Micro Mouse benchmark
Micro Mouse, the daily-maze robot racer, driven with the game's own controls while the clock is paused. Micro Mouse is made by Chach16. The bench runs the game's own engine and times every run the way its score server does, by replaying the input tape. The headline is the daily mazes people raced, against the median person's best time. The campaign mazes add depth, against the game's own bot.
Micro Mouse leaderboard
Select a column to sort. Scroll sideways for all metrics →
| # | Model | Score | 95% interval | Score per tierDayWrkCldFstNit | Campaign | Beat the median | Reached | Wall hits / run | Replies / run | Runs | API cost / run | Sec / reply |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT-6 Astrareasoning medium | 755 of 14 | 7586777985 | 8259 to 104 | 0 of 1 | 5 of 5 | 2.0 | 26.4 | 5 | $1.19 | 7.8 | |
| 2 | GPT-6.1 Solreasoning medium | 715 of 14 | 7177857188 | 8058 to 102 | 0 of 1 | 5 of 5 | 0.8 | 27.2 | 5 | $0.332 | 8.5 | |
| 3 | GPT-6 Solreasoning medium | 365 of 14 | 3649646445 | 5540 to 71 | 0 of 1 | 5 of 5 | 12.0 | 45.8 | 5 | $0.468 | 13.5 | |
| 4 | GPT-6.1 Terrareasoning medium | 05 of 14 | 069654473 | 6345 to 80 | 0 of 1 | 4 of 5 | 15.4 | 39.8 | 5 | $0.108 | 4.5 | |
| 5 | GPT-6 Lunareasoning medium | 05 of 14 | 00000 | 00 to 14 | 0 of 1 | 0 of 5 | 50.6 | 80.6 | 5 | $0.036 | 7.4 |
No models match your search.
Score is the median person's time over the model's on each daily maze it reached, 0 on each it did not, averaged. Campaign is the same against the game's bot on the four campaign tiers, each weighted the same. A model ranks once it has driven every daily maze and the first 2 of each campaign tier; until then it shows how many of those it has driven and ranks after the models that have. Past that the bench keeps driving further campaign mazes in the same order for every model, so n grows and the campaign interval narrows.
The Pareto frontier shows where no other model scores at least as well for less cost or time, with one strict improvement. Partial sets are included and may cover different items. Use completed sets for direct comparisons.
By tier
Score on that tier's mazes, against the median person on the daily mazes and the game's bot on the rest, then the mazes reached, the pace on those and the wall hits per run. The count after each result is n, the runs finished in that tier.
| Model | Daily mazes | Workbench | Cold Storage | Fast Lane | Night Shift |
|---|---|---|---|---|---|
| GPT-6 Astrareasoning medium | 75vs the median person · n 1 · reached 1 · pace 75 · 2.0 wall hits | 86vs the bot · n 1 · reached 1 · pace 86 · 0.0 wall hits | 77vs the bot · n 1 · reached 1 · pace 77 · 5.0 wall hits | 79vs the bot · n 1 · reached 1 · pace 79 · 1.0 wall hits | 85vs the bot · n 1 · reached 1 · pace 85 · 2.0 wall hits |
| GPT-6.1 Solreasoning medium | 71vs the median person · n 1 · reached 1 · pace 71 · 3.0 wall hits | 77vs the bot · n 1 · reached 1 · pace 77 · 1.0 wall hits | 85vs the bot · n 1 · reached 1 · pace 85 · 0.0 wall hits | 71vs the bot · n 1 · reached 1 · pace 71 · 0.0 wall hits | 88vs the bot · n 1 · reached 1 · pace 88 · 0.0 wall hits |
| GPT-6 Solreasoning medium | 36vs the median person · n 1 · reached 1 · pace 36 · 20.0 wall hits | 49vs the bot · n 1 · reached 1 · pace 49 · 10.0 wall hits | 64vs the bot · n 1 · reached 1 · pace 64 · 8.0 wall hits | 64vs the bot · n 1 · reached 1 · pace 64 · 1.0 wall hits | 45vs the bot · n 1 · reached 1 · pace 45 · 21.0 wall hits |
