ARC Prize verifies GPT-6.2 Astra at 100% on ARC-AGI-3 with provider adapter; 70.3% on standard harness

AI Benchmarks Laurenceo Intelligence ARC Prize Foundation / OpenAI · Global

A 30-point spread between two verified runs of the same model means scaffolding and state management now move headline scores as much as the model does. Enterprise buyers and procurement teams comparing vendor claims should require the evaluation setup, cost per run and harness type alongside any score. Watch whether the standard-harness leaderboard, not the adapter result, remains the figure that the market treats as the reference.

The Sierra supercomputer at Lawrence Livermore National Laboratory
The Sierra supercomputer at Lawrence Livermore National Laboratory. Photo: U.S. Department of Energy from United States / Wikimedia Commons, Public domain

What Happened

ARC Prize posted verified results for OpenAI's GPT-6.2 Astra: 100.0% on ARC-AGI-3 using a provider adapter at about $4.1K, and 70.3% on the standard harness at $24.2K. It also lists 95.8% on ARC-AGI-2 at $0.50 per task and 98.5% on ARC-AGI-1, and says Astra is the first model to reach 100% on ARC-AGI-3.

Why It Matters

A 30-point spread between two verified runs of the same model means scaffolding and state management now move headline scores as much as the model does. Enterprise buyers and procurement teams comparing vendor claims should require the evaluation setup, cost per run and harness type alongside any score. Watch whether the standard-harness leaderboard, not the adapter result, remains the figure that the market treats as the reference.

“The same AI model scored 100% in one test and 70.3% in another. The way a test is set up can move the score just as much as the model itself does.”

Jorai LaurenceoJorai Laurenceo
Executive Director, Laurenceo

Sources

More From This Briefing

More on Frontier AI

All Frontier AI coverage →

Key Terms

  • Artificial general intelligence (AGI): Artificial general intelligence refers to AI systems that match or exceed human ability across most cognitive tasks. There is no single...
  • Model evaluation: Model evaluation is the testing of an AI model's capabilities, reliability and risks, such as accuracy, bias, security and dangerous...

Filed under Frontier AI Models, ARC Prize Foundation, OpenAI.

This story is part of the Laurenceo Intelligence briefing for 19 November 2026, a daily sourced briefing on AI, technology and enterprise risk published in Bermuda by Laurenceo and edited by Jorai Laurenceo. It is produced by Laurenceo's proprietary analytical systems, with every source verified before publication; how this is made. General information only, not legal, regulatory or financial advice.