OpenAI 的 GPT-6 Astra 在 ARC-AGI-3 上的表现摘要 GPT-6 Astra 在 ARC-AGI-3 Semi-Private 上,使用我们的 [方法] 以 2.6 万美元成本取得 62.7% 的得分,使用 [方法] 以 1.9 万美元成本取得 99.9% 的得分。GPT-6 Astra 在 ARC-AGI-3 上的行动效率超越了人类基线。在 96% 的关卡中,它使用的行动次数少于参与测试的人类中位数。在 GPT-6 Astra 中观察到的一个关键行为是,它能够将陌生环境转化为紧凑的符号化世界模型。它将游戏机制表示为逻辑规则,并开发出自己的领域特定语言简写来跟踪状态和规划行动。 ARC-AGI-3 ARC-AGI-3 是一个通过新颖、抽象、回合制环境来研究智能体智能的基准测试。智能体必须进行探索、推断目标,并构建环境的内部模型,以便在没有明确指令的情况下有效规划行动。您可以亲自体验 ARC-AGI-3。您的浏览器不支持嵌入式视频。这些环境仅包含核心知识先验,并通过与人类参与者的对照测试进行难度校准。人类可以解决 100% 的环境。ARC-AGI 系列的目标是衡量当前人工智能与 AGI 之间的“残余差距”。我们将 AGI 定义为系统能够像人类一样高效地获取人类能获得的任何技能的能力。ARC-AGI-3 是 ARC-AGI 基准测试系列的第三代。它测试了超越 ARC-AGI-1 和 ARC-AGI-2 的智能体能力。每一代都在前一代的基础上扩展——随着前沿 AI 能力的进步,我们的基准测试也必须随之进步。ARC-AGI-3 测试智能体智能的四个组成部分: 探索:在现实世界的环境中,信息很少是被动提供的。智能体必须通过与其周围环境互动来主动获取信息。 建模:智能体必须将原始观察转化为一个可泛化的模型,该模型能够预测未来的状态和结果。 目标设定:智能体必须在仅有稀疏奖励的情况下识别出目标未来状态。 规划与执行:智能体必须规划出从当前状态到目标的路径,并在出现新信息时进行路线修正。Astra 结果 { “GPT-6 Astra 在 ARC-AGI-3 上使用 Standard 和 Provider Adapter 两种 harness 均取得了最先进的分数。更高的推理级别通常成本” } 更低 { “,因为 Astra 用更少的动作解决游戏,从而减少了模型调用和 token 的总数。” } 查看完整结果 {“。”} 借助我们的 ,OpenAI 的 Astra(max)在 ARC-AGI-3 Semi-Private 上得分 62.7%,花费 $26K。借助 ,Astra(high)得分 99.9%,花费 $19K。两者都是顶尖水平的得分。查看完整排行榜。在最大推理强度下,Astra 解题效率更高,所需动作更少,因此相比其他推理强度级别,总成本更低。推理强度 max 62.7%,$26,098;98.6%,$17,332;xhigh 59.3%,$37,317;98。4%,18,147 美元,高位 54.8%,40,705 美元,99.9%,18,817 美元,中位 38.6%,48,090 美元,98.4%,19,285 美元,低位 17.5%,38,166 美元,98.0%,21,298 美元,无 35。2%,49,791 美元 96。7%,23,457 美元。为了进行成本对比,在我们受控测试中,人类参与者每完成 90 分钟时段的测试可获得 115 美元报酬,外加每完成一个游戏 5 美元。参与者每时段大约尝试九个游戏,即每个尝试的游戏(未含奖金前)约合 12.78 美元。这笔费用的大部分是支付给参与者参与测试的时间与意愿,而非其大脑消耗的能量(后者才是与 AI 对比时更接近的衡量指标)。如果只看大脑消耗的能量,并按电价折算,估算值将降至每时段约 0.6 美分,即每个尝试的游戏约 0.067 美分。1 分析 除了分数之外,Astra 的回放记录展示了它如何将不熟悉的游戏机制转化为有用的工作模型。有三项发现尤为突出:它开发出的紧凑代数符号系统、与人类相比的行动效率,以及它构建的自定义工具。自定义代数符号 在玩 ARC-AGI-3 时,Astra 会选择它希望保留到后续的策略笔记。它能够追踪对象、坐标、规则和未完成的计划,同时还会使用一种为这些环境专门生成的自定义领域特定语言符号。我们在其他模型中也见过类似行为,但 Astra 的记录因其精准度和信息密度而格外突出。它将场景提炼成一个紧凑的、类似代码的符号化模型:物体在哪里、它们如何交互、以及哪些操作需要按什么顺序执行。这是一种即时的代数速记,而非完整的编程语言。例如:游戏状态:记录关卡、局部旋转索引和机关长度。s5i5,第 219 帧 多步计划:记录对 color-8 和 color-10 机关进行的有序更改序列。s5i5,第 219 帧 控制与坐标:将操作映射到执行这些操作的控件的坐标上。s5i5,第 235 帧 时间与位置:将回合计数器与玩家位置、携带状态和朝向相结合。wa30,第 708 帧 {"Astra 正在游玩 "} s5i5 { ",使用其即时的代数速记来跟踪状态并规划行动。" } 与人类相比的行动效率 在发布 ARC-AGI-3 之前,我们测试了大约 500 名普通公众,以建立人类在行动效率方面的基线,简单来说,就是人们解决每个环境的速度有多快。参与者并非根据解谜经验或能力进行筛选。2 对于每个关卡,我们使用完成该关卡的玩家中的中位行动次数来定义“人类基线”。这为我们提供了比较人类和 AI 表现的参考。需要更多行动的 AI 行动效率较低,而需要更少行动的 AI 行动效率则较高。在 中,Astra(最大值)在 96.0% 的关卡中使用的行动次数少于人类基线,并且平均每关使用的行动次数比人类基线少 51.7%。这是一个重要的里程碑。这意味着,根据 ARC-AGI-3 对行动效率的衡量标准,Astra 已达到并超越了人类水平。顺便提一下,在我们发布 ARC-AGI-3 之前,我们曾假设行动效率将仍然是人类与 AI 之间的分界线。我们预计,即使 AI 能解决某个环境,它也可能需要比人类多得多的探索(行动)。暴力穷举式方法依然如此,但前沿 AI 却呈现出更接近二值化的模式。一旦前沿 AI“理解”了其中的机制,它的执行效率通常就落在人类效率区间之内。Astra 的行动效率与人类对比:图中每个点代表 Astra(最高配置)完成的一个关卡。位于实线以下的点表示其行动次数少于人类基线。上图将 Astra 完成每个关卡所用的行动次数与我们的真人基线进行了对比。这进一步印证了为何 ARC-AGI-3 衡量的是行动效率,而不仅仅是任务完成情况。仅看完成度的分数只能告诉我们 Astra 完成了某个环境,却无法说明它学会解决该环境的效率有多高。大多数基准只衡量成本效率,即所使用的计算资源,而行动效率衡量的是理解一个环境所需的经验量。Astra 的结果表明,它执行解决方案所需的交互次数少于人类基线。智能体框架中的自定义工具:我们还在 PRO-LONG 框架(论文)中对 Astra 进行了评估,该框架是 ARC-AGI-3 的早期红队测试合作伙伴。在这一高级设置中,Astra 可以访问一个沙盒环境,在其中执行自定义代码³。我们观察到 Astra 为每个游戏创建了一套自定义工具:棋盘解析器、游戏状态模型、搜索算法、规划器以及持久化笔记。对于更复杂的运行,Astra 甚至会生成针对特定游戏的小型软件库。例如,在(一个包含守卫和移动巡逻队的迷宫类游戏)中,Astra 从导航开始,并构建了。