Forrester says the first of four AGI stages has been reached. It calls the current frontier "competent AGI", not superintelligence. Jensen Huang of Nvidia posted that AGI had arrived on 6 September. Gary Marcus replied that Huang gave no evidence and no definitions. Forrester argues both sides failed to state criteria. Its own criteria, published in August 2025, require self-learning, collaboration and tool-building. It scores OpenAI's GPT-6 Astra as passing tool-building and, narrowly, self-learning. Astra wrote its own Python code mid-game on the ARC-AGI-3 test. That learning vanished at the end of each run and needed a special memory setup. Forrester calls collaboration the weakest leg, since models do not negotiate trade-offs or adapt over months.
What changed
Forrester had forecast stage-one AGI arriving somewhere between 2026 and 2030.
What it unlocks
Labelling current frontier systems as competent AGI in planning discussions, using published criteria.
- ARC-AGI-3: 62.7% with ARC harness
- 99.9% with OpenAI harness
- prior frontier models scored 1%
- stage one of four, forecast 2026-2030
What you need to act on it
- Forrester client access for the full report
Sources