RoboCurve times Astra's zero-shot picks at two and a half minutes each

Two and a half minutes a cycle at 2.1k output tokens a run comes to roughly 50,400 output tokens an hour of arm time ^[3][11]. Run that arm pair around the clock and it emits about 1.21 million output tokens a day ^[18]. The account has the token count but no price per token, so the hourly inference bill cannot be worked out from it ^[17].

The physical evidence base is twenty attempts ^[2]. Eighteen successes out of twenty is 90% ^[12]. One more miss would have produced a much duller figure. cryptobriefing.com attributes the three points between the simulator and the bench to unexpected lighting and imperfect surfaces ^[15].

The humanoid item in the account is a zero-shot cola-bottle pickup in simulated control, translated from human demonstrations using camera input only ^[6]. The article's headline says Astra demonstrated zero-shot pick-and-place on a full humanoid robot ^[16]. Its text says comprehensive documentation of performance on full humanoids had yet to be revealed, with explorations still on arms and simulators ^[9]. The 95% belongs to the arms.

What a per-task pipeline buys is data. According to cryptobriefing.com, Astra used none of the usual kind: no privileged access to the robot's internal state, no specialized robotic datasets, just camera views and proprioception fed into actions ^[5]. The new spending line is tokens per pick plus somebody to clear the failures. A one-in-twenty miss rate at this cadence works out to 1.2 interventions an hour per arm pair ^[13].

RoboCurve also ran the predecessor Claude Fable variants and reports Astra ahead of them on speed, token efficiency and reliability, particularly on gross motor tasks ^[14].

In my view the binding number here is the two and a half minutes. The 95% is not. A dedicated pick cell is cheap per pick once it is running and expensive to set up for each new object. A general model that needs no dataset per object competes against that setup cost, and at this cadence setup cost is the only place it competes. The counter-thesis is that cycle time and token count per task fall faster than robot hardware costs do, and this evaluation already shows a generation-over-generation gain on both ^[14]. What would change my view is a physical sample in the hundreds holding at the same rate, or a published cycle time under a minute.