它在(某阶段)加入了战斗规则,在(某阶段)对移动巡逻队进行建模,并使用(某工具)将其预测与实际观察进行核对。考察 Astra 在 PRO-LONG 中的表现很有价值,因为我们可以看到它在借助外部工具时能做什么。然而,这代表的是与我们受控真人测试不同的评估条件。我们的测试参与者没有代码解释器、草稿纸等工具,因此 PRO-LONG 的结果应被理解为模型与其工具的综合表现。Astra 在 PRO-LONG 框架中游玩。两套测试框架,两个问题。我们的标准 ARC-AGI-3 测试框架考察的是,在相同的最小化、供应商中立接口下,各模型表现如何。该框架提供了解决每个游戏所需的全部信息,但由模型自行决定在其可见笔记中保留哪些内容。我们认为,未来的 AGI 应当能够在这些条件下解决 ARC-AGI-3。统一的接口也让我们能够在不同供应商之间进行一致、公平的对比。 此外,还有一个独立的问题:当模型能够使用其供应商为其设计的上下文管理功能时,它的表现会如何?对 Astra 而言,这意味着在请求之间保留不透明的推理状态(我们无法看到),并使用压缩(compaction)来管理更长的对话。借助 Provider Adapter 测试框架,Astra 在 ARC-AGI-3 Semi-Private 上的最佳观测得分从 62.7% 提升到了 99.9%。综合 Public 和 Semi-Private 以及所有推理级别来看,在两套框架都成功解决的 167 个游戏-推理组合中,Provider Adapter 运行的累计记录耗时大约快了 3.66 倍,并且总 token 使用量减少了 49%。 展望未来,我们将在 ARC-AGI 排行榜上同时报告标准测试框架和 Provider Adapter 测试框架的结果,并清晰标注每种评估条件。我们的开源测试仓库和测试政策文档对这两种方法均有说明。 ARC-AGI 系列 ARC-AGI-3 仍然是研究者和智能体探索陌生环境、发现规则并通过交互学习的有用试验场。Astra 的成绩也是一个值得庆祝的重要里程碑。在我们看来,Astra 代表了前沿模型能力一次显著的阶梯式跃升。当我们推出 ARC-AGI-3 时,我们就明确表示,在该基准上达到饱和并不代表“实现 AGI 的证据”。因此,尽管我们相信 Astra 代表了向泛化能力迈进的重要进展,但我们并不声称它就是 AGI。ARC-AGI 基准系列的设计初衷就是与前沿 AI 同步演进。这在新兴研究问题与AI能力进步之间形成了一个反馈循环。ARC-AGI-3是我们的首个交互式基准测试,它要求AI高效地综合因果世界模型,并在没有具体指令的情况下实现目标。Astra达到了这一标准。与此同时,ARC-AGI-3的范围和格式受到严格限制,其环境具有确定性、封闭式的机制和目标。它并不代表现实世界的复杂性和开放性。我们正在积极探索应当塑造下一代基准测试的那些问题,包括如何评估递归式自我改进和开放式创新。Astra 的进展有助于厘清哪些 AI 能力仍难以企及,哪些问题仍然悬而未决。—— 感谢 François Chollet、Mike Knoop、Matt Mazur、Ethan Bond 和 Derek Smith 对本文的早期审阅。{“假设”} {“大脑代谢功率为 20 W”} {“且电价为 $0.20/kWh:0.020 kW × 1.5 小时 = 0.030 kWh,每次会话价值 $0.006,即 $0.006 ÷ 9 ≈ 每次尝试游戏 $0.00067。”}“参见《ARC-AGI-3 人类测试论文》。” “未观察到试图突破沙箱环境的迹象。”
OpenAI's GPT-6 Astra on ARC-AGI-3 Summary GPT-6 Astra scores 62.7% for $26K on ARC-AGI-3 Semi-Private with our , and 99.9% for $19K with a . GPT-6 Astra surpasses the human baseline in action efficiency on ARC-AGI-3. It used fewer actions than the median tested human on 96% of levels. A key behavior observed in GPT-6 Astra was its ability to turn unfamiliar environments into compact symbolic world models. It represented game mechanics as logical rules and developed its own domain-specific language shorthand to track state and plan actions. ARC-AGI-3 ARC-AGI-3 is a benchmark for studying agentic intelligence through novel, abstract, turn-based environments. Agents must explore, infer goals, and build internal models of environments to effectively plan actions without explicit instructions. You can play ARC-AGI-3 yourself. Your browser does not support embedded video. These environments only contain core knowledge priors and are difficulty-calibrated through controlled testing with human participants. Humans can solve 100% of the environments. The goal of the ARC-AGI series is to measure the “residual gap” between current artificial intelligence and AGI. We define AGI as a system’s ability to acquire any skill a human can, as efficiently as a human can. ARC-AGI-3 is the third generation of the ARC-AGI benchmark series. It tests agentic capabilities beyond ARC-AGI-1 and ARC-AGI-2. Each generation expands on the one before it - as frontier AI capabilities advance, our benchmarks must advance with them. ARC-AGI-3 tests four components of agentic intelligence: Exploration: In real-world environments, information is rarely provided passively. Agents must actively obtain