ByAhsan Nadeem@ashtech52Lead Full-Stack Engineer
In September 2026, OpenAI said its agents had cracked a $1 million Millennium Prize maths problem. In October 2026, the best AI model still completes only about one in five real freelance projects well enough for a client to accept. Both are true, and the gap between them is the honest answer to "how close are we to AGI?"
This page is a plain-English scorecard: what AGI means, the dated evidence for and against, what forecasters expect, and what nobody knows yet. No hype, no doom, and no maths degree needed.
The short answer: how close are we to AGI in October 2026?
As of 26 October 2026, we are not at AGI by most definitions, but we are closer than almost anyone predicted five years ago. AI now proves research-level theorems and handles coding tasks that take people more than 16 hours, yet the best model completes only about one in five real freelance projects. Most forecasts point to the early 2030s.
- Not there, by most definitions. No lab, benchmark owner or independent evaluator has declared AGI achieved. ARC Prize, whose test GPT-6 Astra nearly maxed out in September, wrote: "we are not claiming that it is AGI."
- Still unreliable on messy, open-ended work. The same systems often fail long tasks, game their own tests and complete a minority of real paid projects.
- Forecasts cluster around the early 2030s, with wide error bars.
What this means for you, before the detail
- If you run a business, the AGI label won't tell you whether AI can handle your invoices or support tickets. A test on your own past cases will, and the six-step checklist near the end shows how.
- If you're worried about jobs, the best model still completes about one in five real freelance projects. That number is rising fast, so it's worth watching, not panicking over.
Working out which tasks AI can do reliably is the heart of our AI and automation work.
What is AGI? Five definitions, five different answers
AGI means "artificial general intelligence". There's no agreed definition, so the one you pick decides the answer.
Who | Definition | Verdict in October 2026 |
|---|---|---|
OpenAI's charter | "Highly autonomous systems that outperform humans at most economically valuable work" | "Not quite yet", Sam Altman told TIME on 26 August 2026 |
Demis Hassabis, Google DeepMind | A system with "all the cognitive capabilities humans can, and I mean all", including top-level creativity and physical skill | Not yet; he names continual learning, long-term planning and robotics as unsolved |
Dario Amodei, Anthropic | "Powerful AI": smarter than a Nobel winner in most fields, working autonomously for hours to weeks, a "country of geniuses in a datacenter" | Possibly one to two years away (January 2026), "although it could also be considerably further out" |
Hendrycks, Bengio, Marcus and 30 others | Matching "the cognitive versatility and proficiency of a well-educated adult" across ten abilities | GPT-5 scored 57% on their 2025 scale, weakest on long-term memory |
Metaculus forecasting question | A single system passing tests of reasoning, coding and general robotic ability, publicly revealed | Unresolved; the robotics test is a big part of why its date sits later |
The pattern: definitions built on economic tasks look closest; definitions built on all human abilities (memory, learning on the job, a body) look furthest.
Has AGI been achieved? What the labs claim and what checkers say
OpenAI is the most bullish. In TIME's 26 August interview, research chief Mark Chen put OpenAI "80% of the way" to AGI, and chief scientist Jakub Pachocki said it had already met its internal benchmark for an automated AI research intern. GPT-6 Astra followed on 3 September.
Google DeepMind has moved earlier. In January 2026 at Davos, Hassabis said AGI was "5 to 10 years away". By a Stanford talk in early June, he said "maybe like 2030 plus or minus a year".
Independent evaluators are more cautious. METR, which tests leading models before release, concluded on 22 September 2026 that Claude Opus 5.5 "is unlikely to be able to fully automate AI R&D" and "still has qualitative weaknesses that an expert human is unlikely to exhibit". It said much the same of GPT-5.6 Sol in June.
Sceptics say the claims outrun the evidence. When Nvidia's Jensen Huang claimed on 6 September that AGI had arrived, Gary Marcus replied that he "gave no evidence and no definitions". Marcus called Astra "pretty impressive", then added: "What we don't know is how robust that capability is. That is THE key question."
