Claude Fable 5.1 versus Codex (Astra) on the ESC G431

Two AIs, the same board, the same spec, the same prompt, the same flow (Adom's AI Flow), measured the same way: the AI's own time from the prompt to "done, here is your video", every return to an earlier step, every clip, and the analysis each AI did itself. Nobody had put two engines through one measured board flow before; this page is the first pair.

The side by side, 186 s: Claude Fable 5.1 on the left, Codex (Astra) on the right, step by step in the flow's order; each run's own clocks in its own corner (AI time, the idle cut out); both AIs' analysis drawings side by side; each step's AI-minute bars below; the words cut to the picture.

The number: prompt to "done, here is your video"

AI time is the sum of the run's sessions, one per human prompt, from the prompt to the AI's done, with every idle gap over 15 minutes (the human away) cut out. Follow-up prompts after delivery (the walkthroughs, the drawings, the re-cuts) are on each run's page under Sessions and are not in the delivery number.

Per step, AI time

The analysis each AI did itself

Each AI solved the current and the heat on its own board and drew it; the pictures are theirs, not the binary's. Left, then right, per step.

Claude Fable 5.1: Current density on the copper, the AI's own DC conduction solve on the filled polygons of every loaded net at its spec current: the 20 A nets peak at 230 to 320 amps per square millimetre where a one millimetre track carries a phase, the pours sit far lower.

Codex (Astra): Independent four-layer DC conduction solve of native filled polygons, tracks, pads and 473 plated barrels: F.Cu and B.Cu current density, with each loaded net peak annotated. The 0.1 mm stress-case solve gives 2.327 W copper loss; +VBAT peaks at 494 A/mm² and GND_OUT at 436 A/mm², with mesh-sensitive corner peaks and 0.3% total-loss change on refinement.

Claude Fable 5.1: Heat flow in the copper, the AI's own four-layer heat solve read on F.Cu and B.Cu like the current picture: each hot tab's watts spreading along its layer and going down its vias, the vias drawn. Q5 sends 35 percent of its heat down three vias; Q3 has no thermal via within 3 mm and sends none, so it is the hottest tab.

Codex (Astra): Own four-layer heat solve carries 6.427 W from the hot tabs and copper losses through native copper, dielectric and 473 plated barrels. F.Cu and B.Cu show watts per millimetre of width with via dots; each chip is labeled with its own along-layer and down-via heat within 4 mm, with full balanced splits and mesh sensitivity in the numerical report.

Claude Fable 5.1: Temperature rise over the board, the same four-layer solve read as temperature on the top face, the FR4 spreading: 4.1 W from the seven hot tabs plus 3 W of Joule loss. In still air the FET tabs sit 105 to 121 degrees above ambient, over the 60 degree budget; in moving air the peak is far lower.

Codex (Astra): The independent coupled solve fails the 60 C hot-tab rise budget: Q6 reaches +121 C in still air and +70 C with moving air; board peaks are +122.5 C and +70.7 C. Top-face temperature from the same 6.427 W solve uses h=10 and 25 W/m²K; results remain mesh-sensitive and are predictions under stated assumptions, not measured junction temperatures.

The two runs, in full

• Claude Fable 5.1: the run page: every step's clip, screenshots, sessions, returns, decisions, plan usage, what it gave back.
• Codex (Astra): the run page: the same for the other engine.
• The flow: Adom's AI Flow, the tool both ran.

Plan usage on each run page is per account, not per run: every AI thread signed in as that user draws on the same limits.