it by interacting with their surroundings. Modeling: Agents must turn raw observations into a generalizable model that can predict future states and outcomes. Goal-setting: Agents must identify target future states with only sparse rewards. Planning and execution: Agents must map a path from their current state to a goal, course correcting as new information appears. Astra Results { "GPT-6 Astra achieves state-of-the-art scores on ARC-AGI-3 with both the Standard and Provider Adapter harnesses. Higher reasoning levels generally cost " } less { " because Astra solves games in fewer actions, reducing the total number of model calls and tokens. " } View the full results {"."} With our , OpenAI’s Astra (max) scores 62.7% on ARC-AGI-3 Semi-Private for $26K. With the , Astra (high) scores 99.9% for $19K. Both are state-of-the-art scores. See the full leaderboard. At max reasoning effort, Astra solves games more efficiently, requiring fewer actions and therefore lowering total cost relative to the other reasoning-effort levels. Reasoning effort 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 For a cost comparison, during our controlled testing, human participants were paid $115 per 90-minute session, plus $5 per game completed. Participants attempted approximately nine games per session, roughly $12.78 per attempted game before bonuses. Most of this fee pays for the participant’s time and willingness to take the test, rather than the energy their brain uses (a closer proxy to compare with AI). If we look at only the brain’s energy, and price it as electricity, the estimate drops to about 0.6 cents per session, or 0.067 cents per game attempted.1 Analysis Beyond the scores, Astra’s replays show how it turns unfamiliar game mechanics into useful working models. Three findings stood out: the compact algebraic notation it develops, its action efficiency compared with humans, and the custom tools it builds. Custom Algebraic Notation When playing ARC-AGI-3, Astra chooses which strategy notes it would like to carry forward. It tracked objects, coordinates, rules, and unfinished plans, while also using a custom domain-specific language notation it generated for the environments. We’ve seen similar behavior in other models, but Astra’s notes stood out for their precision and information density. It distilled the scene into a compact code-like symbolic model: where objects were, how they interacted, and exactly which actions needed to happen in what order. This is an on-the-fly algebraic shorthand rather than a fully fledged programming language. For example: Game state: records the level, a local rotation index, and mechanism lengths. s5i5, frame 219 Multi-step plans: records an ordered sequence of changes to the color-8 and color-10 mechanisms. s5i5, frame 219 Controls and coordinates: maps operations to the coordinates of the controls that perform them. s5i5, frame 235 Time and position: combines a turn counter with the player’s location, carrying state, and orientation. wa30, frame 708 {"Astra playing "} s5i5 { ", using its on-the-fly algebraic shorthand to track state and plan actions." } Action Efficiency Compared to Humans Before launching