GPT-6-Astra-Max hits 62.7% on ARC-AGI-3, up from 0% six months ago at $26k per run
BrennusSokol · reddit · 2026-09-26
A Reddit user followed up on a six-month-old RemindMe: back on March 25, no model scored even 1% on ARC-AGI-3, and a commenter asked whether 75% at $2 would ever happen.
Six months later:
- GPT-6-Astra-Max scored 62.7% with no harness, though at a cost of about $26,100 per run.
- With a lightweight memory adapter the benchmark is effectively saturated, still at $17k+; the adapter preserves opaque reasoning state between requests and uses compaction for longer conversations, letting the model reuse prior work instead of reconstructing it.
- Citing Epoch AI's research on plunging reasoning prices, the author argues models will beat ARC-AGI-3 cheaply soon if the trend holds.
More from Models
- Micius-Flash: 27B Multimodal Model for Hardware and Semiconductor Agents — zhiman-ai · 2026-11-11
- Claude IP risk-check site goes viral as users audit their IPs ahead of new rules — vista8 · 2026-11-11
- Baidu says RSI auto-optimized ERNIE inference, 6.24x decode speedup on Kunxin P800 — xiaohu · 2026-11-11
- Baidu's ECHO: multi-token prediction reusing the last layer as a drafter — xiaohu · 2026-11-11
- ERNIE 5.5 Flash Preview: 280B-param multi-router MoE activating only 14B per token — xiaohu · 2026-11-11
- ERNIE 5.5 recreates the Pelican test: 117-second animation drawn frame-by-frame in Python — xiaohu · 2026-11-11