Atlas

Benchmarks

Singularity-Bench

Singularity-Bench (CMU Infini-AI Lab and others) asks a coding agent to build six AI-development systems from a minimal starting codebase: a pre-training run, a mid-training data mixture, post-training, an inference engine, an agent harness and a serverless inference-cluster orchestrator. Each system is scored on identical hardware against a mature human-built reference such as OLMo 2, and the Singularity-Index is the mean over the six tasks of the agent's score divided by the reference's (inverted for perplexity), so 1.0 matches the human references on average. A failed task scores 0 and ratios are not capped. Runs use the Breakwater extension of Harbor to block reward hacking.

Models
6
Top score
1.3
Leader
Claude Opus 5.5
Updated
Nov 12, 2026
Resolution
—
Saturation
—

No signal or difficulty estimate. This benchmark has no declared score range, so it cannot be placed on the 0-1 scale the capability curve requires.

Organizations
6 dated models · 3 frontier modelsClosed weightsOpen weights
Data: Carnegie Mellon Universityatlas.village.dev

Singularity-Bench leaderboard

Loading Atlas data…