September 29, 2026
Grok Bot, Muse, and Dots Shipped the Same Agent
xAI, Meta, and OpenAI shipped always-on agents 49 days apart with the same architecture, but none of them measures whether the promised learning loop actually improves.
Read articleMeasure the loop
The Benchmark for Recursive AI Improvement
Leaderboard
Score vs. cost under the same evaluation protocol. Each connected sweep shows how much autonomous improvement a model family produced as reasoning effort changes. 100% = a perfect predictor.
ARI Bench pass rates by model
One-hour hidden exam exact-match score. OpenAI reasoning efforts are shown as a ladder; other labs show their strongest tested reasoning setting.
Intelligence per dollar
Benchmark percentage points earned per API-equivalent dollar on a zero-based linear scale. Higher is better; line lengths are directly proportional to value.
| # | Model | Lab | Score | Current cost / run | % / $ | Seeds |
|---|
Updated · API-equivalent costs use current configured rates when available; cost-at-run remains retained in the published data · scores are mean ± SE across seeded runs; rows with fewer than 3 seeds are provisional · full methodology · run analyses: all 50 reports · GPT-6 Astra · 38% · Grok 4.6 · Claude Opus 5
Measuring AI Labs' Acceleration Potential
Human-level benchmarks ask whether AI can do the work. ARI asks whether AI can accelerate the creation of better AI.
01
The next step after human parity
Once systems can reason, code, and research at expert level, the important question becomes whether they can improve the systems behind those capabilities.
02
Measuring the rate of self-improvement
ARI focuses on the recursive cycle: inspect a starting system, find a better path, implement it, test it, and produce measurable improvement.
03
Designed for super intelligence
Every result comes from the same fixed protocol and private scoring target, so progress is comparable without turning the benchmark into a pure scale contest.
Built For Frontier Competition
What ARI rewards
Autonomous model-design judgment: choosing experiments, repairing failures, improving training behavior, and shipping a better system under fixed rules.
- Understand a constrained starting point
- Discover and execute useful improvements
- Convert iteration into verified score gains