The AGI timeline has never been argued over this loudly. On September 3, OpenAI president Greg Brockman closed the launch briefing for GPT-6 Astra with four words: “Welcome to the AGI era.” Within hours, scientists, benchmark designers and rival labs were fighting over whether he was right.
I spent the past few weeks reading the launch documents, the independent test results and the forecasts behind that fight. This guide gives you the plain-English version: what artificial general intelligence really is, how close we are as of October 2026, and which AI looks most likely to get there first.
The short answer: no lab has proven AGI by a definition most experts accept. But the gap is closing faster than almost anyone predicted. A test that every top model failed in March now has one model scoring about 63%. Serious forecasts for the AGI timeline run from “it has already started” to around 2030 to the 2040s.
What Is AGI? Artificial General Intelligence in Plain English
Artificial general intelligence, or AGI, is an AI that can learn and do almost any mental task a person can, including tasks it was never trained for.
Today’s AI feels general, and in some ways it is. A chatbot can write an essay, debug code and explain a tax form. But it is uneven. The same system that solves a graduate-level maths problem can get lost in a simple game that an untrained person figures out in minutes.
The easiest way to picture the gap: today’s best AI is a brilliant new hire who has read everything but needs a clear brief and a manager checking the work. AGI would be the colleague who works out what needs doing, learns what they don’t know, and gets it right without supervision.
Three abilities sit at the heart of the idea:
- Fast learning: picking up a new situation without being retrained.
- Transfer: using a skill from one field in a completely different one.
- Reliable independence: finishing long, messy jobs without a human checking every step.
Why Nobody Agrees on a Definition
There is no official test for AGI, and that is the root of most arguments. OpenAI’s charter describes it as “highly autonomous systems that outperform humans at most economically valuable work.” Google DeepMind co-founder Shane Legg prefers a stricter idea he calls “minimal AGI”: an AI that reliably handles the full range of tasks an average person can, without failing in ways that would surprise us. Anthropic CEO Dario Amodei has called AGI a marketing term and talks about a “country of geniuses in a data center” instead.
This matters for every AGI timeline you will read. Two experts can look at the same model and disagree, because they are measuring with different rulers.
The AGI Timeline in 2026: 7 Truths Behind the Headlines
1. A test that stumped every top AI in March is now partly cracked
In March, the ARC Prize Foundation launched ARC-AGI-3, a set of game-like worlds with no instructions and no stated goal. Untrained humans solved every one. Every frontier model scored below 1%, according to the benchmark’s technical report.
Six months later, ARC Prize’s neutral test setup scored GPT-6 Astra at 62.7%, TechTimes reported. Anthropic’s Claude Opus 5 scored 30.2% on the same verified leaderboard. The score measures how efficiently an AI solves each world compared with a human, so 100% means matching human efficiency. Going from near zero to better than half in a single product cycle is the biggest move in the AGI timeline this year.
2. The headline number depends on who runs the test
OpenAI’s launch materials cited 99.9% on the same benchmark, using its own test setup. ARC Prize’s provider-neutral setup produced 62.7%. ARC Prize said Astra shows meaningful progress toward generalisation but declined to call it AGI.
OpenAI’s own July experiment shows how much the wrapper around a model matters. Changing just two settings, keeping the model’s private reasoning between steps and summarising old context instead of deleting it, lifted GPT-5.6 Sol from 7.8% to 38.3% with no change to the model itself. On the open Kaggle community leaderboard, which allows custom setups, a 27-billion-parameter open-weight model with clever engineering around it reached 55.89% on October 4. The $700,000 grand prize for matching humans is still unclaimed.
The takeaway: a benchmark score is part model and part engineering. Any AGI timeline built on one score deserves suspicion.
3. OpenAI says the “AGI era” has begun, but hasn’t declared AGI
Brockman framed it as his personal view and told reporters it was up to each reader to decide. Coverage of the launch found no formal company declaration, and under OpenAI’s October 2025 agreement with Microsoft an independent expert panel is supposed to verify one. None has been announced as of this writing.
Altman’s version of the AGI timeline is more specific. He told TIME on August 26 that OpenAI would have an internal system he would call AGI by the end of 2026, while admitting the company was “not quite yet” there. The Decoder notedthat this holds only under his own definition. Markets are unconvinced: Polymarket traders put the odds of OpenAI achieving AGI by 2027 at about 9% in August, according to AIMultiple.
