1
ARC-AGI-1
- Human baseline
- 98%
- Baseline protocol
- Original ARC human baseline
- Machine results
- 12
- Result span
- 2020–2025
Appendix D · Edition registry
Three benchmarks. Three distinct protocols.
ARC-AGI-1, ARC-AGI-2, and ARC-AGI-3 are related generations, but their scores are not interchangeable. This index keeps each benchmark with its own human baseline, protocol, evaluation set, and chapter.
3
Benchmark editions
25
Structured evaluations
22
Machine results
3
Human baselines
BenchmarkEdition index
Each benchmark on its own terms.
Open a chapter for the full argument, or use the Evaluations appendix for individual protocols and evidence.
2
ARC-AGI-2
- Human baseline
- ~100%
- Baseline protocol
- Every evaluation task solved by at least two humans
- Machine results
- 6
- Result span
- 2025
3
ARC-AGI-3
- Human baseline
- 100%
- Baseline protocol
- First-contact human calibration
- Machine results
- 4
- Result span
- 2025