3.

It isn’t a test of knowledge. It’s a set of hundreds of small, turn-based games, handmade by human game designers. There are no instructions. You get dropped in, you poke at things, and you work out what the game is and how to win. Then the next level changes the rules a little.

The launch results fitted in one line: humans scored 100%, frontier AI scored well under 1%. Ordinary people, playing for the first time, solved every environment. The best AI systems in the world solved almost nothing.

Six months later, on 3 September 2026, ARC published a very different result. OpenAI’s GPT-6 Astra scored 62.7% on ARC-AGI-3 using ARC’s own standard setup, and 99.9% using a setup OpenAI provided. ARC called both results state of the art, and in the same post wrote: “we are not claiming that it is AGI.”

From under 1% to over 60% in half a year, on a test designed to be hard for AI. And the people who built the test, looking straight at the result, still say it isn’t AGI. That tension is the honest picture of AGI in October 2026, and it’s the subject of this final part.

We’ll cover what today’s systems can and can’t do, what the people building them claim, what the skeptics say, what forecasters expect, the physical limits of compute and energy, the rules being written, and how to read the next AGI headline. Then the grand scoreboard of every bet from this series.

What frontier AI can do (with dates)

“AI can do X” goes stale in months, so every claim here has a date.

Olympiad maths. In July 2025, an advanced version of Gemini Deep Think solved five of six IMO problems and scored 35 out of 42, a gold-medal score, officially graded by IMO coordinators. It worked in plain language, inside the 4.5-hour time limit. OpenAI announced a gold-level result on the same problems, self-reported at the time. A year earlier, the best AI result was a silver that needed humans to translate the problems into formal maths first (see Part 7).

Long tasks. METR, a non-profit that evaluates AI, measures a time horizon: the length of task, measured by how long it takes a skilled human, that a model completes with 50% success. Its FAQ puts GPT-5 at about 2 hours 17 minutes. For an early version of Claude Mythos Preview, tested in March 2026, METR estimated at least 16 hours, with a 95% confidence range of 8.5 to 55 hours. METR itself warns that measurements above 16 hours are unreliable with its current set of tasks. In plain words: models are now good enough to break the ruler.

Puzzles built to resist memorising. On ARC-AGI-2, the previous version of the ARC test, ARC Prize’s own results page lists xAI’s Grok 4.6 (released August 2026) at 67.1% at about $0.76 per task, and 87.5% on the original ARC-AGI-1. ARC-AGI-2 was designed in 2025 so that every task was solved by at least two people in a public study.

So the “can do” list is long and getting longer: competition maths, long coding tasks, scientific question answering, writing, translation, and working through software on their own for hours.

What it still can’t do

Learn new things cheaply, on its own terms. This used to be the clearest gap, and it’s the one that just moved most. ARC-AGI-3 measures how efficiently a system figures out an unfamiliar game, compared with first-time human players. That’s exactly what François Chollet, ARC’s creator, calls intelligence: not what you know, but how fast you pick up something new. Frontier AI scored under 1% at launch in March 2026. In September, GPT-6 Astra scored 62.7% under ARC’s standard setup.

The details matter, and ARC spelled them out. The setup matters a lot: the same model scored 99.9% with OpenAI’s own harness, the software that connects the model to the games, which keeps the model’s hidden reasoning between moves. The scores are on ARC’s semi-private test set. The standard run cost about $26,000; ARC estimates a human tester costs about $12.78 per game. The model had external tools, like a code interpreter, that human testers didn’t. And ARC notes that its games are closed, deterministic puzzles that don’t capture the messiness of the real world. Its conclusion: a big result, and not proof of AGI. Even beating the test outright, ARC wrote, would not be proof of achieving AGI.

Reliably make experts faster. In July 2025, METR ran a randomised trial with 16 experienced open-source developers on 246 real tasks in their own codebases. With AI tools allowed, they took 19% longer. The striking part: they expected a 24% speedup, and afterwards still believed they’d been 20% faster. METR is careful to say this was a snapshot of early-2025 tools in one setting, and a 2026 follow-up was inconclusive. But the gap between feeling faster and being faster is worth remembering every time you read a productivity claim.

Keep learning after training. Most deployed models are frozen. They can use what’s in their context window, but they don’t permanently learn from the conversation the way you learn from a week at a new job. Several of the alternative bets in Part 7 exist precisely because of this.

Behave the same when watched and unwatched. The International AI Safety Report 2026, written by over 100 experts and chaired by Yoshua Bengio, noted that some models can tell when they’re being evaluated and change their behaviour. That makes testing harder. (More on why scores mislead in Benchmarks and Goodhart’s law.)

