YC Paper Club reveals how the same Claude Opus model scores 30% on ARC-AGI bare but reaches 95% with proper Harness, and NVIDIA's AVO hits 100%, proving agent runtime—not model weights—determines the performance ceiling.
OpenAI's GPT-6 Astra integrates reasoning, computer operation, and continuous execution into a single model, scoring 72.6% on OSWorld 2.0, 57.9% on Terminal-Bench 4.0, and 99.9% on ARC-AGI-3 with a Provider Adapter, while demonstrating autonomous tax filing, CRM updates, code migration, and zero-day vulnerability discovery — all with new cross-context memory and safety boundaries.
A 10‑million‑parameter GRAM model from Bengio, KAIST, Mila and NYU achieves 97% accuracy on Sudoku‑Extreme and competitive scores on ARC‑AGI tasks by replacing deterministic recursive updates with a probabilistic multi‑trajectory process, and extensive ablations show that both random guidance and depth‑supervised training are essential for its performance.
The GRAM model introduced by Yoshua Bengio’s team replaces deterministic recursive updates with probabilistic multi‑trajectory sampling, enabling a 10 M‑parameter network to achieve 97 % accuracy on Sudoku‑Extreme, 52 %/11 % on ARC‑AGI, and near‑perfect results on N‑Queens and graph‑coloring, while also supporting unconditional generation tasks.
While recent model releases boast near‑perfect scores on benchmarks like MMLU and HumanEval, the ARC‑AGI‑2 and ARC‑AGI‑3 leaderboards expose a stark gap between headline numbers and genuine programming intelligence, highlighting cost, fluid reasoning, and real‑world applicability.
Anthropic’s Claude Opus 4.6 launches with a 68% ARC‑AGI score, a 1 million‑token context window, top rankings on Terminal‑Bench 2.0, Humanity’s Last Exam, and GDPval‑AA, unchanged pricing, enhanced safety, and new API features such as adaptive thinking and context compression.
