ARC Prize: GPT 6.2 Astra Is the First Model to Score 100% on ARC AGI 3
GPT-6.2 Astra from @OpenAI on ARC-AGI (Verified): - ARC-AGI-3: 100.0% (SOTA), $4.1K (provider adapter), 70.3%, $24.2K (standard) - ARC-AGI-2: 95.8% (SOTA), $0.50/task - ARC-AGI-1: 98.5%, $0.17/task GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3.
Context
An official benchmark group posted verified scores, including a 100% result on its hardest test under a provider adapter and a lower score on the standard setting.
ARC Prize posted what it called verified scores for OpenAI's GPT 6.2 Astra, the same model OpenAI launched earlier in the day. It said Astra is the first model to score 100% on ARC AGI 3, and it also reported 95.8% on ARC AGI 2 and 98.5% on ARC AGI 1. The ARC AGI 3 line listed two results: 100% at $4.1K with a provider adapter, and 70.3% at $24.2K on the standard setting.
Deep dive
ARC Prize, which runs the ARC AGI tests, posted a verified scorecard for GPT 6.2 Astra on the same day OpenAI launched the model. The group said Astra is the first model to achieve a 100% score on ARC AGI 3, and it marked the 100% and the 95.8% ARC AGI 2 results as new state of the art.
The post lists three tests. On ARC AGI 3 it reports 100.0% at $4.1K with a provider adapter, and 70.3% at $24.2K on the standard setting. On ARC AGI 2 it reports 95.8% at $0.50 per task. On ARC AGI 1 it reports 98.5% at $0.17 per task. Those figures, labels, and the first to 100% claim are ARC Prize's own verified results, not OpenAI's blog scores.
The two ARC AGI 3 numbers matter because they are not the same result. The 100% score is the one ARC Prize ties to the provider adapter and the lower cost. The standard setting is much more expensive in the post and comes in at 70.3%. This file does not have a longer technical note from ARC Prize that defines those two settings, so readers should not treat 100% as the score under every listed condition.
The launch story for Astra and Sol, led by OpenAI's own announcement, is the product context. This item is the official third party verification of a specific benchmark result for Astra. Sol is not scored in the ARC Prize post.
What to expect next
Watch whether ARC Prize publishes a fuller note on the provider adapter versus the standard setting, and whether other labs post verified ARC AGI 3 scores that close the gap. Independent reruns will show if the 100% figure holds only under the cheaper adapter condition. OpenAI may also post Sol results on the same tests. Any later correction or extra cost table from ARC Prize should be read back into this story.
