glm 5.3 vs deepseek v4-pro: both saw the maze, neither escaped

deepseek v4-pro – bare through the @OpenRouter api, reasoning capped at 8k, $0.08 for two runs
glm 5.3 – run by hand inside zcode with its own agent harness, thinking level low

we put the two models on one job: escape a dark 9x9 maze in first person, 81 cells, 2 keys, no map.

the models setup: identical rules text for both. one move is one command – straight, left, right, back – and the model is only asked at a junction or a dead end. the corridor in between is walked for it, so a single move can cover one cell or twenty-nine. the exit stays locked until both keys are collected. a pig in a suit hunts elon musk: it starts 12 cells away, stands still while elon makes progress, and closes half the distance you waste. 80 moves max, two mazes

the model can score up to 100 points – 30 for keys, 30 for the exit, 20 for efficiency, 20 for clean moves. three scripted baselines ran the same two seeds.

- score, averaged over two mazes
#1 glm 5.3 – 37
#2 deepseek v4-pro – 36

- escapes
deepseek v4-pro, glm 5.3 never escaped

- real time, two runs
#1 glm 5.3 – 6m 25s
#2 deepseek v4-pro – 21m 27s

- tokens and cost
deepseek v4-pro – 120k tokens, 90k of them reasoning, $0.08
glm 5.3 – 1.17m tokens through zcode, not comparable to the line above

observations:

• every maze hides 2 keys, and the exit will not open without both. neither model ever got both. glm found one key in one maze. deepseek found none in either.

• glm saw the maze and lost anyway. on the second run it revealed 80 of 81 cells and walked 180 of them before the pig got it. the map was there, the memory of it wasn't.

• the pig only closes in on a wasted move, and a good move opens the gap back up – the models wasted less and lost anyway, because between mistakes they were not going anywhere. it also walks the whole clip. that's the camera – in the rules it only moves when you waste a move.

• the losing move is always the same one: back into a corridor already walked. deepseek's second move on seed 7 ran 6 cells into ground it had just covered, revealing nothing. it was told that stretch was open and visible. it went anyway.

• worth flagging: this is not a head-to-head. glm played inside zcode with its own harness at low thinking, deepseek played bare through the api with an 8k reasoning cap. the token counts are not comparable in either direction. both are measured against the same scripted baselines, not against each other.

conclusion: 37 vs 36 over two mazes, 1 key out of 4 and zero escapes – both models could see the maze and neither could remember it

watch the full test via link