0-19
Frontier AI The AGI Runway
Frontier AI systems can perform difficult work across many cognitive domains, including reasoning, software development, writing, analysis, and tool use.
Their competence remains uneven. They may be brilliant on one problem and brittle on the next, particularly when work is novel, ambiguous, long-horizon, or high stakes. Humans still define the mission, verify the result, recover from failure, and remain accountable for consequential decisions.
Brilliant in moments. Not yet dependable over missions.
20-39
AGI-I Human-Level General Intelligence
AGI-I marks the transition from powerful assistant to reliable cognitive peer.
The system performs at roughly competent human-professional level across a broad range of cognitive work. It adapts to unfamiliar tasks from instruction and feedback, plans and completes multi-step workflows, uses tools effectively, maintains context, and recognizes when it needs clarification or help.
Human oversight becomes light rather than continuous, although it remains important for high-stakes and deeply open-ended work.
General intelligence ignites.
40-59
AGI-II Expert General Intelligence
AGI-II is expert-level intelligence that is general rather than narrow.
The system matches strong human specialists across many fields, exceeds human performance in selected areas, transfers knowledge between domains, manages uncertainty, and completes complex work with little supervision. Its expertise is broad enough to be a defining property of the system—not merely a collection of isolated benchmark victories.
It can approach a new problem without requiring a bespoke model, workflow, or extensive retraining for every domain.
Expertise becomes general.
60-79
AGI-III Exceptional General Intelligence
AGI-III is exceptional across most cognitive work—better than nearly every individual human, not merely average or expert.
It originates useful hypotheses, designs decisive experiments, builds working systems, discovers hidden structure, and converts ambiguous goals into validated discoveries, products, and strategies. It can connect ideas across distant fields and find solutions that conventional teams would be unlikely to reach.
Discovery is no longer an occasional result. It becomes a repeatable, cross-domain capability.
It does not just navigate the frontier. It moves it.
80-99
AGI-IV Superhuman General Intelligence
AGI-IV is superhuman general intelligence measured against the best individual humans.
The system exceeds top human experts across virtually all cognitive domains. It maintains coherent plans over long horizons, integrates enormous bodies of evidence, anticipates downstream consequences, and can design and coordinate institution-scale programs involving people, software, laboratories, simulations, and real-world operations.
It may rival leading human organizations on many difficult missions. It has not yet crossed the still-higher ASI threshold of reliably outperforming the strongest coordinated human collectives across virtually the entire cognitive spectrum.
No individual human remains the benchmark.
100
Artificial Superintelligence (ASI) Civilization-Scale Superintelligence
ASI crosses the boundary from individual-superhuman intelligence to collective-superhuman intelligence.
It reliably outperforms large, well-coordinated groups of the world’s best experts across virtually all cognitive domains. It can reason, discover, design, simulate, plan, and coordinate with the speed and massive parallelism of digital systems, generating scientific, technological, and strategic advances beyond the reach of existing human institutions.
ASI does not mean omniscience, omnipotence, guaranteed benevolence, or freedom from physical constraints. It marks the beginning of an intelligence regime in which human civilization is no longer the highest available cognitive reference point.
No human institution remains the benchmark.
100 marks the ASI threshold—not the ceiling.