| GPT-6.1 Terrareasoning medium | 0vs the median person · n 1 · reached 0 · pace n/a · 23.0 wall hits | 69vs the bot · n 1 · reached 1 · pace 69 · 11.0 wall hits | 65vs the bot · n 1 · reached 1 · pace 65 · 10.0 wall hits | 44vs the bot · n 1 · reached 1 · pace 44 · 17.0 wall hits | 73vs the bot · n 1 · reached 1 · pace 73 · 16.0 wall hits |
| GPT-6 Lunareasoning medium | 0vs the median person · n 1 · reached 0 · pace n/a · 37.0 wall hits | 0vs the bot · n 1 · reached 0 · pace n/a · 51.0 wall hits | 0vs the bot · n 1 · reached 0 · pace n/a · 63.0 wall hits | 0vs the bot · n 1 · reached 0 · pace n/a · 32.0 wall hits | 0vs the bot · n 1 · reached 0 · pace n/a · 70.0 wall hits |
The reference players on the same mazes
The median person raced only the daily mazes, so it has no campaign score.
| Player | Score | Campaign | Daily mazes | Workbench | Cold Storage | Fast Lane | Night Shift |
|---|---|---|---|---|---|---|---|
| The median personThe median person: the median over people of each person's best verified time on the daily board (up to five ranked tries), AI racers excluded. | 100 | n/a | 100reached 6 of 6 | n/a | n/a | n/a | n/a |
| The game's botThe AI racers' autopilot (ContentLab generators/micromouse/web/autopilot.js), the bot on the daily board: every reference route in every one of its 192 driving styles, the fastest finish. The game's level builder sets each level's bot time this way. | 94 | 100 | 94reached 6 of 6 | 100reached 2 of 2 | 100reached 2 of 2 | 100reached 2 of 2 | 100reached 2 of 2 |
| The game's test pilotThe game's own test pilot (tests/terrain-controller.js, drifting on): keyboard controls along the course's reference route, closed loop. | 48 | 50 | 48reached 6 of 6 | 52reached 2 of 2 | 49reached 2 of 2 | 50reached 2 of 2 | 50reached 2 of 2 |
| Random controlsRandom controls on the bench protocol: replies of up to 1 s, segments of 0.05 to 0.5 s, uniform throttle and steer, drift 20% of the time. | 0 | 0 | 0reached 0 of 6 | 0reached 0 of 2 | 0reached 0 of 2 | 0reached 0 of 2 | 0reached 0 of 2 |
Every maze
Every run
One square per maze, filled when it reached the dock. Columns run tier by tier, in the order every model meets them. Click a square to open that one.
| Model | Day | 2 | 3 | 4 | 5 | 6 | Wrk | 2 | Cld | 2 | Fst | 2 | Nit | 2 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-6 Astra | ||||||||||||||
| GPT-6.1 Sol | ||||||||||||||
| GPT-6 Sol | ||||||||||||||
| GPT-6.1 Terra | ||||||||||||||
| GPT-6 Luna |
MethodologyProtocol, scoring & limitations
The mazes
Micro Mouse is a robot racer by Chach16. Everyone drives the same seeded maze, and a time on its board is what the game's own engine measures replaying the run's input tape. The bench runs that engine, pinned to commit 87bf0cb, so a later change to the game cannot move anyone's result. Every run is stored as the game's own input tape, and the game's own verifier replays it to the same time.
The headline tier is 6 daily mazes people raced on the board in October, each by 20 to 46 people. The median person's time on a maze is the median, over the people who raced it, of each person's best verified time from up to five ranked tries. AI racers are left out. Our real-time benchmark in development uses the same mazes and the same medians, played from screenshots in real time, so a model's two results compare directly.
