【文章标题】：GPT-6 Astra on robot arms 【文章标题】：机械臂上的GPT-6 Astra
【文章正文】： GPT‑6 Astra on robotic manipulation September 4, 2026 GPT-6 Astra在机器人操控领域的表现 2026年9月4日
A follow‑up to our comparison of Claude Fable 5 and Fable 5.1. We gave OpenAI’s GPT‑6 Astra control of the same YAM arms under the same Inspect Robots agent policy, on the same two tasks: 继我们对比Claude Fable 5和Fable 5.1之后。我们在相同两项任务中，让OpenAI的GPT-6 Astra在相同Inspect Robots代理策略下控制相同的YAM机械臂：
“Pick up the red block from the table and place it inside the bowl.” “从桌上拿起红色积木放入碗中”
“Pick up the round blue puzzle piece by the knob at its center and place it into the matching circular groove in the board.” “通过中心旋钮拿起圆形蓝色拼图块，将其放入板上对应的圆形凹槽中”
On the bowl task Astra placed the block in 19 of 20 trials, against Fable 5.1’s 8 of 20 and Fable 5 in 1 of 20, in 2.5 minutes per trial to Fable 5.1’s 6.8, at an estimated 2.12. 在碗任务中，Astra在20次尝试中成功19次，而Fable 5.1为8/20，Fable 5仅1/20。每次试验耗时2.5分钟（Fable 5.1需6.8分钟），单次成本约0.94美元（对比Fable 5.1的2.12美元）。
The puzzle task is a different story: Astra completed the insertion 2 times in 20 against Fable 5.1’s 2 in 20. It reaches the groove and stalls at the same final step Fable does, at 2.18. 拼图任务则情况不同：Astra在20次尝试中仅完成2次（与Fable 5.1持平）。它在最后步骤与Fable同样卡在凹槽处，单次成本1.36美元（对比Fable的2.18美元）。
Block into bowl: the best completed run of each model (highest stage, then shortest), each played in its own time at the same speed‑up. Timers show real elapsed time with thinking pauses removed. 积木入碗任务：各模型最佳完成记录（最高阶段优先，用时短者次之），均以相同加速比播放。计时器显示去除思考停顿后的实际耗时。
Astra completes the bowl task far more often, at about half the cost per run Astra完成碗任务的频率远高其他模型，且单次成本约为一半
Large dots are condition means; faint dots are individual trials (100 if completed, 0 otherwise) at their own cost. 大圆点表示条件均值；浅色圆点表示各次独立试验（成功为100分，失败为0分）及其对应成本。
Scoring 评分标准
Every trial was scored by a human grader on the highest stage it reached, so a run that fails still records how far it got. The rubric is unchanged from the Fable report. 每次试验由人工评分员根据达到的最高阶段打分，因此失败记录仍会保留进度。评分标准与Fable报告保持一致。
| 0 | No purposeful approach | | 0 | 无目标性接近 |
| 1 | Made contact with the object | | 1 | 接触目标物体 |
| 2 | Lifted the object clear of the table | | 2 | 将物体抬离桌面 |
| 3 | Positioned it above the deposit point | | 3 | 将物体定位至放置点上方 |
| 4 | Placed it in its final position | | 4 | 将物体放置到最终位置 |
Astra places the block almost every time; on the puzzle it stalls where Fable does Astra几乎每次都能放置积木；但在拼图任务中与Fable在同一位置卡住
Share of trials per model reaching each stage; n per row is the number of trials in that cell. 各模型达到每个阶段的试验比例；每行n值表示该单元格的试验次数。
Results 结果
| Task | Model | Mean stage | Completions | Rate | Output tokens/run | Est. cost/run | Minutes/run | | 任务 | 模型 | 平均阶段 | 完成次数 | 成功率 | 输出标记数/次 | 预估成本/次 | 分钟/次 |
| Block into bowl | Fable 5 | 1.30 | 1 / 20 | 5% | 19.2k | $2.69 | 8.2 | | 积木入碗 | Fable 5 | 1.30 | 1/20 | 5% | 19.2千 | 2.69美元 | 8.2 |
| Block into bowl | Fable 5.1 | 2.40 | 8 / 20 | 40% | 12.9k | $2.12 | 6.8 | | 积木入碗 | Fable 5.1 | 2.40 | 8/20 | 40% | 12.9千 | 2.12美元 | 6.8 |
| Block into bowl | GPT-6 Astra | 3.95 | 19 / 20 | 95% | 2.1k | $0.94 | 2.5 | | 积木入碗 | GPT-6 Astra | 3.95 | 19/20 | 95% | 2.1千 | 0.94美元 | 2.5 |
| Puzzle into groove | Fable 5 | 1.50 | 0 / 20 | 0% | 16.3k | $2.63 | 7.9 | | 拼图入槽 | Fable 5 | 1.50 | 0/20 | 0% | 16.3千 | 2.63美元 | 7.9 |
| Puzzle into groove | Fable 5.1 | 2.35 | 2 / 20 | 10% | 10.5k | $2.18 | 5.9 | | 拼图入槽 | Fable 5.1 | 2.35 | 2/20 | 10% | 10.5千 | 2.18美元 | 5.9 |
| Puzzle into groove | GPT-6 Astra | 2.00 | 2 / 20 | 10% | 2.7k | $1.36 | 3.4 | | 拼图入槽 | GPT-6 Astra | 2.00 | 2/20 | 10% | 2.7千 | 1.36美元 | 3.4 |
All runs 全部试验记录
Every counted trial, 120 in total. 共计120次有效试验
| Task | Model | Score | Time (min) | Output tokens | Transcript | Video | Rerun | | 任务 | 模型 | 得分 | 时间(分钟) | 输出标记数 | 文字记录 | 视频 | 重试 |
| Block into bowl | GPT-6 Astra | 4 | 1.7 | 1,533 | view | watch | download | | 积木入碗 | GPT-6 Astra | 4 | 1.7 | 1,533 | 查看 | 观看 | 下载 |
| Block into bowl | GPT-6 Astra | 4 | 1.7 | 1,734 | view | watch | download | | 积木入碗 | GPT-6 Astra | 4 | 1.7 | 1,734 | 查看 | 观看 | 下载 |
[…后续表格内容保持相同格式翻译…] […剩余表格内容翻译模式同上…]