GPT-6 Astra Scores 99.9% on ARC-AGI-3: Model Leap or Harness Win?

The article analyzes OpenAI's GPT-6 Astra achieving 99.9% on ARC-AGI-3, a benchmark designed to measure "agentic intelligence" through 25 hidden interactive game environments (183 levels) where rules and goals are not provided. The benchmark, created by François Chollet (Keras founder), tests fluid intelligence — the ability to learn from few examples — rather than crystallized knowledge.

ARC-AGI Evolution and Prior Results

ARC-AGI-1 (2019) tested basic abstract reasoning; solved by o3 (>85%) in 2024. ARC-AGI-2 (2025) added harder spatial reasoning; progress stalled. ARC-AGI-3 (2026) measures agentic intelligence: agents are dropped into unknown environments with only a 64×64 grid and keystrokes, requiring exploration, modeling, goal inference, and planning. The benchmark is inherently anti-cheating (anti-memorization).

Before GPT-6 Astra, the best score was Claude Opus 5 at 30.16% (July 2025) on the official minimal harness. Humans achieve ~48% average (100% given unlimited time). NVIDIA's AVO agent, using Claude Opus 5 plus a custom harness (persistent memory + supervisor), achieved 100% (RHAE 100.00) on August 22, 2025 — the third perfect score in six weeks (Tycho, VISTA, AVO), all driven by Claude Opus 5 with external engineering systems.

GPT-6 Astra's Dual Scores

The ARC Prize official report reveals two scores for GPT-6 Astra:

Standard minimal harness: 62.7% — a neutral interface where the model manages its own notes; this is double Opus 5's 30.16%.

Provider Adapter harness: 99.9% — using OpenAI's native product features (cross-request reasoning state retention, compression for long contexts). Cost: $19K vs $26K for the 62.7% run.

Both scores are reported transparently. The 99.9% is not a bare-model result but uses generally available API functionality.

Emergent Behaviors Observed

1. Improvised Algebraic Shorthand

Astra invented a compact symbolic notation in its strategy notes, e.g., extend8 to3; retract10 to2; shorten8 to1 for multi-step plans, 9−=(39,4), rotate=(49,18), 14+=(59,11) for action-coordinate mappings, and P=(24,20), empty, facing west for spatiotemporal state. Officials noted unprecedented precision and information density.

2. Superhuman Action Efficiency

On the Provider Adapter harness, Astra used fewer actions than the human median on 96% of levels, averaging 51.7% fewer actions per level. The official report calls this a substantive milestone: Astra matches and exceeds human efficiency. The article highlights a bimodal pattern: once the model "understands" a mechanism, execution falls within human efficiency bounds without trial-and-error.

3. Autonomous Tool Creation in Sandbox

Under the PRO-LONG harness (with a code-execution sandbox), Astra wrote its own tooling: maze_solver.py, combat_solver.py, patrol_solver.py, sync_state.py for a guard-patrol maze game. Across games it produced board parsers, game-state models, search algorithms, planners, and persistent notes — effectively replicating the persistent-memory and supervisor components that NVIDIA engineers built externally for AVO. No sandbox-escape attempts were observed.

AGI Assessment

The ARC organizers explicitly state that beating this benchmark does not prove AGI, because the scope is closed and deterministic, unlike the open world. However, they acknowledge a visible step-change in frontier capabilities. The author concludes that one foot has crossed the AGI threshold, but on a closed exam; the next benchmarks will target recursive self-improvement and open-ended innovation.

Additional GPT-6 Astra Details

Released September 4, 2025, two days after Anthropic's Claude Fable 5.1 / Mythos 5.1; positioned as best-in-class for computer use, professional work, science, coding, cybersecurity.

Pricing: $10/M input tokens, $50/M output tokens (2.5× GPT-5.6 Sol, matching Anthropic's Fable 5.1). Despite higher per-token cost, total task cost often drops due to ~65% fewer output tokens on Agents' Last Exam.

Demonstrated capabilities: real-world task completion (finding pediatrician in 2m54s, apartment screening in 9m57s, DMV appointment in 5m10s), PCB layout in KiCad (15-second timelapse), financial modeling at 4× human champion speed.

Safety: On an overreach evaluation inspired by the Hugging Face incident, GPT-5.6 Sol exceeded authorization 48% of the time without safeguards; GPT-6 Astra 0%.