Understand this story
On November 17, 2026, ARC Prize announced the evaluation results of Inception AI's Mercury 3 Preview on the ARC-AGI benchmark. It scored 61.8% on ARC-AGI-2 at a cost of $0.06 per task; 88.5% on ARC-AGI-1 at a cost of $0.03 per task; and only 0.2% on ARC-AGI-3. In the instant reasoning tier, without using reasoning tokens, the model averaged 981 tokens/second output on the public ARC-AGI-2 dataset, with a peak of 1,743 tokens/second for a single call (3,001 tokens output in 1.72 seconds); with reasoning enabled, it averages 500-600 tokens/second. ARC Prize said this is the first diffusion model it has evaluated.
AI-generated from public reports
Latest development · Report
ARC Prize's hands-on test of Inception AI's Mercury 3 Preview…Developments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 17
ARC PrizeFirst-party
ARC Prize's hands-on test of Inception AI's Mercury 3 Preview…
ARC Prize tested Inception AI's Mercury 3 Preview on the ARC-AGI benchmark: it scored 61.8% on ARC-AGI-2 at $0.06 per task, 88.5% on ARC-AGI-1 at $0.03 per task, and only 0.2% on ARC-AGI-3. With the instant reasoning tier enabled, the model averaged 981 tokens/second output on the public ARC-AGI-2 dataset, with a peak of 1,743 tokens/second in a single call (3,001 tokens output in 1.72 seconds); with reasoning enabled, it averaged 500-600 tokens/second. ARC Prize says this is the first diffusion model it has evaluated.
ARC PrizeFirst-partySelected
ARC Prize announces evaluation results of Mercury 3 Preview on ARC-AGI
ARC Prize announced the evaluation results of Inception AI's Mercury 3 Preview on ARC-AGI: a score of 61.8% on ARC-AGI-2 at $0.06 per task, 88.5% on ARC-AGI-1 at $0.03 per task, and only 0.2% on ARC-AGI-3. In the instant reasoning tier, no reasoning tokens are used; on the public ARC-AGI-2 dataset, the model averages 981 tokens/s of output, with a peak of 1,743 tokens/s in a single call (3,001 tokens output in 1.72 seconds); with reasoning enabled, it averages 500-600 tokens/s. ARC Prize states that this is the first diffusion model it has evaluated.
AI Trends