StateBench
Reported model scores on StateBench, best parseable score first. Each row keeps its verbatim score, test conditions and provenance, and links to the source coverage it was extracted from. Benchmark profile →
| Model | Score | Conditions | Provenance | Measured | Source |
|---|---|---|---|---|---|
| Gated DeltaNet variant (275-million-parameter models) | near perfect | Direct training tests: stayed near perfect up to roughly 10 times the longest training length on the hardest general finite state-tracking level | vendor-reported | — | coverage → |
| GPT-6 Luna | 13% or lower | Accuracy dropped to 13% or lower on the longest general-finite state-tracking sequences | vendor-reported | — | coverage → |
| GPT-6 Sol | 13% or lower | Accuracy dropped to 13% or lower on the longest general-finite state-tracking sequences | vendor-reported | — | coverage → |
| Claude Opus 5.5 | 13% or lower | Accuracy dropped to 13% or lower on the longest general-finite state-tracking sequences | vendor-reported | — | coverage → |
| Gemini 3.8 Flash | 13% or lower | Accuracy dropped to 13% or lower on the longest general-finite state-tracking sequences | vendor-reported | — | coverage → |
Scores are only comparable within one benchmark under matching conditions — results under different test setups, and scores from other benchmarks, are not directly comparable.