GPT-6 Astra saturates ARC-AGI-3 with near-perfect scores and near-zero reasoning tokens

GPT-6 Astra achieved a breakthrough on the interactive reasoning benchmark ARC-AGI-3, described by ARC Prize and François Chollet as a step-change. It scored roughly 63–66% under the standard harness and 99–99.9% with the provider-adapter (continuous conversation) setup, completing tasks in fewer steps than the average human player. The result draws attention for two reasons: scores swing dramatically depending on the harness, and at low reasoning settings most actions consume no reasoning tokens, hinting at a leap in implicit reasoning.

Confirmed

• ARC Prize's official blog and François Chollet announced GPT-6 Astra's SOTA on ARC-AGI(-3): 63–66% on the standard unbiased harness, 99–99.9% via the provider adapter harness, with strong performance on 96% of tasks (m1, m5, m9, m10).
• OpenAI engineer Steven Heidel introduced compaction in the Responses API, which compresses long conversation state to help agents manage context and cost in long tasks; with compaction enabled, GPT-6 Astra reached a perfect score, beating Opus 5 and GPT-5.6 Sol (m3, m6, m8, m9).
• Astra used fewer steps than the average human player (m1).
• Charts from Mike Knoop show that at lower reasoning settings Astra often outputs zero reasoning tokens per action while doubling the accuracy of GPT-5.6 Sol at max reasoning (m2); mhmazur's behavioral stats report finer numbers — under the standard environment at low reasoning effort, 98% of GPT-5.6 Sol's actions used reasoning tokens versus only 27% for GPT-6 Astra (m7).
• One user posted a 97% score for GPT-6 Astra without CoT, noting it was measured on the provider-adapter harness yet still calling it "pretty absurd" (m4).

Unconfirmed

• mhmazur and others note the perfect-score evaluation setup is not entirely fair; the cause of the spread from 62.7% to 99.9% across harnesses awaits a more transparent explanation (m5, m8, m9).
• Whether the zero-reasoning-token phenomenon confirms a major leap in implicit reasoning remains an inference from behavioral statistics (m7).

Why it matters

• Scores that swing wildly with the harness highlight how much evaluation design shapes conclusions on interactive agent benchmarks; claims of "saturation" must specify the setup.
• High scores with most actions producing no reasoning tokens, if implicit reasoning holds, would challenge the assumption that explicit chain-of-thought is required for strong reasoning.
• Responses API compaction shows context management is becoming a key engineering lever for long-horizon agent tasks.

2026-09-04 ~ 2026-09-05 · 10 related posts

Full story(7 episodes)→

• Episode 1: OpenAI launches flagship GPT-6 Astra to rave reviews and pricing debates(2026-09-04, 190 posts)
• Episode 2: GPT-6 Astra saturates ARC-AGI-3 with near-perfect scores and near-zero reasoning tokens(2026-09-04, 10 posts)
• Episode 3: OpenAI Engineer: Token-Based Pricing Is Broken, Cost Should Be Measured Per Task(2026-09-04, 4 posts)
• Episode 4: OpenAI reportedly quietly edited GPT-6 Astra benchmark results(2026-09-05, 4 posts)
• Episode 5: Users Report GPT-6 Astra Hits New SOTA in Agentic CAD(2026-09-05, 2 posts)
• Episode 6: GPT-6 Astra reportedly runs Blender for 5 hours on 3D modeling(2026-09-06, 2 posts)
• Episode 7: Astra's Record Debut Reportedly Pushes OpenAI to Accelerate Roadmap(2026-09-06, 2 posts)

• [source] Astra Fully Saturates the ARC-AGI-3 Benchmark Using Fewer Moves Than the Average Human — ObiWanCanownme · 2026-09-04
• OpenAI Ships Responses API Compaction as GPT-6 Astra Maxes Out ARC-AGI-3 — stevenheidel · 2026-09-04
• GPT-6 Astra maxes ARC-AGI-3 with compaction, but harnesses differ — mhmazur · 2026-09-04
• Astra reportedly saturates ARC-AGI-3 with a 99.9% score — smlpth · 2026-09-04
• GPT-6 Astra reportedly hits 97% on ARC-AGI-3 without CoT, even under a provider-adapter harness — haider1 · 2026-09-04
• [source] GPT-6 Astra uses reasoning tokens in only 27% of ARC-AGI-3 actions, hinting at a leap in opaque reasoning — mhmazur · 2026-09-05
• ARC v3: Astra Low Emits Zero Reasoning Tokens Yet 2x More Accurate Than Sol Max — rbhar90 · 2026-09-05
• GPT-6 Astra hits 99% on ARC-AGI-3 via provider adapter harness, up from 62.7% — rohanpaul_ai · 2026-09-05
• [source] GPT-6 Astra hits 66% on ARC-AGI-3, near-100% with custom harness at ~$360 per game — AccBalanced · 2026-09-05

1 near-duplicate retellings: stevenheidel