When AI plays Pac-Man, response time becomes part of the game

Pac-Man can wait for an answer. The ghosts keep moving.

AI & games | 3-minute read | Sources checked 10 October 2026 (Japan time)

Turn right, or escape to the left? While Pac-Man waits for a decision at a junction, the ghosts can keep getting closer.

That situation is built into jevman, a Pac-Man game published by AI service provider Opper. Models choose which direction to take. The result depends partly on when an answer arrives, as well as which direction it selects. The jevman project

Pac-Man waits while the ghosts move

The game asks about the next junction while Pac-Man is still travelling along a corridor. An answer that arrives early can be ready when he gets there. If he reaches the junction first, he stops to wait. The ghosts continue moving. How movement works

There is a deadline. More than two seconds after a question is sent, a simple set of rules takes over the decision. This backup also handles errors and unusable answers. The decision scheduler

Waiting in the wrong place is a risk of its own. That wait includes the practical work of getting an answer through an API, including communication and service response time, as well as the model's processing.

The maze comes with helpful explanations

What does the model receive? Instead of a screenshot, it gets structured data describing the maze and the characters' positions. The information also includes lives, score and the remaining time during which frightened ghosts can be eaten.

Each available direction comes with an explanation. It describes distances to pellets, how many pellets lie in the next corridor and whether a dangerous ghost is approaching. The game can add a DANGER warning for a ghost in the corridor, or a TRAP warning when a ghost can reach the next junction at least as soon as Pac-Man. The information sent to each model

A separate program is therefore doing useful work before the model chooses. It organises the surroundings and explains the options. The model's task is to use that information to select a direction. This differs from a person watching the screen and controlling every turn.

The highest average can still be a tie

Opper's published results cover 100 games for each of seven models, against ghosts following scripted rules. Each game ends after three lives are lost or a five-minute limit is reached. Benchmark conditions

The reported averages put jev 1.13 at 2,750 points, GPT-6 Luna at 2,568 and Clef Flash at 2,538. But under the organiser's rule for accounting for variation between games, these three share first place among the seven models Opper tested. Being listed first does not establish a clear win. Reported results · How ties are calculated

The same results report average response times of 290 milliseconds for jev and 179 milliseconds for GPT-6 Luna. Score and speed describe different parts of the performance. These figures alone cannot isolate how much of the score difference was caused by response time.

What this particular contest can tell us

Jevman compares performance with a particular maze, a particular description of the situation and a particular deadline. The ranking need not carry over to writing or answering questions at work. The descriptions supplied to the model and the backup rules also contribute to the game being measured.

Even so, Pac-Man makes the cost of waiting easy to understand. A direction can look safe when a question is asked, then become dangerous while the answer is on its way. The next time an AI score catches your attention, it is worth asking what information the model received, and how long it was allowed to respond.

Sources and verification

This article is based on Opper's documentation, published results and source code. We did not play-test the game or run independent API benchmarks. All scores and response times are reported by the organiser. The results file was updated on 9 October 2026; the six models other than Aplomb 1 retained their 7 October measurements. Update record