Is ARC-AGI-3 still meaningful?

Known problems

• Every frontier model scored under 1% at launch in March 2026.^{[3][3] ARC-AGI-3 launch OfficeChai Secondary}
• A provider adapter keeps the model's hidden reasoning state between turns. With it, the same model reached 99.9%.^{[2][2] GPT-6 Astra on ARC-AGI-3 ARC Prize Primary}
• OpenAI's launch table uses OpenAI's own adapter.^{[4][4] GPT-6 Astra launch OpenAI Vendor}
• Open and Chinese models stay low, with DeepSeek V4.1 Flash at 5.8%.^{[1][1] ARC-AGI-3 leaderboard data ARC Prize Primary}
• ARC Prize says the result is not proof of AGI, and it names no ARC-AGI-4 yet.^{[2][2] GPT-6 Astra on ARC-AGI-3 ARC Prize Primary}

What to look at instead

The report names no direct successor.

Where it appears in the report

Chapter 1, Benchmarks

• The harness and the host can move a score more than the model
• Figure 7. The same model gets very different scores in different harnesses
• Table 2. First model to cross each threshold in “Go deeper: The full saturation timeline (14 benchmarks)”
• Table 5. Reasoning and knowledge benchmarks: what they measure, who maintains them, the top score on 5 November 2026 and the known flaw in “Go deeper: Reasoning and knowledge benchmarks (17)”
• Figure 14. OpenAI and Anthropic share only five headline benchmarks
• Figure 15. What the 2026 system cards add