ARC Prize reports GPT-6 Astra scores 62.7% or 99.9% on ARC-AGI-3 by harness

• Astra's best reported ARC-AGI-3 runs are separated by 37.2 points: 62.7% under the Standard harness and 99.9% under the Provider Adapter, as ARC Prize's harness update reports.
• The Standard interface carries forward visible notes selected by the model, while the Provider Adapter preserves opaque reasoning state across requests and compacts long conversations, as explained in a harness explainer.
• The Provider Adapter also recorded 3.66 times less aggregate elapsed time and 49% fewer tokens across game-reasoning pairs both setups solved, according to an ARC leaderboard chart and ARC Prize's full analysis.
• OpenAI's launch framing foregrounded a 99.9% ARC-AGI-3 result in the launch announcement, while ARC Prize will now label both harness conditions on its leaderboard.

An ARC Prize replay contains Astra's tiny scratch language, including L8: hub q2 (8↓), for recording state and mechanism lengths. The verified results table splits six reasoning-effort settings across the two harnesses. A pre-release API probe had already found that gpt-6-astra produced a gated endpoint response, and before the launch the prelaunch statement said training had finished while safety and alignment work continued.

ARC Prize frames the two harnesses as two distinct measurements in its analysis: a provider-neutral comparison, and performance with the provider's native context-management features.

• Standard harness: every provider receives the same minimal interface. The model can retain the visible notes it chooses, and must decide what information earns a place in that scratchpad.
• Provider Adapter harness: OpenAI's adapter retains opaque reasoning state between requests and applies compaction as the conversation grows, so prior work stays available without carrying the entire raw transcript forward.

Only the evaluation wrapper changed, a harness explainer argued. The reported result therefore covers a model operating inside a specified context-management system.

ARC Prize's verified result table makes the gap visible at every reasoning effort:

| Reasoning effort | Standard harness | Provider Adapter harness |
| --- | ---: | ---: |
| Max | 62.7%, $26,098 | 98.6%, $17,332 |
| XHigh | 59.3%, $37,317 | 98.4%, $18,147 |
| High | 54.8%, $40,705 | 99.9%, $18,817 |
| Medium | 38.6%, $48,090 | 98.4%, $19,285 |
| Low | 17.5%, $38,166 | 98.0%, $21,298 |
| None | 35.2%, $49,791 | 96.7%, $23,457 |

The headline best-run comparison crosses effort levels, Max for Standard and High for Provider Adapter. At matched Max effort, the adapter increases the score by 35.9 points; at Low, the difference reaches 80.5 points.

Across the 167 game-reasoning pairs both configurations solved, Provider Adapter runs were about 3.66 times faster by recorded elapsed time and used 49% fewer tokens, as ARC Prize's analysis records. The lower cost is part of the system result: retained state and compaction reduce repeated reconstruction of a game model.

In the Standard setup, Astra wrote down a compact symbolic representation of each environment. ARC Prize's replay analysis records four recurring forms:

• Game state: L8: hub q2 (8↓). Lengths: 14=1…
• Ordered plan: extend8 to3; retract10 to2; shorten8 to1
• Control map: 9−=(39,4), rotate=(49,18), 14+=(59,11)
• Position and time: Turn 5: P=(24,20), empty, facing west

The notes encode objects, coordinate bindings, rules, and unfinished plans in a small domain-specific shorthand. Provider Adapter state is opaque, so ARC Prize exposes no corresponding internal record for the near-saturating runs.

ARC Prize's action-efficiency analysis defines its human baseline as the median action count among people who completed each level. It tested roughly 500 members of the public, without selecting for puzzle expertise.

Under the Provider Adapter, Astra used fewer actions than that median on 96% of levels, and 51.7% fewer actions per completed level on average, according to the ARC Prize report. That result measures the number of interactions needed to learn and execute a solution, rather than the API cost of producing it.

OpenAI's launch post calls the 99.9% figure an ARC-AGI-3 saturation result. ARC Prize's reporting identifies it as the High Provider Adapter result, while the best Standard result is 62.7% at Max effort.

A commenter in a Hacker News discussion flagged another comparability issue: the public chart showed GPT-5.6 Sol at 7.8%, while the scorecard text estimated it could land near 30% with a Responses API harness. That estimate is not a rerun.

ARC Prize will publish both Standard and Provider Adapter results for future models, with the evaluation condition labelled, in its testing update.

The Information reported that Astra uses a reasoning technique called recurrent depth in the original report, and its reporter said the headline required more nuance in a follow-up. The claim concerns a looped transformer, where a shared block processes hidden states repeatedly before the next token is emitted.

Looped transformers reuse parameters to add effective computation per token, according to rasbt's technical breakdown; that post also argues weight reuse alone does not remove emitted chain-of-thought tokens. ARC Prize ties the observed ARC score split to cross-request opaque state and compaction in its harness analysis, a separate part of the system from the reported architecture.