Two benchmarks. 286 everyday DROID tasks, and bin-to-bin pick and place of four household objects timed against a human teleoperator.
As of November 17, 2026, Cosmos3-Nano-Policy leads with 68.1, followed by π0.5 (62.4) and GR00T N1.7 (62.3). Official leaderboard ↗ · Paper ↗
Leaderboard
- Tasks
- 410
- Attempts / task
- 1
- Median cost / task
- $0.042
- Median time / task
- 46 s
DROID tasks vs speed
Published time per task attempt
Cost isn’t published
PhAIL doesn’t report cost per attempt for enough models to chart. The table shows Playground’s estimates, marked est.; Methodology explains how they’re made.
Reliability (%). DROID tasks scores how much of each task was done within 240 s, PhAIL's headline, from its ranking model over 286 tasks. Pick and place scores the items placed within 120 s each. The index uses DROID tasks by 480 s, the longest limit PhAIL publishes: the headline plus each model's gain from 240 s to 480 s on PhAIL's own success curve. Only GPT-6 Astra ran past 240 s.
Cosmos3-Nano-Policy68.1—68.1$0.045est.49 s2
π0.562.475.162.4$0.031est.34 s3
GR00T N1.762.3—62.3$0.031est.34 s4
FLUX 3 Action61.7—61.7$0.038est.42 s5
GPT-6 Astra55.9—68.5$17.74est.2 min 26 s6
MolmoAct244.7—44.7$0.042est.46 s7
GalaxeaVLA G0.539.0—39.0$0.055est.60 s—
ACT—63.7————
SmolVLA—32.0————
GR00T N1.6—72.3———Tasks
410 tasks. Hover a task to play its clip (397 of 410).
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ ClipMethodology
- Scoring
- Generic ladder
- Time cap / attempt
- 4 min 00 s
- Control rate
- —
- Models in index
- 7 of 10
Scoring. Operator places each episode on a ladder: contact 0.1, control 0.3, at target 0.4, success 1.
Who runs it. Positronic runs every episode and publishes its video.
Robots. Franka FR3 (DROID setup), with a Robotiq gripper for pick and place.
Note. DROID scores come from PhAIL's ranking model over all 286 tasks. The index counts DROID at 480 s, the longest limit PhAIL publishes, rather than its 240 s headline. The pick-and-place GR00T is GR00T N1.6.
Time per attempt. Published or measured for 7 of 7 models.
Cost per attempt. Not published. Estimated otherwise: GPT-6 Astra's median published cost per attempt on other boards; one H100 at $3.29 per hour (Lambda on-demand, Oct 2026) × time. GPU price source ↗