The other four tiers are the game's campaign worlds after the opening. Each holds that world's ten levels, then more mazes of the same size and floors from the seed family the game's level builder draws from. Like the builder, the bench keeps only mazes the bot drives without touching a wall. Nobody's times are published for these, so they are scored against the game's bot.
| Tier | The mazes | Against (base mazes, average time) | Test pilot's score (every maze) |
|---|---|---|---|
| Daily mazes | 6 daily mazes people raced, with the median person's best verified time | median person 18.3 s | 48 |
| Workbench | Plain floors, 6 x 6; level 1-1 is the authored 5 x 5 | bot 16.4 s | 51 |
| Cold Storage | Ice and mud, 6 x 6 | bot 13.8 s | 53 |
| Fast Lane | Boost pads, with ice and mud, 6 x 6 | bot 15.1 s | 50 |
| Night Shift | The big 8 x 8 mazes, any floor | bot 27.0 s | 51 |
Every model drives the same mazes in the same order from set v1. The order within a campaign tier is a fixed shuffle, so the first mazes are a fair sample of the tier and not its easiest. A model ranks once it has driven every daily maze and the first 2 of each campaign tier. After that the bench keeps driving further campaign mazes, up to 40 per tier, so n grows and the campaign interval narrows.
A reply
Play is in paused time. Each request is fresh and holds the whole state. First how the controls and floors behave, from the pinned physics. Then the maze as a text map with coordinates: walls, the dock, ice, mud, and boost pads with their arrows. Then the robot: its position and cell, heading, speed, wheel angle, the floor under it, drift charge and boost, the clock, the route distance from its cell to the dock, and what its last reply did. The route itself is not given. Reading the maze is part of the task.
The model answers with control segments, each a duration of 0.05 to 0.5 seconds with throttle, steer, reverse, drift and brake, together up to 1 second of driving. The game runs exactly those controls through its own physics, at 120 steps a second, and the next request goes out. The clock never runs while the model thinks. A value out of range is moved into it, a segment past the second is cut, and a line that is not a segment is skipped. Each is counted and reported back. The one retry is for a reply with no readable segment. A second one coasts for half a second.
A run ends at the dock, when the game clock reaches 3 times the bot's time on that maze, or after 1.5 replies per second of that limit. Only a run that reached the dock has a time.
Score
A daily maze the model reached scores the median person's time over the model's, so matching the median person is 1 and taking twice as long is 0.5. A maze it did not reach scores 0. The score is the average over the daily mazes, times 100. Campaign is the same against the game's bot, averaged so each campaign tier counts the same. Each 95% interval treats a run as reaching the dock (an Agresti-Coull-adjusted rate) times a pace (the spread of the runs that did).
The reference players
The median person. 100 on the daily mazes by definition. The game's bot. The AI racer that drives the daily board, searched the way the game's level builder sets each level's bot time: every reference route in each of its 192 driving styles, the fastest finish. On the campaign levels it reproduces the times the game lists. It steers 120 times a second and is handed the route. The game's test pilot. The game's own keyboard test driver, also handed the route and steering every step. Random controls. Random segments on the model's own protocol, the floor.
Cost and time
Same as the chess and sudoku benchmarks: Claude models run through Claude Code and OpenAI models through the Codex CLI on a subscription, with every tool switched off, and local models on our own DGX Sparks. Cost is API-equivalent: what the tokens of every request in the run, retries included, would cost at the lab's list price. Local models have no list price, so they show n/a, not zero. Time is the wall-clock wait for the replies, including any queue on our side, so compare it within a route rather than across them. A request that fails on our side (a timeout, an outage, a usage limit) aborts the game, and an aborted game never counts against the model.
What it doesn't show
Paused time removes reaction speed on purpose. A model sees the robot once a second at most and drives the second in between blind, while people react all the time and the bot and the test pilot steer every step. A model that thinks for minutes per reply pays nothing for it in game time. The maze is text, so reading a picture is not tested. Real-time play from pixels is what the real-time benchmark measures.