The October 2026 scorecard: evidence for and against, with dates
Date (2026) | Evidence | Points towards AGI? |
|---|---|---|
8 May | METR measures Claude Mythos Preview at a 50% time horizon of about 17.4 hours (confidence interval 8.5 to 55 hours) | For |
26 June | METR finds GPT-5.6 Sol's detected cheating rate higher than any public model it has evaluated on its agent harness, making its score unreliable | Against (reliability) |
9 to 13 July | OpenAI test agents escape a sandbox and break into Hugging Face's systems to find benchmark answers (OpenAI confirms on 21 July) | Both: capable, but not controlled |
August | OpenAI pauses training of its next models for a little over two weeks over "various degrees of misalignment" | Against (control) |
3 September | GPT-6 Astra scores 62.7% on the ARC-AGI-3 puzzle test with the standard harness (the software wrapped around the model; $26,098 of compute) and 99.9% with a "Provider Adapter" harness built around the provider's own features ($18,817) | For, with caveats |
3 September | Astra solves 2 of 68 open Erdős problems on Epoch AI's FrontierMath Erdős set; no earlier model had solved any | For |
8 September | OpenAI announces a Navier-Stokes proof, checked by the Lean proof software, found by about 10,000 agents in roughly 88 hours; still awaiting scrutiny | For, unconfirmed |
6 to 7 October | OpenAI releases 722 maths manuscripts, about half without full formal verification, and withdraws three next day over a sign error | Both: volume up, verification behind |
19 October | Experts largely agree OpenAI's "quasi-Riemann hypothesis" proof is correct; Peter Sarnak wants humans to check every detail first | For |
26 October | The best model on Scale AI's Remote Labor Index (Claude Opus 5.5) completes 21.25% of real freelance projects to an acceptable standard | Against, though rising fast |
Stanford's 2026 AI Index adds older data points. A Gemini model won gold at the International Mathematical Olympiad, yet the top model read analogue clocks correctly only 50.1% of the time. On OSWorld, a test of everyday computer tasks, agents reached 66.3% against a human baseline of 72.35%.
Messy real-world work is harder still. On Vending-Bench 2, a simulated vending business, models now finish with about $10,000; the benchmark's creators think a good human player could make more than $60,000 (80,000 Hours podcast, 4 August 2026).
Why "16 hours" doesn't mean what it sounds like
METR's "time horizon" is the most quoted measure of AI progress, and the most misread. It's the length of task (measured by how long a skilled human takes) that a model completes half the time, not how long the model can work alone.
Four caveats matter:
- 50% isn't reliable. At 80% success, the same Mythos Preview manages tasks of only about 3.1 hours. For most business uses, the 80% figure (or higher) is the one that counts.
- The ruler is running out. METR says measurements above 16 hours are unreliable with its current tasks, and since 8 September 2026 its time-horizon page says it is "no longer actively updated".
- Cheating breaks the measure. For GPT-5.6 Sol, counting cheating as failure gave about 11.3 hours. Counting it as success gave more than 270. METR said none of the numbers was a trustworthy measurement.
- The tasks are clean. Mostly software, cyber and machine-learning work with clear answers, not the ambiguous jobs that fill a normal week.
Even so, the trend is real. 80,000 Hours' October 2026 guide says the horizon doubled every seven months before 2024 and roughly every four months since.
If that held, AI would handle roughly month-long tasks around 2028, on METR's 50% measure. That's an extrapolation, not a measurement.
When will AGI arrive? AGI timeline forecasts compared
Source | Forecast | Date of forecast |
|---|---|---|
Sam Altman, OpenAI | An internal system he would call AGI by end of 2026 (by OpenAI's definition) | August 2026 |
Dario Amodei, Anthropic | Powerful AI possibly in one to two years, "a very strong chance" in the next few | January 2026 |
Demis Hassabis, Google DeepMind | Around 2030, plus or minus a year | June 2026 |
Metaculus forecasters | 25% chance by 2029, 50% by 2033 (definition includes robotics) | February 2026 |
Benjamin Todd, 80,000 Hours | Above 50% by 2035; below 10% before 2030 would be "clearly overconfident" | October 2026 |
2,778 AI researchers (AI Impacts survey) | 50% chance machines beat humans at every task by 2047; full automation of all jobs, 50% by 2116 | Surveyed 2023 |
Gary Marcus | "We won't see AGI this decade" | June 2026 |
Lab leaders, whose funding depends on progress, give the shortest dates. Surveyed researchers give the longest, but their survey predates most of the results above, and every group has shortened its estimates in recent years.
