Forrester Says 'Stage One' AGI Arrived in 2026 — a Year Early, and Still Just 'Barely Competent'

Forrester researcher Brian Hopkins is calling it: “stage one” AGI arrived in 2026, a year earlier than Forrester’s own 2025 prediction of a 2026-2030 window, and he’s holding the industry to Forrester’s predefined criteria rather than letting the goalposts move now that a system has cleared them. The headline evidence is OpenAI’s Astra system on the ARC-AGI-3 benchmark: prior frontier models scored around 1% accuracy, Astra reached 62.7%, then 99.9% once paired with an optimized memory harness. Forrester’s competent-AGI bar requires software to “autonomously act in pursuit of goals across domains by learning new skills” while “collaborating with humans and machines,” and Hopkins scores Astra against three dimensions — tool-building (passed, it wrote Python libraries independently), self-learning (barely passed, confined to a single session), and collaboration (weakest, no long-term preference learning). His recommendation: retire “frontier model” language for “competent AGI,” and start building contingency plans for “stage two” independent AGI, which Forrester expects within roughly 12 months.

Not everyone reaches for the same label with the same confidence. Jason Lemkin’s own recap of a 20VC crossover conversation the same week has investor Rory O’Driscoll and Jensen Huang’s AGI declaration treated as far less settled — the more meaningful signal, in that conversation, is code-automation economics, a roughly half-trillion-dollar market where AI’s value is actually measurable today rather than a benchmark score.

That gap between benchmark claim and measured reality is exactly what OpenAI’s own audit of a different benchmark, SWE-Bench Pro, warned about two months earlier: OpenAI found roughly a third of that benchmark’s 731 tasks were broken in ways that misrepresented what a model could actually do, and formally retracted its own recommendation to adopt it. Forrester’s ARC-AGI-3 numbers may hold up better under scrutiny than SWE-Bench Pro did — but the lesson generalizes regardless of which specific benchmark is in play: a leaderboard score is not a capability claim until someone has checked whether the underlying tasks measure what they say they measure. Any sales org setting AI-capability roadmaps off a benchmark number owes itself that same audit before it recalibrates a hiring plan or a vendor bet around one.