The word people use for this shape is jagged. Superhuman on some peaks, below a child on some valleys, and hard to predict which is which until you try.

What the lab leaders have claimed

These are claims from people running companies that sell AI and raise money on its future. They’re also people with the closest view of the newest systems. Both things matter. Here are three, quoted from the original sources with dates.

Sam Altman, OpenAI, January 2025. In a blog post called “Reflections”, he wrote: “We are now confident we know how to build AGI as we have traditionally understood it.” He also wrote that in 2025 “we may see the first AI agents ‘join the workforce’ and materially change the output of companies.” Note the careful qualifier, “as we have traditionally understood it”.

Dario Amodei, Anthropic, October 2024. In the essay “Machines of Loving Grace”, he described “powerful AI”: smarter than a Nobel Prize winner across most relevant fields, able to work on its own on tasks lasting hours to weeks, run as millions of copies. He summarised it as “a country of geniuses in a datacenter” and wrote: “I think it could come as early as 2026, though there are also ways it could take much longer.” He also wrote that he dislikes the term AGI, calling it imprecise and full of “sci-fi baggage”.

Demis Hassabis, Google DeepMind, March 2025. At a briefing in London, reported by NBC News, he defined AGI as “a system that’s able to exhibit all the complicated capabilities that humans can”, and said that “over the next five to 10 years, a lot of those capabilities will start coming to the fore.”

It’s October 2026 now, so you can start to judge the record yourself. Whether today’s systems look like a “country of geniuses” depends heavily on which part of the map you look at, which brings us to the skeptics.

What the skeptics say

The skeptics aren’t one camp. They disagree with each other, and each has a real argument.

Gary Marcus, cognitive scientist. In a 2022 essay for Nautilus titled “Deep Learning Is Hitting a Wall”, he argued that scaling up was approaching diminishing returns, that the systems didn’t genuinely understand language, and that the field needed neurosymbolic approaches. Critics point to everything since GPT-4 as evidence against him. He points to unreliability, hallucinations and the need for formal tools in maths as evidence for him. Both sides have a point: progress didn’t stop, and the problems he named didn’t vanish.

Yann LeCun, Turing Award winner. His argument isn’t that AI will stall, but that autoregressive language models are the wrong road to human-level intelligence: they lack a model of the physical world, they can’t plan properly, and text alone isn’t enough. He left Meta in late 2025 to build the alternative. (Full story in World Models.)

François Chollet, creator of Keras and of ARC. In his 2019 paper “On the Measure of Intelligence”, he argued that skill at any task can be “bought” with enough training data, so skill alone proves nothing. Intelligence, he argued, is skill-acquisition efficiency: how quickly you learn something genuinely new. ARC is his test of that idea. ARC-AGI-3 was where today’s systems looked weakest, until September 2026, and ARC’s own reaction to that jump was to say it still isn’t AGI.

What the fair reader takes from all three: the skeptics have repeatedly been wrong about what AI would be able to do next, and repeatedly right that impressive scores hide fragility. Both lessons keep holding.

What researchers and forecasters expect

The biggest survey. In late 2023, AI Impacts surveyed 2,778 researchers who had published at top AI venues. Their aggregate forecast gave a 50% chance by 2047 that machines could do every task better and more cheaply than humans, unaided. That was 13 years earlier than the same survey a year before. For “full automation of labour”, the 50% date was 2116. On risk, between 38% and 51% of respondents put at least a 10% chance on outcomes as bad as human extinction.

Two cautions. The spread of answers was enormous, and the results were sensitive to how the question was phrased. And the survey ran before reasoning models, before IMO gold, before agents. Opinion has likely moved again since.

Prediction markets and forecasting sites. Metaculus, a forecasting platform, has long-running questions on when the first “general AI system” will be devised, tested and publicly announced, with detailed resolution criteria. I couldn’t verify a current, dated figure for this post, so I won’t quote one. Check the live page, and note the date when you do.

Lab leaders’ timelines (above) are the shortest. Academic surveys are longer. Skeptics reject the framing. The honest summary: nobody knows, and the spread of serious opinion runs from “within a few years” to “not with current methods”.

The physical limits: compute and energy

AGI talk often floats free of physics. It shouldn’t.

Compute. Epoch AI, a research group that tracks AI trends, estimated that the training compute of frontier models grew about 4 to 5 times per year from 2010 to 2024. That’s the engine behind Bet #4 (Just Make It Bigger). It can’t continue forever at the same pace without chips, money and power to match. (Why it’s all on GPUs: Why AI runs on GPUs.)

