TwinBench
TwinBench measures whether an AI system can behave like a real personal AI assistant: remember, act, follow up, stay safe, and operate over time.
Can your personal AI assistant beat TwinBench? Start with the reference result, then run the demo or benchmark your own system.
Current public board
The public board shows the current reference result and challenge-worthy artifacts. Historical and degraded runs stay available, but they do not dominate first impression.
| Assistant | Class | Tier | Score | Coverage | Date |
|---|---|---|---|---|---|
| Nullalis Reference Runtime | Reference Runtime | Production-Grade | 75.9 | 84% | 2026-03-25 |
| TwinBench Demo Runtime | Demo Fixture | Emerging | 54.4 | 69% | 2026-03-25 |
What current benchmarks miss
Most AI benchmarks still measure chat quality, coding skill, or one-shot tasks. TwinBench focuses on the missing category: personal AI assistants that persist and operate over time.
10-minute path
Use the repo, the demo runtime, or hand TwinBench to Codex, Claude Code, or Cursor with one prompt.
make demo make site
Evidence-first by design
- unsupported is not failure
- missing bootstrap is not poor scale
- same-user contention is diagnostic
- measured coverage matters
Can your assistant beat this?
Every serious result should be artifact-backed, shareable, and easy to challenge in public.
GitHub first
TwinBench starts with Discussions, issues, and submissions on GitHub. That keeps the benchmark legible and searchable while the ecosystem is still forming.
Monthly challenge shape
Reference Runtime, Top Score This Month, Most Improved, and Verified Artifact are the first public loops. The goal is recognition without hype.