GPT-6 Luna scores 59.3% on ARC-AGI-2, matching GPT-5.6 at ~62% lower cost
scottleibrand · x · 2026-09-23
ARC Prize published verified ARC-AGI results for OpenAI's GPT-6 Luna:
- ARC-AGI-3: 0.19% ($241, standard harness) / 0.59% ($237, provider adapter harness)
- ARC-AGI-2: 59.3% at $0.062/task
- ARC-AGI-1: 86.7% at $0.018/task
Overall, GPT-6 Luna matches GPT-5.6 Luna's scores at roughly 62% lower cost.
More from Models
- Andriy Burkov: Codex Is Infinitely Faster Than Any Open-Weight Agent, But Priced Out of Reach — burkov · 2026-09-23
- Viral demo claims 'GPT-6' can drive browser Paint to draw, unverified — alexcovo_eth · 2026-09-23
- Opus 5.5-generated three.js spell demo wows with procedural VFX and sound — majidmanzarpour · 2026-09-23
- Early hands-on with rumored Opus 5.5 in Scenario's Blender plugin stuns users — repligate · 2026-09-23
- GPT-6 Sol's price cut means it should be compared to Sonnet, not Opus — TraditionalHome8852 · 2026-09-23
- User burns two OpenAI 20x accounts down to 13%, sizing up GPT-6 efficiency — jdjohnson · 2026-09-23