Artificial Analysis published AA-Robotics v0.1 on 14 October 2026, a research test of language models controlling a robot arm through coding agents. Its first version contains eight tasks. GPT-6 Astra receives an average score of about 69 out of 100, taking approximately 30 minutes and US$16.36 per task according to the evaluator. This is not a 69% success rate for arbitrary real-world work.
The average combines full and partial completion. That distinction matters: moving an object approximately toward its destination and satisfying every condition are different outcomes. The ranking measures a particular setup with particular models and software, rather than certifying a universal robot.
Useful questions behind the overall score
Our editorial interpretation is that physical tests should show more than the final arrangement. Practical work depends on elapsed time, recovery requirements and whether a person intervened. A system that eventually reaches the objective after a long search can be interesting research while remaining unsuitable for a particular operating pace.
Cost is another axis. A more expensive attempt might be justified for a rare, difficult task but not necessarily for a constantly repeated action. These are conditional considerations, not a calculated business outcome for the tested models. A comparison should begin with what constitutes an accepted final result in the user's own process.
Knowing when to stop is particularly important. Once an objective has been reached, an unnecessary adjustment can spoil the arrangement. An evaluation should therefore distinguish reaching an acceptable state from correctly recognising that the task is complete. That criterion matters regardless of which model leads the aggregate table.
Reading an early version
Eight tasks are an initial sample, not coverage of all robotics. Before drawing an application-specific conclusion, readers need to examine objects, equipment, acceptable errors and repeatability. A change to one of those elements can change the difficulty even when the natural-language instruction looks much the same.
We have not reproduced these trials and do not present the data as our own measurement. The publication is dated 14 October and is included as a valuable uncovered test, not a launch today. In later versions, useful progress would mean more than a higher total score: it would explain more clearly under which conditions an agent finishes reliably, how long it takes and what assistance it needs. That makes an early benchmark informative without asking it to answer questions outside its measured scope.
Editorially checked: 2026-10-17T06:16:31.145379+00:00 (UTC). Cover: conceptual AI illustration; not a photograph of an actual event.
