Physical AI Analysis

Two robot arms, two interlocked claw pieces: get them apart without breaking either. 11 of 13 models we've run have failed it. GPT-6 Astra: apart in 53 min, $14.36. Kimi K3.1 Pro: still locked after 101 min, $13.45.

Spot, a lidded bin, three objects: open it, bin them, close the lid, stand clear. 20 random starts. GPT-6 Astra: 20/20, 68 min, $30.20. Kimi K3.1 Pro: 0/20. Bins all three on seed 0, never shuts the lid. 179 min, $32.09. 2× speed.

Pick up an apple by the stem, five-finger hand, one attempt each. Both lifted it. Kimi K3.1 Pro: lifted at 6.8 s, 9 min to build, 34 cents. GPT-6 Astra: lifted at 13.9 s, 11 min, $3.48. All 17 models we've run on it ranked at the end.

DeepSeek V4.1 Flash 1017 vs 4 models: carry an apple over a wall and set it on the plate, 5 random starts. GPT-6 Astra, Haiku 5.5, Gemini 4 Argon, Opus 5.5, one after another. Everyone lands it. DeepSeek does it for 7 cents. 2× speed.

Six models, one apple, one rule: pick it up by the stem. Same prompt, same five-finger hand, one attempt each. Three rounds, then where all 15 models we’ve run on it landed. Gemini 4 Argon lifted first at 4.7 s. Grok 4.7 never got it off the table.

Three face turns on a Rubik’s cube, five-finger robot hand. Both landed all three. GPT-6 Astra: 43 min to build, last turn at 49.8 s, $15.97. Claude Opus 5.5: 32 min, last turn at 30.8 s, $4.75. All 14 models we’ve run on this task ranked at the end.

Two robot arms, two interlocked claw pieces, one rule: get them apart without breaking either. 9 of 11 models we’ve run have failed this one. GPT-6.1 Sol: apart in 61 min, $1.63. Claude Opus 5.5: still locked after 93 min, $17.39.

GPT-6 Astra wrote a controller for Boston Dynamics Spot (MuJoCo) that walks to a table, picks up a block, carries it to a mat and sets it down. 20/20 randomised seeds, first attempt, 21 minutes of agent time, $6.98 of API cost.

GPT-6 Astra screwing a light bulb into a lamp socket with a Franka arm, in simulation, until it lights. One attempt, 108 minutes of agent time, $38 of API cost. Task from EmbodiedSWE. Score 1.0.

Pinch an apple by its stem and lift it clear of the table, with a five-finger robot hand. Both models managed it. GPT-6 Astra: built in 11 min, lifted at 13.9 s. Gemini 4 Argon: built in 23 min, lifted at 4.7 s, cheaper.

Three face turns on a Rubik’s cube with a five-finger robot hand. Both models landed all three. GPT-6 Astra: 43 min to build, last turn at 49.8 s. Gemini 4 Argon: 86 min to build, last turn at 24.2 s, half the cost.

GPT-6 Astra (high) vs Gemini 4 Argon (high) Astra: 11m40s Argon: 24m05s 3D assembly, PCB routing and schematics. Source circuit checks: 0 errors each; Astra has 1 warning.

Same task, same 60-min budget, one attempt each: build and animate a cutaway turbofan in Onshape. GPT-6 Astra (max): 58 min. Gemini 4 Argon (high): hit 60 min with required exports and reports unfinished.

Same task, same 120-min budget, one attempt each: separate two interlocked claw pieces without breaking either. GPT-6 Astra: done in 53 min, interlock 71→0. Gemini 4 Argon: still interlocked after 85 min, 71→22.