Jev / GPT-6 Astra / GPT-4.1 mini

One apple-to-plate trial per controller in the same seed-0 MuJoCo scene. The video combines the latest Jev/GPT-6 pair with the earlier GPT-4.1 mini recording. No new model calls were made for this edit.

Controller	Outcome	Cycles	API calls	Model cost	Wall time	Simulation time
Jev 1.13	Placed	113	226	$0.018825	181.847 s	36.16 s
GPT-6 Astra	Placed	106	212	$5.933624	707.274 s	33.92 s
GPT-4.1 mini	160-cycle limit reached	160	320	$0.288512	704.253 s	51.20 s

The three recordings use the same initial state, physics implementation, action options, control instructions and 160-cycle limit. Every recorded intent and motor question was checked against the current shared interface. Source hashes, model response IDs, cumulative costs and prior exact physical replay audits were verified.

GPT-6 uses low reasoning effort and 4,096 maximum completion tokens. Mini uses temperature 0 and 850 maximum tokens. The HTTP timeout was 90 seconds for the latest pair and 35 seconds for the earlier mini run; no recorded trial ended in a request error.

Playback uses a shared simulation-time clock. Network waiting is excluded from playback but included in the displayed wall time. Finished controllers hold their actual final frames. The video shows the complete 51.2-second mini trial, with a three-second intro and five-second final hold.

Inputs are simulator geometry and contacts. The controllers choose intent, X/Y/Z direction and gripper commands; the shared executor uses bounded increments and inverse kinematics. No VLA or full pick-and-place skill is used.

Jev probabilities are native model outputs; GPT probabilities are self-reported JSON values. Neither is calibrated task-success confidence. This one-trial-per-controller comparison cannot establish general success rates.

Sources:

• Jev 1.13: pair 20260919-200239-715974-0, run 20260919-200239-0-jev.
• GPT-6 Astra: pair 20260919-200239-715974-0, run 20260919-200239-0-llm.
• GPT-4.1 mini: pair 20260919-193012-198478-0, run 20260919-193012-0-llm.