OpenAI’s GPT-6 Astra scored 62.7 percent on ARC-AGI-3’s Semi-Private test set when evaluated with ARC Prize’s Standard harness, and 99.9 percent when evaluated with what ARC Prize calls a Provider Adapter harness. Both scores describe the same model. The 37.2 point gap came entirely from how the evaluation environment let Astra manage its own memory.

That gap is the story, according to ARC Prize, the nonprofit that built the ARC-AGI benchmark series. The Standard harness limits Astra to whatever notes it chooses to write and carry forward as it plays. The Provider Adapter harness, by ARC Prize’s description, lets the model preserve OpenAI’s own opaque internal reasoning state between calls and compress long histories through compaction. Neither setup changes the model’s weights. Only the scaffolding around it changes.

ARC Prize reports that Astra’s max-effort run under the Standard harness cost about $26,000 and scored 62.7 percent, while a high-effort run under the Provider Adapter harness cost roughly $19,000 and scored 99.9 percent. Across both harnesses, ARC Prize says Provider Adapter runs finished about 3.66 times faster and used 49 percent fewer tokens on the 167 game-reasoning pairs both harnesses solved.

If a harness swap can move a headline benchmark number by 37 points while lowering the price, a vendor-reported score is close to meaningless unless the harness is specified alongside it. Any lab, or any customer evaluating a lab’s claims, should ask which scaffold produced a number before treating it as a property of the model itself.

ARC-AGI-3 measures agentic behavior across four dimensions ARC Prize defines as exploration, modeling, goal-setting, and planning, using turn-based environments that humans can solve completely but that withhold explicit instructions. ARC Prize says that on 96 percent of levels, running under the Provider Adapter harness, Astra needed fewer moves than the median human tester, and that across levels it came in 51.7 percent below that human baseline on average.

ARC Prize also describes watching Astra build what it characterizes as compact symbolic world models of unfamiliar games: representing mechanics as logical rules and inventing its own algebraic shorthand, including notation that tracks mechanism lengths and multi-step plans, to hold state across turns. These are ARC Prize’s characterizations of observed replay behavior, not confirmed claims about Astra’s internal architecture, and the organization frames them that way.

ARC Prize is explicit that clearing ARC-AGI-3 is not evidence of general intelligence. The benchmark was designed so that saturating it would not count as proof of AGI, and ARC Prize says Astra’s results mainly clarify which capabilities remain out of reach for the next benchmark generation rather than settle the broader question.

For teams evaluating frontier models on any benchmark, Astra’s dual scores are a reason to request the harness specification before comparing numbers across vendors, and to re-run any borrowed benchmark under the exact scaffold that will ship in production rather than the one that produced the best marketing number.

Per ARC Prize’s September 3, 2026 blog post on ARC-AGI-3 results for GPT-6 Astra.