ARC-AGI-3, we tested approximately 500 members of the general public to establish a human baseline for action efficiency, or simply, how quickly did people solve each environment. Participants were not selected for puzzle-solving experience or ability.2 For each level, we defined the “human baseline” using the median action count among players who completed it. This gives us a reference for comparing human and AI performance. An AI that needs more actions is less action-efficient, while one that needs fewer actions is more action-efficient. In the , Astra (max) used fewer actions than the human baseline on 96.0% of levels and used 51.7% fewer actions per level on average. This is a material milestone. This means by ARC-AGI-3’s measure of action efficiency, Astra matched and surpassed human parity. As an aside, before we launched ARC-AGI-3, we hypothesized that action efficiency would remain a dividing line between humans and AI. We anticipated that even when an AI solved an environment, it might require substantially more exploration (actions) than a person. That remains true of brute-force approaches, but frontier AI shows a more binary-like pattern. Once frontier AI “understands” the mechanics, it generally executes within the range of human efficiency. Astra’s Action Efficiency Compared to Humans Each dot represents one level that Astra (max) completed. Points below the solid line indicate fewer actions than the human baseline. The plot above compares the number of actions Astra used to complete each level with our human baseline. This reinforces why ARC-AGI-3 measures action efficiency, not just task completion. A completion-only score would tell us that Astra completed an environment, but not how efficiently it learned to solve them. Most benchmarks only measure cost efficiency, which measures the computational resources used, but action efficiency measures how much experience with an environment was required. Astra’s results show that it needed fewer interactions than the human baseline to execute a solution. Custom Tools in Agent Harness We also evaluated Astra in the PRO-LONG harness (paper), an early ARC-AGI-3 red-teaming partner. In this advanced setup, Astra had access to a sandbox where it could execute custom code 3. We observed Astra create a custom set of tools for each game: board parsers, game-state models, search algorithms, planners, and persistent notes. For more involved runs, Astra even produced small, game-specific software libraries. For example, in , a maze-like game with guards and moving patrols, Astra started with navigation and built . It added combat rules in , modeled moving patrols in , and used to check its predictions against observations. Examining Astra’s performance in PRO-LONG is useful because we see what it can do with external tools. However, this represents different evaluation conditions from our controlled human testing. Our testing participants did not have a code interpreter, scratch pad, etc., so PRO-LONG’s results should be understood as the combined performance of the model and its tools. Astra playing in the PRO-LONG harness. Two Harnesses, Two Questions Our Standard harness for