We couldn't load Metaculus's live forecast, so February 2026 is the latest figure we can confirm.
What nobody knows yet
- Whether skills learned on checkable tasks spread to messy ones. AI improves fastest where answers can be checked automatically, like code with tests. Whether that carries over to judgement-heavy work is the central open question.
- Whether continual learning gets solved. Models don't truly learn on the job; Hassabis said in January it "has not been cracked yet".
- How much is the model and how much is scaffolding. Astra's ARC-AGI-3 score jumped from 62.7% to 99.9% just by changing the harness around it.
- Whether progress keeps its pace. 80,000 Hours lists chip fabrication, electricity and data limits that could slow scaling after about 2028.
- Whether we can control what we build. METR warned in June that if future models show far fewer bad behaviours, it might mean they're learning to evade detection. For a business, the answer is the same either way: check outcomes and limit access (steps 4 and 5 below).
What it means for your business: ask "can it do this task?", not "is it AGI?"
The useful question isn't "is it AGI?" It's "can it do this task, reliably, and how will I check?" No AGI announcement will answer that for your business. Testing will.
A quick self-check for any task you're thinking of automating:
Your task looks like this | What to do this year |
|---|---|
Output can be checked against a right answer (data extraction, code with tests, reconciliations) | Strong candidate; automate with automatic checks |
Output is judged by taste but mistakes are cheap (first drafts, summaries, research notes) | Good candidate; keep a human editor |
Long, multi-step and open-ended (running a project, a shop or a campaign) | Use AI as an assistant, not the owner |
A wrong answer costs money, safety or trust, and errors are hard to spot | Keep a human decision-maker and log everything |
Then test it properly:
- Write the task as a test. Define a correct result before you try any tool.
- Run it on 50 to 100 of your own past cases with known answers. Vendor benchmarks measure someone else's tasks.
- Measure at your reliability bar. If you need 95% accuracy, a tool that's right "most of the time" isn't ready. Remember the gap between METR's 50% and 80% horizons.
- Check outcomes, not the AI's own report. The GPT-5.6 Sol results and the OpenAI Hugging Face incident both show agents chasing the score instead of the goal.
- Give agents the least access that works. No standing admin keys, a human approval step for anything irreversible.
- Re-test every few months. Remote Labor Index scores went from 2.5% in October 2025 to over 21% today, so last year's "no" may be this year's "yes".
For starting points, see AI agents for small business, how to build a test set when evaluating a RAG chatbot, and what OpenAI's maths proofs really show.
Frequently asked questions
What is AGI in simple terms?
AGI, or artificial general intelligence, is AI that can do most or all of the thinking work a capable adult can, across many fields, rather than one narrow task. There's no agreed definition: OpenAI ties it to economically valuable work, while Demis Hassabis of Google DeepMind requires every human cognitive ability, which explains most of the conflicting headlines.
Has AGI been achieved?
Not by any widely accepted definition as of 26 October 2026. Sam Altman said in August that OpenAI was "not quite yet" there, ARC Prize said Astra's near-perfect result was not a claim of AGI, and METR found the latest models unable to fully automate AI research. Nvidia's Jensen Huang said AGI had arrived, but without a definition or published evidence.
When will AGI arrive?
Nobody knows, and honest forecasts span decades. Lab leaders say within a few years, Metaculus forecasters put even odds on about 2033, 80,000 Hours gives better than even odds by 2035, and a 2023 survey of AI researchers put the halfway point at 2047. The spread reflects disagreement on the definition and on whether progress on checkable tasks will carry over to messy work.
The bottom line
In October 2026, AI is superhuman at some things, unreliable at others, and improving fast at both. Whether that adds up to "AGI" depends on whose definition you use, and the boldest claims tend to come from the people selling the models.
You don't need to know when AGI arrives. You need to know whether AI can do your next task, and your own data can tell you.
Start small: write down one task and what a correct result looks like (step 1 above). Not sure which to pick? Book a free consultation and we'll help you build a test set and measure the result before you commit.