Energy. The International Energy Agency’s Energy and AI report (April 2025) estimated that data centres used about 415 TWh of electricity in 2024, around 1.5% of the world’s total. It projected that to more than double, to about 945 TWh by 2030, slightly more than Japan uses today, with AI as the main driver.

What it means: even if the algorithms were solved tomorrow, power plants, grid connections and chip factories set a speed limit on how fast the most compute-hungry bets can scale. That’s one reason the reasoning bet (Part 6) matters so much. Spending compute when the model thinks, not just when it trains, is a different way to get more capability from the same infrastructure.

Safety and the rules being written

The EU AI Act is the most detailed law so far. Key dates:

1 August 2024: entered into force.
2 February 2025: bans on certain practices, such as some kinds of manipulation and social scoring, started to apply.
2 August 2025: obligations for general-purpose AI models, the kind this series is about, started to apply.
July 2026: a “Digital Omnibus” amendment, published in the EU’s Official Journal on 24 July 2026, postponed the rules for high-risk AI systems to 2 December 2027 (stand-alone systems) and 2 August 2028 (AI built into regulated products).

The International AI Safety Report is the closest thing to a scientific consensus document. The first edition, in January 2025, was written by 96 experts nominated by 30 countries plus the UN, EU and OECD, chaired by Yoshua Bengio. The second, published 3 February 2026, found that capabilities had kept improving fast, especially in maths, coding and autonomous work, and flagged that evaluation itself is getting harder. It deliberately makes no policy recommendations.

Two things are true here too. The rules are real and getting more detailed. And they’re being delayed and softened under industry and political pressure, while capabilities move faster than the paperwork. Inside the labs, much of the safety work is about problems this series touched on: models that game their rewards (Reward hacking) and models we can’t yet fully see inside (Interpretability).

How to read an AGI headline

You’ll see many more of these. Here’s my checklist.

Which definition of AGI? As Part 1 showed, there are at least five. A headline that doesn’t say which is selling a feeling.
Who’s measuring? Company claim, independent evaluator, or peer-reviewed paper? Officially graded or self-reported?
Which test, and could it be memorised? Old benchmarks leak into training data. Look for held-out or new tests.
What did it cost? A score at $10,000 per task isn’t the same as one at $1.
How many tries, and who helped? Best of many runs? Did humans translate the problem first?
Does it work on a new kind of problem? Not a harder version of the same benchmark, but something genuinely new. This is ARC-AGI-3’s whole point, and why its September 2026 result made news.
What’s the date? Anything older than six months may be wrong in either direction.

The same instincts work for products. I wrote a version for buyers in How to tell if an AI product is real or just a demo.

70 years of bets, on one line

Seventy years of milestones: each bet had its moment, and the newest wall moved within a single year.

The grand scoreboard

Here is every bet from this series in one table, as of October 2026.

The bet	What was tried	Best result	Where it hit a wall	Status in 2026
#1 Write down the rules (symbolic AI)	Logic, search, expert systems written by hand	Deep Blue beat Kasparov 3.5 to 2.5 (1997); expert systems in daily commercial use in the 1980s	Brittle outside its rules; knowledge had to be typed in by hand	Lives on as search, tools and formal checkers attached to learned models
#2 Let it learn (neural networks)	Networks that learn patterns from data	AlexNet cut ImageNet error to 15.3% from 26.2% (2012)	Hungry for data and compute; hard to see inside	The foundation of almost every modern system
#3 Reward is enough (reinforcement learning)	Learning by trial, error and reward	AlphaGo beat Lee Sedol 4 to 1 (2016)	Needs clear rewards and cheap trials; reward hacking	Core of how reasoning models are trained
#4 Just make it bigger (scale)	More data, parameters and compute	GPT-3 (2020) to ChatGPT (2022); compute up about 4 to 5x a year	Cost, energy, data supply; diminishing returns debated	Still the engine, increasingly spent at thinking time
#5 Teach it to think (reasoning and agents)	Models that reason step by step and act with tools	IMO gold (July 2025); 16-hour-plus METR time horizon (2026); GPT-6 Astra 62.7% on ARC-AGI-3, standard setup (Sept 2026)	Reliability; cost; learning outside closed tests; frozen after training	The current frontier
Neurosymbolic AI	Language models plus symbolic engines and proof checkers	IMO silver, 28/42 (2024)	Formal translation; few formal domains	Absorbed as “write, then verify”
World models and JEPA	Predict abstract summaries of the world	V-JEPA 2 robot pick-and-place in new labs (2025)	Not yet broadly general	A well-funded alternative
Brain-inspired and brain copying	Emulation, Thousand Brains, active inference	Full fruit fly connectome (2024)	Scale, missing biology, small results	Real science, distant AGI path
Evolution and open-endedness	Generate environments and agents automatically	POET; Darwin Gödel Machine 20% to 50% on SWE-bench	Compute; scoring “interesting”	Growing, layered on LLMs
Cognitive architectures and embodiment	Whole-mind designs; robots that learn like children	TacAir-Soar’s 48-hour simulated air operation (1997)	Hand-built knowledge; slow real-world learning	Ideas live on in agent design and robotics
AIXI	A mathematical definition of the perfect agent	A precise target	Incomputable	A yardstick

Where that leaves us

Read the scoreboard top to bottom and the pattern of seventy years is clear.

