AGI: 5% for 2026 and 2029 on the horizon

On 6 November, specialist publication The Decoder published its Frontier Radar #5: 45 neolabs, research labs founded by researchers who have left OpenAI, Google DeepMind or Meta, with billions raised and, in several cases, no product. Their shared doubt is not whether powerful AI will arrive, but whether the current recipe — ever-larger language models trained on internet data and refined with reinforcement learning — is enough to reach AGI.

The paradox is in the tone. After the launch of GPT-6 Astra in September, Nvidia chief executive Jensen Huang wrote on X that «AGI has arrived», and OpenAI president Greg Brockman welcomed the «AGI era». There is no agreed definition of AGI. That is why those statements coexist with prediction markets that give only a 5.1% chance that OpenAI announces AGI before 2027, and with the ARC Prize Foundation, which called Astra meaningful progress but made clear it was not claiming the model was AGI.

What AGI means, and why the dates do not match

The first difficulty is definitional. Demis Hassabis, chair of Google DeepMind and Alphabet's chief scientist, talks of «a system with the full cognitive range of the human brain». OpenAI's charter describes AGI as «highly autonomous systems that outperform humans at most economically valuable work». Dario Amodei, chief executive of Anthropic, points to AI «smarter than a Nobel Prize winner across most relevant fields». Yann LeCun, founder of AMI Labs, rules out human-level AI coming from today's language models.

This is not an academic distinction. A system that beats a human at most digital work is not the same as one that can research like a Nobel laureate, and neither is the same as general intelligence in Hassabis's sense. That is why a 2027 date can be optimistic under one definition and reasonable under another. The FourWeekMBA tracker, which collects 35 dated forecasts, puts it plainly: the people building AI are not forecasting the same thing.

Where we are

In capability, the frontier is moving. In 2026 came models such as Gemini 4 Carbon, which Google first limited to cyber defenders, and GPT-6 Cyber, which OpenAI restricted to verified customers. These are jumps in specific tasks — cybersecurity, code, multi-step engineering — not proof of general intelligence.

Measurement does not help either. In March, RAND published a report synthesising expert surveys, prediction markets, compute-centric models and scenario analysis. Its conclusion: forecasts have shifted earlier, but the forecasting infrastructure is «immature». There is a lack of resolved forecasts for calibration, benchmarks that do not saturate and independent validation of influential models.

The indicator RAND considers most relevant is different: the extent to which AI accelerates AI research itself. That is the signal that could produce discontinuities, and frontier labs are best placed to measure it.

What the forecasts say

With those caveats, the map of dates is useful. The FourWeekMBA tracker, updated on 2 November, collects 35 dated forecasts. The median of the latest from ten people who gave a year is 2029. Hassabis told the Wall Street Journal on 21 October that AGI would arrive «within two to three years»; Altman told TIME in August that by the end of 2026 OpenAI would have «an internal system he would call AGI»; Amodei speaks of «one to two years, maybe one to three years» for «a country of geniuses in a data center».

The Metaculus community, which requires tests and a robotics component, had its median at 18 September 2030 (middle half: March 2028 to June 2036), according to FourWeekMBA's reading on 2 November. The largest survey of AI researchers — 2,778 authors, published in 2024 — put a 10% chance on machines outperforming humans at every task by 2027 and a 50% chance by 2047.

The map: who can pay for the race

The GDP-per-capita map does not measure AGI. It shows the economic base of the countries where labs and investment concentrate: most of the labs profiled in detail by The Decoder are in the United States and Europe. The choropleth helps contextualise the economic base of those labs and the investment that funds them; it does not explain on its own when AGI will arrive or who will control it.

The data come from the World Bank (NY.GDP.PCAP.CD, 2025), via Our World in Data. Singapore leads at $139,593 current per capita; Burundi is bottom at $1,067. The gap between them is more than a hundredfold.

What comes next

Ten individual forecasts in the agitracker.io dataset close their window before the end of 2027; eight come from lab leaders. They include Altman's internal system by the end of 2026 and Musk's AI «smarter than any human» for 2026-2027. The next 18 months will put those forecasts to the test.

In the meantime, the race will keep being measured in corporate announcements, not an independent test. Prediction markets already warn that a company declaring AGI is one thing and a system the scientific community recognises as AGI is another. That difference, not the date, is what to watch.