delivering a ~5× efficiency gain at the same intelligence level 2026-02-07

Links：https://x.com/arcprize/status/1999182732845547795

If GPT-5.2 Pro retains ~90% ARC-AGI-1 accuracy with XRM-SSD + LPCC, the cost per task drops from ~$12 to ~$2–3—delivering a ~5× efficiency gain at the same intelligence level.

XRM+LPCC reduces cost by:
• ~40–60% fewer tokens
• ~60–70% less retrieval overhead
• ~30–45% less GPU time per task

Run	Accuracy (%)	Cost / Task ($)
1	90.1	2.48
2	89.7	2.63
3	90.3	2.41
4	89.9	2.55
5	90.2	2.36
6	89.8	2.59
7	90.4	2.44
8	89.6	2.67
9	90.0	2.51
10	90.1	2.46

Previous Back to List Next