The best general-purpose AI models got about twice as good at steering a real robot between July and October, with performance doubling roughly every 54 days. That is the headline finding of a LessWrong post published on 26 October 2026 by chrisleet and bira, who also submitted a paper to a robot-learning workshop at CoRL.
The authors put nine multimodal models from OpenAI, Anthropic, and Google DeepMind (fifteen configurations in all) in charge of a two-armed I2RT YAM robot. Each model saw three camera feeds and the arms’ own position readings, then issued commands that set where each gripper should go, how it should be angled, and how far it should open. Six tasks came from the RoboDojo benchmark, including stacking bowls, capping a pen, and sorting objects. Every model got 20 attempts per task, with a budget of 40 actions each.
The scale of improvement is easy to state. The best execution rate grew 4.4 times over, while the cheapest cost per completed subtask dropped to under a sixth, a 6.2-fold cut. The 95 percent confidence interval on the 53.9-day doubling time spans 36.1 days at the quick end and 59.5 at the slow end, so the true figure could be considerably faster than the headline. The authors also note that it is unclear whether Gemini 4 Argon should count as a new record, since its lead over Opus 5.5 is only 0.13 points.
Absolute performance is still modest. The strongest models averaged 45 to 48 percent on the task scores, which the authors describe as steady partial progress rather than frequent finished jobs. Speed and price pull in different directions. GPT-6 Astra in its fastest setting finishes a subtask in 1.1 minutes at $4.51 apiece. Gemini 3.8 Flash charges $0.38 apiece and takes 4.18 minutes. Nobody wins on both.
What separates good from weak is mostly economy of motion, not understanding. In the bowl-stacking task, every model worked out that the gripper could not span the bowls and had to pinch the rim from the side. The sensor value after a good pinch resembled an empty grasp, though. The five stronger models lifted the bowl to check, while the weaker ones burned their action budget regrasping. Those five completed the task in at least 70 percent of trials, and the four weaker models in at most 25 percent. Failures were physical and small: misjudged distances, drooping arms, camera angles that distorted depth. In one trial Opus 5.5 squeezed a pen cap so that it shot out of reach.
The post also covers a safety record. It cites prior work in which GPT-6 Astra scored 80 percent or better in seven of eight configurations on real-arm tasks, and another team that stopped real-arm experiments with the same model after 33 trials because it kept commanding unsafe moves that damaged the hardware. The authors raise a larger worry too: wet-lab work is a practical barrier that keeps lone actors from synthesizing bioweapons, and a model that can drive lab arms lowers it. Their estimate is that a typical cell of two arms with cameras can handle basic chemistry, and ten fixed arms can run a wet lab.
In the comments, chrisleet added results for three purpose-built robot models. The best of them, Dream Lab’s Dream 1, scored 1.3 times higher and ran 1.7 times quicker than the top general-purpose systems, while NVIDIA’s GR00T N1 and the Allen Institute’s MalmoAct2 were beaten by the top chatbots. That matters commercially: a chatbot you already pay for is within reach of a specialist built for the job.
RoboDojo’s human teleoperators finish a task in 25 to 40 seconds. The authors call their own extrapolation “very naive,” and it puts language models level with those operators by mid-2027. Robotics teams choosing between a custom controller and a general model API should re-run that comparison every quarter, because the cost side of the trade is falling faster than the capability side is rising.
Reported by LessWrong on 26 October 2026, in a post by chrisleet and bira.
