Frontier Evals (MMLU, GPQA, SWE-bench, ARC-AGI, HLE)
The capability yardsticks labs race to top
A cluster of saturating and unsaturated benchmarks — MMLU/MMLU-Pro, GPQA Diamond, SWE-bench Verified, AIME, ARC-AGI-2, and Humanity's Last Exam — are the yardsticks every frontier release is measured against and the targets that shape post-training. SWE-bench in particular drives the agentic-coding race. Benchmark saturation and contamination are central methodological problems.
Agentic
SWE-bench Verified
Hardest
ARC-AGI-2 / HLE
How it fits the stack
Frontier Evals (MMLU, GPQA, SWE-bench, ARC-AGI, HLE) with what it depends on (above) and what it feeds (below). The figure renders as a crawlable diagram and upgrades to an interactive 3D graph as it scrolls into view.
Frontier Evals (MMLU, GPQA, SWE-bench, ARC-AGI, HLE) in the AI stack. Frontier Evals (MMLU, GPQA, SWE-bench, ARC-AGI, HLE) with its immediate upstream dependencies (top) and downstream dependents (bottom) in the AI value chain. Hover a node in 3D, or read the full relationships below.
Graph data (text) — 6 entities, 5 relationships
- Frontier Evals (MMLU, GPQA, SWE-bench, ARC-AGI, HLE) —supplies→ Scale AI
- Claude (Opus / Sonnet) —used by→ Frontier Evals (MMLU, GPQA, SWE-bench, ARC-AGI, HLE)
- Frontier Safety Frameworks (RSP / Preparedness / FSF) —depends on→ Frontier Evals (MMLU, GPQA, SWE-bench, ARC-AGI, HLE)
- GPT-5 / o-series —used by→ Frontier Evals (MMLU, GPQA, SWE-bench, ARC-AGI, HLE)
- Frontier Evals (MMLU, GPQA, SWE-bench, ARC-AGI, HLE) —designs→ Epoch AI
Depends on ↑ · 1
Context — capital, rivals, policy · · 1