Claude Code + Opus 5 saturates ARC-AGI-3 via falsifiable predictions
scaling01 · x · 2026-08-20
A developer achieved 100% RHAE on the ARC-AGI-3 benchmark using a general-purpose coding agent (Claude Code + Opus 5) without additional harness.
The key idea was simple: force a falsifiable prediction before every action.
This turns every move into an experiment and every miss into a precise correction, allowing the agent to learn the game mechanics by being wrong on the record, effectively saturating the benchmark.
Related event: Claude Opus 5 with Claude Code Achieves Perfect Score on ARC-AGI 3(2 posts)→
More from coding & agent
- Decision models match LLMs on routing and safety — and are 7-8x faster — randal_olson · 2026-10-16
- Perplexity open-sources SPACE sandbox, partners with CoreWeave, Akamai, Crusoe — AravSrinivas · 2026-10-16
- LLMs beat decision models on knowledge, but not on routing or safety — randal_olson · 2026-10-16
- Temporal Serverless Workers on Cloud Run: queue-driven autoscaling, scale to zero — rseroter · 2026-10-16
- Meta's Superintelligence Lab Runs 200+ Direct Reports—A Lesson for Flat Multi-Agent Systems — peterjliu · 2026-10-16
- LangChain CEO: the most interesting agent harness work lives at the harness-infra boundary — hwchase17 · 2026-10-16