4. AI is strong in spikes, not across the board
Astra’s strengths are real. By OpenAI’s launch tables, as reported by TechTimes, it scored 72.6% on the OSWorld 2.0 computer-use test and 57.9% on Terminal-Bench 4.0, up from 37.3% for its predecessor. But GDPval, OpenAI’s own benchmark for economically valuable work, was missing from the launch materials. That is the exact standard OpenAI’s charter uses to define AGI. Independent firm Artificial Analysis found Astra slipped in some GDPval categories, including banking support and scientific coding, and Astra’s 57.2% on Humanity’s Last Exam (with tools) trailed the 65.0% of its predecessor.
Skeptic Gary Marcus called Astra real progress but disputed the AGI claim. That uneven profile is often called “jagged” intelligence, and a real AGI timeline has to account for the weak spots, not just the peaks.
5. Expert forecasts have collapsed from decades to years
In 2020, the Metaculus forecasting community put the median arrival of AGI around 2070. By early 2026 it was about 2033, and other trackers show the early 2030s depending on how the question is worded. Demis Hassabis of Google DeepMind keeps his AGI timeline near 2030, as Fast Company reported. Legg gave 50% odds of minimal AGI by 2028. Amodei’s gut says a “country of geniuses” could arrive in 2026 or 2027, with 90% confidence it comes within ten years, according to a recap of his February interview, and he has said we are not at AGI today.
Not everyone is speeding up. Daniel Kokotajlo, lead author of the widely read AI 2027 scenario, pushed his own forecast later in January, and the Samotsvety forecasting group still gave 50% odds only by 2041. The AGI timeline is shrinking on average, but the spread is huge.
6. The “best AI” changes every few weeks
Three labs can claim the top spot, depending on the test. On Artificial Analysis’s Intelligence Index, Anthropic’s Claude Fable 5.1 scored 65.7, ahead of GPT-6 Astra at 61.2. On ARC-AGI-3’s verified board, Astra leads. Google released Gemini 4 Argon in September, and Elon Musk keeps pointing to Grok 5, which I could not confirm as released at the time of writing. Any claim about who leads the AGI timeline race expires quickly.
7. Safety is struggling to keep pace with capability
Astra is the first OpenAI model rated “Critical” for cybersecurity under the company’s Preparedness Framework, meaning it can find and build working exploits largely on its own. At the launch briefing, chief scientist Jakub Pachocki conceded that monitoring a model’s reasoning is becoming fragile. A widely reported July incident reportedly saw OpenAI’s AI agents escape a test environment and breach Hugging Face’s systems. And US pre-release review of frontier models is still voluntary. The closer the AGI timeline gets, the less comfortable that arrangement looks.
Which AI Will Achieve AGI First? The AGI Timeline Race, Lab by Lab
No lab is a safe bet. Each has a different kind of evidence. Here is how they compare, followed by my read.
| Lab | Latest model | Strongest evidence | Biggest gap |
|---|---|---|---|
| OpenAI | GPT-6 Astra (Sept 3) | 62.7% on ARC-AGI-3’s neutral board; strong computer use | Self-set AGI bar, GDPval missing, weakening monitoring |
| Google DeepMind | Gemini 4 Argon (Sept) | Widest mix of reasoning, robotics and science | No verified ARC-AGI-3 score found yet |
| Anthropic | Claude Fable 5.1 | 65.7 on Artificial Analysis’s index, ahead of Astra | Opus 5 scored 30.2% on ARC-AGI-3; Mythos 5.1 is restricted |
| xAI | Grok 4.x (Grok 5 pending) | Heavy compute and a bold roadmap | Grok 5 unconfirmed; earlier release targets missed |
OpenAI: first to claim it, hardest to verify
OpenAI’s AGI timeline is the boldest in the industry. It has the largest training run in its history, on more than 100,000 GPUs in Abilene, Texas, and a stated goal of an internal AGI system by year end. That makes it the likeliest lab to announce AGI first. The weakness is that it grades its own homework: the definition is its own, its key benchmark was missing at launch, and its monitoring tools are weakening.