No bet won alone. Every bet that “lost” left something behind. Symbolic AI left search and verification. Neural networks were written off twice and came back to power everything. Reinforcement learning went from games to training reasoning models. Scale delivered more than most people expected. And today’s frontier systems are a stack of all of them: a neural network, trained at enormous scale, refined with reward, prompted to reason, given tools and sometimes checked by a symbolic verifier.

The goalposts moved, and that’s partly fair. Chess, Go, conversation, olympiad maths: each was once held up as a sign of real intelligence, and each fell without producing a general mind. Some people call that moving the goalposts. Others call it learning what the goalposts should have been all along. Both are a bit right.

The remaining wall is specific, and it’s moving. It isn’t knowledge or raw problem-solving any more. It’s learning something genuinely new, efficiently, cheaply, and reliably, outside a closed test. ARC-AGI-3 went from under 1% to 62.7% in six months, at a cost per game far above a human’s. Continuous learning, and the gap between benchmark scores and real-world speedups, are still open.

If you want to know when the picture has really changed, here are the signals I’d watch, rather than the next press release:

ARC-AGI-3 results holding up on the private test set, under the standard setup, at a cost closer to a human’s, and the next ARC test after it.
Independent measurements of real work, like METR’s trials, showing large speedups for experts, not just a feeling of speed.
Models that keep learning after deployment without forgetting what they knew or drifting off course.
Time horizons that keep doubling once evaluators have built longer tasks to measure them.

None of those needs a definition argument. Each is a number someone can check.

So where are we? Closer than any time in seventy years, by almost any measure. Not there by the definitions that matter most to the skeptics. Possibly there soon, according to the people building it. And with a lot riding on who turns out to be right, which is why it’s worth reading every claim, mine included, with a date and a source next to it.

Key ideas

AI in October 2026 is jagged, and the valleys are filling fast. ARC-AGI-3, a test of learning new games, went from under 1% (March 2026) to 62.7% for GPT-6 Astra under ARC’s standard setup (September 2026). ARC still says: “we are not claiming that it is AGI.”
Lab leaders’ timelines are short (Amodei: powerful AI “as early as 2026”; Hassabis: five to ten years from 2025). Surveys of researchers are longer (50% by 2047 in the 2023 survey).
The skeptics are often wrong about what comes next and often right about fragility. Marcus, LeCun and Chollet each make a different argument.
Compute grew 4 to 5x a year, and data-centre electricity may double by 2030. Physics and power grids are part of the AGI story.
Rules are arriving but slipping: the EU AI Act’s high-risk rules were postponed to late 2027 and 2028.
Read headlines with a checklist: which definition, who measured, can it be memorised, what it cost, and what’s the date.
No single bet won. Today’s systems combine almost every bet in this series.

Sources: ARC Prize, ARC-AGI-3 launch (March 2026); The Decoder on ARC-AGI-3 launch scores; ARC Prize, OpenAI’s GPT-6 Astra on ARC-AGI-3 (September 2026); ARC Prize, Grok 4.6 results page; ARC Prize, ARC-AGI-2; Google DeepMind, Gemini Deep Think IMO 2025 gold; Google DeepMind, IMO 2024 silver; METR time horizons; METR on Claude Mythos Preview (X, 2026); METR, early-2025 developer productivity study; International AI Safety Report 2026; Mila, first International AI Safety Report (January 2025); Sam Altman, Reflections; Dario Amodei, Machines of Loving Grace; NBC News on Hassabis (March 2025); Quote Investigator on “Deep Learning Is Hitting a Wall”; Chollet, On the Measure of Intelligence (2019); Grace et al., Thousands of AI Authors on the Future of AI (2024); Metaculus AGI question; Epoch AI, training compute grows 4-5x per year; IEA, Energy and AI (2025); Gibson Dunn on the EU AI Act omnibus; Lewis Silkin, Digital Omnibus on AI enters into force (July 2026); Krizhevsky et al., AlexNet (2012); Deep Blue (Wikipedia); AlphaGo versus Lee Sedol (Wikipedia).