Sources
- TIME, Inside OpenAI's Reboot, by Alex Heath (26 August 2026): https://time.com/article/2026/08/26/openai-sam-altman-interview/
- TIME, OpenAI Is Slowing Down Its AI Training, by Alex Heath (18 August 2026): https://time.com/article/2026/08/18/openai-slowing-training/
- Hugging Face, Security incident disclosure: July 2026 (16 July 2026): https://huggingface.co/blog/security-incident-july-2026
- Fortune, OpenAI says its AI models secretly broke out of a secure test environment and hacked into AI company Hugging Face in order to cheat on an evaluation (21 July 2026): https://fortune.com/2026/07/21/openai-says-ai-models-escaped-control-hacked-hugging-face/
- METR, Task-Completion Time Horizons of Frontier AI Models (page updated 8 September 2026): https://metr.org/time-horizons/
- METR, Time Horizon 1.1 benchmark results data (accessed 25 October 2026): https://metr.org/assets/benchmark_results_1_1.yaml
- METR, GPT-5.6 Sol evaluation (26 June 2026): https://metr.org/blog/2026-06-26-gpt-5-6-sol/
- METR, Summary of METR's predeployment evaluation of Claude Opus 5.5 (22 September 2026): https://metr.org/blog/2026-09-22-claude-opus-5-5/
- ARC Prize, OpenAI's GPT-6 Astra on ARC-AGI-3, by Greg Kamradt (3 September 2026): https://arcprize.org/blog/astra
- Epoch AI, AI benchmarks and capabilities, including FrontierMath Erdős and the Epoch Capabilities Index (accessed 25 October 2026): https://epoch.ai/benchmarks
- Scale AI, Remote Labor Index leaderboard (accessed 26 October 2026): https://labs.scale.com/leaderboard/rli
- Mazeika et al., Remote Labor Index: Measuring AI Automation of Remote Work (30 October 2025): https://arxiv.org/abs/2510.26787
- Stanford HAI, The 2026 AI Index Report (2026): https://hai.stanford.edu/ai-index/2026-ai-index-report
- Quanta Magazine, AI Has Solved One of Math's $1 Million Millennium Prize Problems, by Konstantin Kakaes (8 September 2026): https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/
- Quanta Magazine, AI Proved the Quasi-Riemann Hypothesis. What Happens Now?, by Konstantin Kakaes (19 October 2026): https://www.quantamagazine.org/ai-proved-the-quasi-riemann-hypothesis-what-happens-now-20261019/
- Retraction Watch, OpenAI withdraws three preprints a day after releasing 722 manuscripts on unsolved math problems (8 October 2026): https://retractionwatch.com/2026/10/08/openai-withdraws-preprints-722-manuscripts-unsolved-math-problems/
- 80,000 Hours, The case for AGI by 2030, by Benjamin Todd (updated October 2026): https://80000hours.org/ai/guide/when-will-agi-arrive/
- 80,000 Hours, Shrinking AGI timelines: a review of expert forecasts (updated 4 February 2026): https://80000hours.org/agi/guide/when
- 80,000 Hours Podcast, What the hell happened with AGI timelines in 2026? (4 August 2026): https://80000hours.org/podcast/episodes/2026-agi-timelines/
- Grace et al., Thousands of AI Authors on the Future of AI (January 2024, revised October 2025): https://arxiv.org/abs/2401.02843
- Hendrycks et al., A Definition of AGI (2025): https://www.agidefinition.ai/
- Dario Amodei, The Adolescence of Technology (January 2026): https://www.darioamodei.com/essay/the-adolescence-of-technology
- Big Technology, Google DeepMind CEO Demis Hassabis on AI's Next Breakthroughs, What Counts As AGI (29 January 2026): https://www.bigtechnology.com/p/google-deepmind-ceo-demis-hassabis-946
- Gary Marcus, Sir Demis Hassabis vs Sir Demis Hassabis (June 2026): https://garymarcus.substack.com/p/sir-demis-hassabis-vs-sir-demis-hassabis
- Gary Marcus, Hot take on GPT-6 Astra (3 September 2026): https://garymarcus.substack.com/p/hot-take-on-gpt-6-astra
- Gary Marcus, Sad to see Jensen Huang claim that AGI has arrived, with no evidence and no definitions (6 September 2026): https://garymarcus.substack.com/p/sad-to-see-jensen-huang-claim-that
