Is Astra really 99.9% on ARC-AGI-3?

What was claimed?

ARC Prize reports 62.7% on ARC-AGI-3 with its standard harness; the 99.9% result used a provider adapter and different reasoning effort.

What does the evidence support?

Independent evidence, scope limited

Compare the same reasoning effort before attributing gains to the model alone.

What are the limits?

Correction, September 26: earlier dashboard labels said ARC-AGI-2. The source specifies ARC-AGI-3. Configuration differences prevent a matched replication-gap conclusion.

Claim or evidence date: 2026-09-03. Last verified: 2026-09-26T20:03:37.888Z.

ARC Prize · Astra evaluation ↗