GPT-6 Scored 99.9% on ARC-AGI-3, at a Cost of $18,000 in API Compute

OpenAI's GPT-6 scored 99.9% on the Semi-Private leaderboard of ARC-AGI-3, a public visual-reasoning benchmark the ARC Prize foundation built to resist memorization. The full run cost about $18,817 in API compute, according to the ARC Prize results page, or roughly $360 per puzzle. A human solver on the same test costs about $12.78 in wages and a fraction of a cent in electricity.

ARC Prize president Greg Kamradt, who runs the benchmark, said on X that "no harness was used in the making of these scores." For the past two years every serious ARC-AGI attempt has leaned on a harness, meaning external scaffolding that holds memory, tracks state, runs code, and lets the model iterate. Kamradt's argument is that GPT-6 has begun absorbing that scaffolding into its own weights.

The model behind the score, called Astra, learned to turn unfamiliar visual puzzles into compact symbolic world models, which are representations of the game as logical rules rather than pixels. It then developed its own internal shorthand to track state and plan moves. Long-time AI skeptic Gary Marcus, who has argued for nearly a decade that the path to general intelligence runs through (neuro)symbolic world models, called the result "pretty impressive" and a genuine advance.

The 99.9% isn't the only number. On the same Semi-Private leaderboard, GPT-6 with the standard harness scores 62.7% at maximum reasoning, at a cost of $26,098. The 99.9% figure requires a configuration ARC Prize calls the "Provider Adapter," which preserves opaque reasoning state between requests and uses a technique called compaction to keep long conversations manageable. That configuration costs less and scores more, because the model can reuse work it has already done. A Chinese-language analysis from LeiPhone called this "钞能力刷分" (money-burning score farming), and the $360-per-puzzle figure is consistent with that critique. The order-of-magnitude cost isn't in dispute; it is on ARC Prize's own page.

Kamradt's caveat is that the Provider Adapter isn't the same as a heavy client-side harness. A client-side harness, in his account, is code the user writes to wrangle a weaker model. The Provider Adapter is closer to plumbing: it preserves state, compacts memory, and lets the model run longer. The shift he is pointing at is that the capabilities the harness used to provide, including lossless memory, programmatic analysis, explicit hypothesis testing, persistent state, and cheap internal compute, are now inside Astra's weights. The harness is shrinking, while the score is staying high.

GPT-6 also used fewer actions than the median tested human on 96% of ARC-AGI-3 levels, averaging 51.7% fewer actions per puzzle. It searches more efficiently rather than guessing more.

OpenAI describes GPT-6 Astra as state-of-the-art on computer use, browsing, software engineering, cyber security, science, and professional work. The Independent's coverage places the release in a competitive context with Claude and Gemini, both of which trail on the same leaderboard.

ARC-AGI-3 is a closed-world puzzle set with a known answer key. If the harness is now inside the model, the score stops measuring what it used to measure. The next question is whether a model that builds a symbolic world model of an ARC puzzle can also build one of a robotics task, a corporate workflow, or a piece of code it has never seen. The $18,000 buys a proof of capability. Generality is the next benchmark.