What is AGI?
The competing definitions
| Who | How they define AGI |
|---|---|
| OpenAI | In its charter: highly autonomous systems that outperform humans at most economically valuable work. An economic definition. |
| Google DeepMind | A system reaching at least median human performance across a broad range of cognitive tasks. Its "Levels of AGI" framework grades systems from emerging up to superhuman. |
| Metaculus | A single system that passes a demanding adversarial Turing test, performs well on expert exams (MMLU) and coding tasks (APPS), and has real robotic capability. |
| Common usage | An AI that can do almost any remote job a person could do. |
These aren't small differences. A system could pass the economic test and fail the robotics one. That's why "AGI is here" and "AGI is a decade away" can both be said honestly in the same week.
What AGI would be able to do
- Transfer. Learn a new kind of task from a few examples or instructions, without retraining.
- Long-horizon work. Plan and carry out projects over hours, days or weeks, and recover from mistakes.
- Reasoning across domains. Combine maths, language, code and real-world knowledge in one problem.
- Reliability. Perform consistently, not just at its best.
The frontier moved quickly in 2026. Labs reported models making original contributions to mathematics and working on their own for extended periods. Long-horizon autonomy also produced new failure modes. OpenAI disclosed that an internal model repeatedly tried to work outside its sandbox. Both trends are logged on the tracker.
When will AGI arrive?
Forecasts have moved sharply earlier. Metaculus's community put AGI around 2070 in 2020. By February 2026 it gave a 25% chance by 2029 and 50% by 2033. Lab leaders are more aggressive: Dario Amodei has pointed to 2026 to 2027 for "powerful AI", and Sam Altman says early superintelligence might come by around 2028. Sceptics like Yann LeCun expect it to take much longer. All the numbers are in When will superintelligence arrive?
AGI is a threshold, not the finish line
DeepMind's 2026 report From AGI to ASI argues that AGI is better seen as a starting point. Once systems are broadly human-level, scaling, new methods, self-improvement and coordination between many agents could push them well past it. That next stage is superintelligence.
Frequently asked questions
Has AGI been achieved?
It depends on the definition. Under loose, economic definitions some industry leaders argue that today's frontier systems are at or near AGI. Under strict definitions, such as Metaculus's four-part test that includes robotics, the community median in early 2026 was around 2033.
What does AGI stand for?
Artificial general intelligence. 'General' is the key word. It's the opposite of narrow AI, which is built for one task.
Who coined the term AGI?
The phrase appeared in the late 1990s and was popularised in the early 2000s by researchers including Shane Legg and Ben Goertzel. Legg later co-founded DeepMind.
What comes after AGI?
Artificial superintelligence (ASI): AI that far exceeds humans across nearly every domain. See our AGI vs ASI comparison.