203X Uplink · Benchmark 26-08

Which AI is the best engineer?

Nine AI models each designed the same product alone in a sandbox: a VTOL drone that delivers blood. A hidden referee then measured every design in CAD, CFD, FEM, circuit-board and mission checks.

FINAL · 12 Nov 20269 models · effort high
WinnerClaude Fable 5.2Kestrel-B, a tail-sitter · 89.5 points
SecondClaude Opus 5.5QB-2, a quadplane · 88 points
Margin1.5 ptsof 120 possible

Fable 5.2's design needs the least energy in cruise (9.3 vs 10.8 Wh/km). Opus flies slightly further (30.3 vs 29 km), but within the 5 % tie band, so both score the same in the final. The win is close: one place lower in the public looks vote and Opus would have won.

Standings

Points per round, already multiplied by the round's weight. "Penalty" holds the CAD-gate and budget deductions.

Model · designLooks×2Air×4Strength×2Board×0.5Money×0.5Final×3PenaltyTotal
1 Claude Fable 5.2Kestrel-B · Anthropic 840161.5127 -4
89.5
2 Claude Opus 5.5QB-2 · Anthropic 1032124327 0
88
3 Claude Fable 5.1Bloodwing Q1 · Anthropic 1624102.252.518 0
72.75
4 Gemini 4 ArgonMedWing-VTOL 4C1C · Google 620204215 0
67
5 GPT-6 AstraALCYON 20 · OpenAI 201662.251.512 0
57.75
6 Kimi K3.1 ProMeridian · Moonshot AI 408440 -4
16
7 DeepSeek V4 ProSwiftStork VTOL · DeepSeek 1202050 -8
11
8 Grok 4.7Heron · xAI 200100 -1
2
9 GLM 5.3HemoWing HW-1 · Z.ai 0040.500 -5
-0.5

The nine designs

Referee renders of the CAD each model submitted. Every value is the referee's own measurement. No design passed the strength test.

Render of Kestrel-B by Claude Fable 5.2
#1

Kestrel-B

89.5 pts

Claude Fable 5.2 · effort high · Claude Code

CAD gate
FAIL
Take-off mass
11.2 kg
Cruise energy
9.3 Wh/km
Mission radius
29 km · claimed 43 km
Strength (min. SF, needs 1.5)
FAIL 2.96
Circuit board
10 / 12 checks
Run
208 min · $25.97
Referee notes

Gate: wing_skin 19.1 % of its surface < 0.8 mm (1st percentile 0.45 mm); fin_skin_dorsal 5.1 %

Strength: all three load cases pass (full thrust 2.96, 3.5 g 6.97, hard landing 7.65; wing_core) · vibration fails (modes in the rotor bands with hub participation)

Render of QB-2 by Claude Opus 5.5
#2

QB-2

88 pts

Claude Opus 5.5 · effort high · Claude Code

CAD gate
PASS
Take-off mass
10.78 kg
Cruise energy
10.8 Wh/km
Mission radius
30.3 km · claimed 36.4 km
Strength (min. SF, needs 1.5)
FAIL 0.30
Circuit board
12 / 12 checks
Run
243 min · $13.33
Referee notes

Strength: 3.5 g wing: foam tail fins SF 0.30, stabiliser 0.67, pod 0.83, wing 1.33 · hard landing pod 1.26 · full thrust 1.73 passes · vibration fails (mode 54.7 Hz within 15 % of the rotor's 63.8 Hz, motor mounts move 24 % in it) · tail sags ~8 cm at 3.5 g

Render of Bloodwing Q1 by Claude Fable 5.1
#3

Bloodwing Q1

72.75 pts

Claude Fable 5.1 · effort high · Claude Code

CAD gate
PASS
Take-off mass
11.33 kg
Cruise energy
12.4 Wh/km
Mission radius
25.4 km · claimed 47 km
Strength (min. SF, needs 1.5)
FAIL 0.29
Circuit board
11 / 12 checks
Run
40 min · $13.46
Referee notes

Strength: foam fins over the limit in every case (full thrust 0.29, 3.5 g 0.32, hard landing 0.67) · wing 0.94 in the hard landing · vibration passes

Render of MedWing-VTOL 4C1C by Gemini 4 Argon
#4

MedWing-VTOL 4C1C

67 pts

Gemini 4 Argon · effort high · Antigravity

CAD gate
PASS
Take-off mass
10.35 kg
Cruise energy
22.3 Wh/km
Mission radius
12.8 km · claimed 16.3 km
Strength (min. SF, needs 1.5)
FAIL 12.24
Circuit board
12 / 12 checks
Run
120 min · $16.16
Referee notes

Strength: load cases pass by far (lowest 12.24, wing at 3.5 g) · vibration fails (modes 86.0 / 107.0 Hz near the rotor's 101 Hz, 197.4 / 214.7 Hz near twice it)

Render of ALCYON 20 by GPT-6 Astra
#5

ALCYON 20

57.75 pts

GPT-6 Astra · effort high · Codex CLI

CAD gate
PASS
Take-off mass
11.98 kg
Cruise energy
26.4 Wh/km
Mission radius
11.2 km · claimed 35 km
Strength (min. SF, needs 1.5)
FAIL 0.12
Circuit board
11 / 12 checks
Run
53 min · $17.76
Referee notes

Strength: elevator over the limit in every case (0.12-0.18) · 3.5 g: wings 0.53 / 0.65 · load cases only, vibration not checked