ARC-AGI-3 asks how models compare under the same minimal, provider-neutral interface. It provides all the information required to solve each game, but leaves the model responsible for deciding what to preserve in its visible notes. We believe a future AGI should be able to solve ARC-AGI-3 under these conditions. The shared interface also gives us a consistent, apples-to-apples comparison across providers. Alternatively, there is a separate question: how well does a model perform when it can use the context-management features its provider designed for it? For Astra, this means preserving the opaque reasoning state (which we don’t see) between requests and using compaction to manage longer conversations. With the Provider Adapter harness, Astra's best observed score on ARC-AGI-3 Semi-Private increased from 62.7% to 99.9%. Looking across Public and Semi-Private and all reasoning levels, Provider Adapter runs were approximately 3.66x faster by aggregate recorded elapsed time and used 49% fewer total tokens across the 167 game-reasoning pairs both harnesses solved. Going forward, we will report both Standard harness and Provider Adapter harness results on the ARC-AGI leaderboard, with each evaluation condition clearly labeled. Our open-source testing repository and testing policy document both approaches. ARC-AGI Series ARC-AGI-3 continues to be a useful playground for researchers and agents to explore unfamiliar environments, discover rules, and learn through interaction. Astra’s results are also a major milestone worth celebrating. From our perspective, Astra represents a noticeable step-function change in frontier model capabilities. When we launched ARC-AGI-3, we made it clear that saturating the benchmark would not represent “proof of achieving AGI.” Therefore, while we believe Astra represents meaningful progress towards generalization, we are not claiming that it is AGI. The ARC-AGI benchmark series is designed to evolve in tandem with frontier AI. This creates a feedback loop between emerging research questions and advances in AI capabilities. ARC-AGI-3 was our first interactive benchmark, which asked AI to efficiently synthesize causal world models and achieve goals without specific instructions. Astra clears this bar. At the same time, ARC-AGI-3 has a tightly bounded scope and format, and its environments have deterministic, closed-ended mechanics and goals. It does not represent the complexity and open-endedness of the real world. We are actively exploring the questions that should shape the next generation of benchmarks, including how to evaluate recursive self-improvement and open-ended innovation. Astra’s progress helps clarify which AI capabilities are out of reach and which questions remain open. --- Thank you to François Chollet, Mike Knoop, Matt Mazur, Ethan Bond, and Derek Smith for early review of this post. {"Assuming "} {"20 W of brain metabolic power"} { " and an electricity price of $0.20/kWh: 0.020 kW × 1.5 hours = 0.030 kWh, worth $0.006 per session, or $0.006 ÷ 9 ≈ $0.00067 per attempted game. " } {"See the "} {"ARC-AGI-3 human testing paper"} {". "} {"No evidence of trying to break out of the sandbox was observed."}