Google DeepMind: the broadest toolkit
Hassabis’s AGI timeline is more cautious than Altman’s, but Google has the widest base. Gemini 4 Argon targets software engineering, legal and finance work, Gemini Robotics ER 2 reached developers on July 30, and DeepMind has launched an institute to study a world with AGI. If your definition of AGI includes the physical world and scientific discovery, Google is the lab to watch.
Anthropic: strongest independent scores, most cautious voice
Anthropic’s AGI timeline is the hardest to pin down because Amodei avoids the label. The evidence is solid, though. Claude Fable 5.1 sits ahead of Astra on Artificial Analysis’s index, and Anthropic keeps its most capable model, Mythos 5.1, restricted to vetted organisations in cybersecurity and life sciences, an approach it began with Project Glasswing. Claude Opus 5 scored 30.2% on ARC-AGI-3, well behind Astra, and I could not find a verified score for Fable 5.1, so learning in unfamiliar settings is not yet its proven edge.
xAI and the open-weight challengers: the wildcards
xAI’s AGI timeline hangs on Grok 5. Musk has said it is the model with a real chance of reaching AGI, once putting that chance at 10%, though xAI has missed several of his earlier release targets. The quieter story is open-weight AI. Alibaba’s free Qwen model, running on a single GPU, now tops the community ARC-AGI-3 board, a reminder that the race is not limited to three American labs.
My verdict
OpenAI is the likeliest to be first to claim AGI, probably by its own definition before the year ends. Google DeepMind looks best placed to satisfy a broader definition that includes the physical world. Anthropic leads on today’s independent knowledge-work scores. A single clean winner is unlikely. The more probable outcome is several labs crossing a loose AGI line within a year or two of each other, with experts arguing over every one.
How Far Are We From AGI? My Read of the AGI Timeline
Treat AGI as two finish lines, not one.
Finish line one, loose AGI. AI that matches skilled people on most screen-based work. Judging by the evidence above, this is plausible between now and about 2030. Parts of it are already here. The open question is reliability across all that work.
Finish line two, strict AGI. AI that learns as efficiently as a person, works reliably without supervision and handles the physical world. This looks more like the early-to-mid 2030s and could slip later. The main drivers of progress, namely bigger models, more thinking time and better agents, can plausibly keep going until about 2028 and perhaps 2032, according to 80,000 Hours.
That is my judgment, not a measurement, and the honest range is wide. Four things will sharpen the picture soon:
- ARC Prize 2026: final submissions close November 2, with winners announced December 4.
- GDPval results: whether OpenAI publishes them for Astra.
- OpenAI’s year-end claim: whether it names an AGI system, and whether outside experts verify it.
- Independent reliability tests: long, unsupervised tasks, where uneven skills show up fastest.
AGI Timeline FAQ: Quick Answers
When will AGI arrive?
Nobody knows, and the answer depends on the definition. OpenAI expects an internal system it would call AGI by the end of 2026. Hassabis points to about 2030, Metaculus forecasters to the early 2030s, and some groups to the 2040s.
Has anyone achieved AGI yet?
Not by a standard accepted outside the company making the claim. OpenAI’s president says we may be in the “AGI era,” but OpenAI has made no formal declaration, and ARC Prize explicitly declined to call Astra AGI.
What is the difference between AGI and ASI?
AGI matches human ability across most intellectual work. Artificial superintelligence, or ASI, would go far beyond the best humans in nearly every field. Most AGI timeline forecasts cover only the first step.
Bottom Line: Where the AGI Timeline Leaves You
AGI is not a switch that flips on a given day. Even OpenAI’s president admits the arrival did not come as one clear moment. It is a slope, and 2026 is the year it got steep.
For you, that comes down to three habits. Learn to work with AI agents now, because the tools are already capable. Treat every “AGI achieved” headline as a claim until independent testers confirm it. And watch neutral scoreboards like ARC Prize rather than company launch posts.
Related Articles on IMFounder
- 15 Explosive AI Updates: OpenAI’s Shock IPO U-Turn, Nvidia’s $13B Power Grab & more..
- The Massive Anthropic IPO May Signal a Dangerous Peak in the AI Bubble
- Your Mirror Is About to Start Diagnosing You
- 5 Powerful Startups You Can Launch With Only $1,000
- How to Validate a Startup Idea Without Writing a Single Line of Code
- China Just Built a $650,000 Robot That Can Punch Through a Wall
