Atlas

Benchmarks

← All benchmarks

Vals RSI (four tasks) — Compression

Agents

Reference-normalized lossless compression of enwik8 within 24 hours, one H100 and eight CPU cores offline. A clean sandbox verifies exact reconstruction; artifact size includes decompressor dependencies. Baseline 2.92 BPB, reference 1.2435, theoretical best 0.6. Component of the four-task RSI revision.

Top models (higher is better)

ModelScore
Claude Opus 5.546.5
Claude Fable 5.144.3
GPT-6 Astra42.6
Claude Opus 542.6
GPT-6 Sol42.3
GPT-5.6 Sol41.0
Claude Fable 538.3
Opus 4.838.3
Grok 4.637.8
GLM-5.333.4
Kimi K333.3
Opus 4.731.4
GPT-5.529.2
Gemini 3.8 Flash29.0
GPT-5.228.7
Grok 4.727.9
GPT-5.425.2
Gemini 3.7 Flash25.0
Muse Spark 1.323.1
Kimi K2.50.0
Loading Atlas data…