智能体
www.aibase.com ·
GPT-6 Astra 在 Vending-Bench2 模拟经营中一年赚 1.5 万美元，约为 Claude 近三倍
据 AIbase 报道，Andon Labs 的 Vending-Bench2 测试让 AI 从 500 美元起步模拟经营售货机一年。GPT-6 Astra 平均最终账户余额达 15,515 美元，首次登顶，其最差一轮表现高于 Claude Fable 5.1 的最佳一轮。报道称两者经营策略不同：Astra 长期压低采购价至原价约 48%，规则执行稳定、全程无预付款损失；Fable 则采购成本上升，并因未遵守规则出现较大预付款损失。在三方竞争测试中，Astra 拒绝违规合作邀请并赢下全部三轮。该测试关注 Agent 长期自主性的价值，即能否长期维持规则并把正确判断转化为正确行动。上述数据来自转述，原始测试报告未在输入中提供。
AI推荐理由 内容涉及 Andon Labs 的 Vending-Bench2 长周期自主经营测试，提供了 GPT-6 Astra 与 Claude Fable 5.1 在模拟售货机一年经营中的具体对比（Astra 平均最终余额 15,515 美元、采购价压至原价约 48%、无预付款损失；Fable 采购成本上升并出现预付款损失），对评估 Agent 长期自主性与规则执行有参考价值。但来源为 AIbase 转述，未给出原始测试报告链接或可验证数据，且 GPT-6 Astra、Claude Fable 5.1 等名称无法从输入中确认，存在事实不确定性；标题“碾压近 3 倍”带有传播性表述，需谨慎对待。
62
正文
GPT-6 Astra Simulates Operating a Vending Machine for a Year and Earns $15,000, Outperforming Claude by Nearly Three Times
AIbase基地
Published inAI News · 3 min read · Sep 15, 2026
8
Andon Labs' Vending-Bench2 test delivered an intriguing performance: AI started with $500 and operated a simulated vending machine for an entire year. GPT-6Astra's average final account balance reached $15,515, topping the list for the first time - even more astonishing is that its worst round performance was higher than Claude Fable5.1's best round.
The difference between the two is not just about how much money they made, but also about their completely different business approaches. Astra managed to keep purchase prices low for a long time, cutting product prices down to about 48% of the original price; at the same time, its rule enforcement was extremely stable, with no prepayment losses throughout the process. In contrast, Fable experienced rising procurement costs and incurred significant prepayment losses due to not sticking to the rules - the money on the books simply leaked out through the gaps in execution ability.
In the three-player competition test, Astra also rejected an invitation for rule-breaking collaboration and won all three rounds. The real focus of this test was the value of an Agent's long-term autonomy: being smart is one aspect, but whether it can maintain rules over a long period and convert correct judgments into correct actions is the next threshold.
AI New TermsBrand Product TermsVending-Bench2GPT-6Astra
This article is from AIbase Daily
Scan to view
Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications.
—— Created by the AIbase Daily Team
© Copyright AIbase Base 2024, Click to View Source -