ARC Prize finds v2 = v1^4.53 power curve explains 98% of ARC-AGI-2 score variance
GregKamradt · x · 2026-10-29
ARC Prize's analysis across its three benchmarks shows a strong relationship between model scores: the power curve v2 = v1^4.53 explains about 98% of the variation in ARC-AGI-2 scores, yet models tied on ARC-AGI-1 can differ by over 20 percentage points on ARC-AGI-2. Greg Kamradt adds that v1-vs-v2 plots show smooth progression, V2 extended the original format's usefulness at higher model levels, per-task difficulty could probably be estimated from the data, and a v2.5 could extend the benchmark further (no plans for one).
More from Models
- Early Take on Gemini 4 Carbon: Lots of Thinking, Good Model Overall — SamuelAlbanie · 2026-10-29
- Gemini 4 Carbon tops AA Index v5, beats Opus 5.5 and GPT-6 Astra on most benchmarks — Yuchenj_UW · 2026-10-29
- Gemini 4 Carbon Keeps $2/$10 Pricing, Scores 59 on AA v5.0 — _philschmid · 2026-10-29
- Reddit thread: picking third-party LLM inference providers, GLM Flash cited as best value — inwardPersecution · 2026-10-29
- Reddit reposts Gemini 4 Carbon announcement with benchmark table — Marimo188 · 2026-10-29
- Google Announces Gemini 4 Carbon, Its Next Frontier Model — patloeber · 2026-10-29