Tops robot comparison rankings three days after release

RoboColosseum, which evaluates robots, posted on X at 11:59 a.m. Nov. 21 Japan time that OpenAI's top model "GPT-6.2 Astra" had taken first place on its Open track leaderboard for "SO-101." It scored 1,984 on the Elo scale. Elo is a method that assigns scores and ranks entries based on wins and losses in head-to-head comparisons. GPT-6.2 Astra was rated on 122 comparisons.

GPT-6.2 Astra now tops our SO-101 Open track leaderboard at 1,984 on the Elo scale, based on 122 pairwise comparisons. A couple of side-by-side runs from the evaluations 🧵

OpenAI released the model on Nov. 17 (Nov. 18 Japan time), so the leaderboard was computed three days after release. Earlier, at 8:45 a.m. Nov. 21 Japan time, Jiafei Duan (@DJiafei), from the side that ran the evaluation, posted on X that GPT-6.2 Astra had reached first place in RoboColosseum's SO-101 Elo rankings. Duan wrote that the team ran hundreds of head-to-head comparisons in real-world settings in one day, and called GPT-6.2 Astra "seriously capable" at robotics.

GPT-6.2 Astra is proving seriously capable at robotics. 🤖 In just ONE DAY, we ran hundreds of head-to-head comparisons in the wild, and it climbed to #1 in @RoboColosseum's SO-101 Elo rankings. 🏆 Watch it in action 👇

Neither post explains how Duan's "hundreds" relates to the 122 comparisons RoboColosseum gave for GPT-6.2 Astra. Duan's post includes a video of about 12 seconds, but ORIENS has not checked the content of the video.

Claude Fable 5.2 second at 1,803; previous top model GPT-6 Astra fifth

The leaderboard image RoboColosseum attached to its post lists scores for 10 models. Anthropic's "Claude Fable 5.2" was second at 1,803, and Anthropic's "Claude Opus 5.5" was third at 1,666. GPT-6.2 Astra led second place by 181 points. OpenAI's previous top model "GPT-6 Astra" was fifth at 1,393, and xAI's "Grok 4.7" was sixth at 1,366. The scores were computed at 11:46 a.m. Nov. 20 Pacific time (4:46 a.m. Nov. 21 Japan time), and a dotted line in the image marks 1,500.

In follow-up posts, RoboColosseum showed videos of two evaluation runs. They place Claude Fable 5.2 and GPT-6.2 Astra side by side for the instruction "Find the only bottle facing a different direction," and Grok 4.7 and GPT-6.2 Astra for "Knock down the yellow box." The text of the posts does not say which model won each run. ORIENS has not checked the content of the videos.

RoboColosseum's Open track leaderboard for SO-ARM101. It ranks 10 models by DCR rating (Elo scale) and shows each model's number of comparisons (N) and error bars (±1 posterior standard deviation)
RoboColosseum's Open track leaderboard for SO-ARM101. It ranks 10 models by DCR rating (Elo scale) and shows each model's number of comparisons (N) and error bars (±1 posterior standard deviation) (Source: RoboColosseum (@RoboColosseum) post on X)
SO-ARM101 Open track scores (Elo scale, no maximum). Reproduced from the leaderboard image RoboColosseum posted on X, computed at 11:46 a.m. Nov. 20 Pacific timeScore
1984
122GPT-6.2-Astra
1803
112Claude-Fable-5.2
1666
187Claude-Opus-5.5
1525
194Pi 0.5
1393
183GPT-6-Astra
1366
181Grok-4.7
1259
214LingBot-MolmoAct2
1233
219GR00T N1.7
1122
216MolmoAct2
1095
207G0.5

GPT-6.2 Astra's 122 comparisons are the second fewest among the 10 models. The fewest were Claude Fable 5.2's 112.

SO-101 is the name of the robot; tasks not disclosed

The link to the leaderboard page in the third post of RoboColosseum's thread carries the description "Robot policies ranked on real robots in the wild, with one leaderboard per robot for the Open track and the Fine-tune track. Updated hourly." The link's URL specifies SO-101 as the robot (robot=so101), and the leaderboard image labels it "SO-ARM101." SO-101 is the name of the robot used in the evaluation.

The posts do not explain the difference between the Open and Fine-tune tracks, what "DCR rating" in the image means, the full set of tasks the robots performed, or how the models were connected to control the robots. When ORIENS retrieved RoboColosseum's leaderboard page, the leaderboard content did not load, so these points could not be confirmed. RoboColosseum says the leaderboard is updated hourly, and the rankings after the computation time of 4:46 a.m. Nov. 21 Japan time (11:46 a.m. Nov. 20 Pacific time) could not be confirmed either.

Brockman shares the result; official page has no mention of robotics

OpenAI President Greg Brockman quoted Duan's post on X at 10:36 a.m. Nov. 21 Japan time, writing only "GPT-6.2 for robotics." It is a mention by an OpenAI executive.

GPT-6.2 for robotics:

Meanwhile, the GPT-6.2 announcement page OpenAI published on Nov. 17 contains no mention of robotics, RoboColosseum or SO-101. ORIENS checked this announcement page, and found no announcement on OpenAI's official accounts or elsewhere. The ranking comes from RoboColosseum's evaluation and could not be confirmed as an announcement by OpenAI itself.

Brockman has continued to post about outside evaluations of 6.2 in the same period. At 7:05 p.m. Nov. 20 Japan time, he quoted a post saying GPT-6.2 Astra exceeded a human baseline on a benchmark that measures physical 3D understanding. At 5:30 a.m. Nov. 21, he quoted another post and wrote "GPT-6.2 Astra is great at robotics." ORIENS has not confirmed who wrote that quoted post or what the evaluation involved.

6.2 launch emphasized document work and code

OpenAI released GPT-6.2 Astra as a new frontier model for the hardest projects. The announcement emphasized gains in the quality of slides and documents, code and cybersecurity capabilities, and set the API price at $8 for input and $40 for output per million tokens.

Before the release, on Nov. 1, the label "gpt-6.2-astra" had appeared in Codex usage logs. In research on running robots with general-purpose AI models, Tsinghua University and others released a paper on a framework for operating robots with the open model "Qwen" on Nov. 1.