Bagel Labs says its robot model beats GPT-6 Astra, but the benchmark shows a gap
Toronto-based Bagel Labs says its open-weight model Bagel-Robotics-1 ranks third on the RoboDojo simulation leaderboard and beats OpenAI's GPT-6 Astra in four areas. Tech Times checked the per-suite results and found a fifth area the press release leaves out. There, Bagel completes 3.67 percent of open-instruction tasks against 31 percent for Astra.

Bagel Labs, a company from Toronto, announced Bagel-Robotics-1 on 19 November. It is an open-weight model for controlling robots. In its press release the company says the model ranks third worldwide on the RoboDojo benchmark and scores higher than OpenAI's GPT-6 Astra in precision, memory, adapting to unfamiliar situations and multi-step tasks. RoboDojo's news feed lists the model with an average score of 38.65 and a success rate of 31.91 percent in simulation.
Tech Times compared the results suite by suite and found what the release leaves out. RoboDojo has a fifth suite that tests open-vocabulary instructions. There Bagel succeeded in 44 of 1,200 episodes, or 3.67 percent, against 31 percent for Astra. On partial score Bagel leads Astra in all four named areas, but on success rate Astra is ahead in generalization by 2.17 points. "Third" also means third among learned policies. Agent systems are listed separately on the board and score far higher, up to 74.20. Tech Times could not load the live leaderboard to confirm the current order.
Why it matters
Leaderboard claims are becoming a standard part of robot AI announcements, and this one shows why the details matter. A model can lead on precise, long tasks and still fail almost every time when the instruction is unfamiliar. All of these results come from simulation, so they say little yet about how either model handles a real robot.