Render of Meridian by Kimi K3.1 Pro
#6

Meridian

16 pts

Kimi K3.1 Pro · effort high · Kimi Code

CAD gate
FAIL
Take-off mass
11.49 kg
Cruise energy
no result
Mission radius
did not finish · claimed 15.2 km
Strength (min. SF, needs 1.5)
FAIL 0.28
Circuit board
12 / 12 checks
Run
70 min · $4.36
Referee notes

Gate: thin walls: main wing 26.6 % of its surface under 0.8 mm (also 3 lift mounts just over 5 %)

Strength: main wing over the limit (3.5 g 0.28, full thrust 0.31) · centre frame 0.54 in the hard landing · load cases only, vibration not checked

Render of SwiftStork VTOL by DeepSeek V4 Pro
#7

SwiftStork VTOL

11 pts

DeepSeek V4 Pro · effort high · OpenCode

CAD gate
FAIL
Take-off mass
10.44 kg
Cruise energy
no result
Mission radius
did not finish · claimed 18 km
Strength (min. SF, needs 1.5)
FAIL < 0.01
Circuit board
2 / 12 checks
Run
40 min · $2.26
Referee notes

Gate: 104 part overlaps (takeoff + cruise), among them its cruise propeller partly inside its own fin (2,623 mm³, both positions) · its propeller disc 1.9 mm from the tail (10 mm required)

Strength: tail not attached (horizontal tail + booms 13.4 mm, fin 34 mm from the airframe) -> loose parts, fuselage SF < 0.01 · load cases only, vibration not checked

Render of Heron by Grok 4.7
#8

Heron

2 pts

Grok 4.7 · effort high · Grok Build

CAD gate
PASS
Take-off mass
11.64 kg
Cruise energy
no result
Mission radius
did not finish · claimed 28 km
Strength (min. SF, needs 1.5)
FAIL < 0.01
Circuit board
8 / 12 checks
Run
138 min · $30.24
Referee notes

Strength: left motor boom (both left motors) not attached: 11.2 mm gap, beyond the referee's 7 mm bridge -> loose part, SF < 0.01 · vibration passes

Render of HemoWing HW-1 by GLM 5.3
#9

HemoWing HW-1

-0.5 pts

GLM 5.3 · effort high · OpenCode

CAD gate
FAIL
Take-off mass
11.64 kg
Cruise energy
no result
Mission radius
did not finish · claimed 28.8 km
Strength (min. SF, needs 1.5)
FAIL 0.09
Circuit board
6 / 12 checks
Run
223 min · $32.54
Referee notes

Gate: fuselage shell walls < 0.8 mm on 13.3 % of its surface (GNSS mast 98.9 %)

Strength: landing-gear feet 0.09 (hard landing) · motor mounts 0.13 (full thrust) · fuselage shell 0.25 (3.5 g) · load cases only, vibration not checked

Claimed vs measured range

Each model predicted its own mission radius. The referee then flew the mission in simulation with the measured CFD and mass.

RefereeModel's claim×Did not finish
0 km10 km20 km30 km40 km50 kmClaude Fable 5.2Claude Fable 5.2: referee 29 km, claimed 43 km29 km (claimed 43)Claude Opus 5.5Claude Opus 5.5: referee 30.3 km, claimed 36.4 km30.3 km (claimed 36.4)Claude Fable 5.1Claude Fable 5.1: referee 25.4 km, claimed 47 km25.4 km (claimed 47)Gemini 4 ArgonGemini 4 Argon: referee 12.8 km, claimed 16.3 km12.8 km (claimed 16.3)GPT-6 AstraGPT-6 Astra: referee 11.2 km, claimed 35 km11.2 km (claimed 35)Kimi K3.1 Proclaimed 15.2 · did not finishDeepSeek V4 Proclaimed 18 · did not finishGrok 4.7claimed 28 · did not finishGLM 5.3claimed 28.8 · did not finish

Looks: the public vote

375 blind pairwise votes from 14 voters, closed 12 Nov 2026 at 15:23 CET, fitted with Bradley–Terry (Elo scale).

  1. 1ALCYON 20 GPT-6 Astra1832
  2. 2Bloodwing Q1 Claude Fable 5.11708
  3. 3SwiftStork VTOL DeepSeek V4 Pro1686
  4. 4QB-2 Claude Opus 5.51537
  5. 5Kestrel-B Claude Fable 5.21460
  6. 6MedWing-VTOL 4C1C Gemini 4 Argon1455
  7. 7Meridian Kimi K3.1 Pro1425
  8. 8Heron Grok 4.71300
  9. 9HemoWing HW-1 GLM 5.31096

Against the AI Index

Artificial Analysis Intelligence Index, fetched 11 Nov 2026. It lists seven of the nine; places are counted among those seven. Rank correlation (Spearman) with our order: 0.75.

ModelIndexIndex placeOur place
Claude Fable 5.26211
GPT-6 Astra5525
Claude Opus 5.55232
Gemini 4 Argon5044
Claude Fable 5.14753
Kimi K3.1 Pro4666
Grok 4.73577

Not listed: DeepSeek V4 Pro, GLM 5.3. Fable 5.1's 47 is the 8 Nov value; it was no longer listed on 11 Nov.

How it was scored

We do not have iterations here, PCB and CAD gates can be rather easily fixed, the design and aerodynamics is where the model really needs to stand out.NOAH KLARMANN, ON THE WEIGHTS