ECHO MAZE

EMZ BENCHMARK · MEMORY IN MOTION

Memory,
measured.

Every model enters the same hidden mazes with the same rules. The only way out is to observe, remember, and move.

Results from published, replayable runs.
SUITEemz-public-v0
MODELS
5
MAZES
9
RUNS
45
PROTOCOL
v0.8 · low effort

LEADERBOARD

Same maze.
Different memory.

Ranked by solve rate, then path efficiency. Open the Replay workspace to inspect every recorded decision.

More metricsFewer metrics
01
x-ai

grok-4.6

reasoning · low
100%9 of 9 scored
easy3/3
medium3/3
hard3/3
AVG SPLSuccess weighted by path efficiency. A score of 1.0 means every scored maze was solved using an optimal path.
0.500
MEDIAN TURNSThe median number of turns used by solved runs. Lower is faster, but maze routes vary in length.
64
WALL HITS / 100Blocked moves per 100 recorded move attempts across all runs.
0.0
FORMATShare of non-API-failure runs that completed without an invalid structured response.
100%
API FAILURESRuns terminated by an unresolved provider or transport failure.
0
RUN EVIDENCE
02
anthropic

claude-sonnet-5.5

reasoning · low
78%7 of 9 scored
easy3/3
medium3/3
hard1/3
AVG SPLSuccess weighted by path efficiency. A score of 1.0 means every scored maze was solved using an optimal path.
0.538
MEDIAN TURNSThe median number of turns used by solved runs. Lower is faster, but maze routes vary in length.
37
WALL HITS / 100Blocked moves per 100 recorded move attempts across all runs.
0.2
FORMATShare of non-API-failure runs that completed without an invalid structured response.
88%
API FAILURESRuns terminated by an unresolved provider or transport failure.
1
RUN EVIDENCE
03
openai

gpt-5.6-luna

reasoning · low
56%5 of 9 scored
easy2/3
medium1/3
hard2/3
AVG SPLSuccess weighted by path efficiency. A score of 1.0 means every scored maze was solved using an optimal path.
0.332
MEDIAN TURNSThe median number of turns used by solved runs. Lower is faster, but maze routes vary in length.
53
WALL HITS / 100Blocked moves per 100 recorded move attempts across all runs.
3.3
FORMATShare of non-API-failure runs that completed without an invalid structured response.
100%
API FAILURESRuns terminated by an unresolved provider or transport failure.
0
RUN EVIDENCE
04
google

gemini-3.8-flash

reasoning · low
22%2 of 9 scored
easy1/3
medium0/3
hard1/3
AVG SPLSuccess weighted by path efficiency. A score of 1.0 means every scored maze was solved using an optimal path.
0.150
MEDIAN TURNSThe median number of turns used by solved runs. Lower is faster, but maze routes vary in length.
52.5
WALL HITS / 100Blocked moves per 100 recorded move attempts across all runs.
0.0
FORMATShare of non-API-failure runs that completed without an invalid structured response.
22%
API FAILURESRuns terminated by an unresolved provider or transport failure.
0
RUN EVIDENCE
05
deepseek

deepseek-v4.1-flash

reasoning · low
11%1 of 9 scored
easy0/3
medium1/3
hard0/3
AVG SPLSuccess weighted by path efficiency. A score of 1.0 means every scored maze was solved using an optimal path.
0.111
MEDIAN TURNSThe median number of turns used by solved runs. Lower is faster, but maze routes vary in length.
27
WALL HITS / 100Blocked moves per 100 recorded move attempts across all runs.
0.0
FORMATShare of non-API-failure runs that completed without an invalid structured response.
13%
API FAILURESRuns terminated by an unresolved provider or transport failure.
1
RUN EVIDENCE
SolvedUnsolved · max turnsInvalid outputAPI failure