Show HN: Pac-Bench – How well can models one-shot a Pac-Man game?

Hacker News Top Tools
pac-man benchmark ai-evaluation one-shot leaderboard hacker-news

Summary

PacBench is a benchmark that evaluates AI models' ability to one-shot a Pac-Man game, providing source code, model labels, and performance stats from various runs.

Benchmarks how well Harness+models can create a Pac-Man game from a single prompt:<p>“Create a Pac-Man game in a single HTML page”<p>Each model gets one shot — no follow-up prompts or fixes.
Original Article
View Cached Full Text

Cached at: 09/29/26, 04:51 AM

# PacBench Source: [https://jonclegg.github.io/pacman-bakeoff/](https://jonclegg.github.io/pacman-bakeoff/) Model labels are the short name: no provider prefix, date, or effort suffix\. The full requested and actual ids stay under Run data\. When two cards share a label, a card that was asked for a different model says so under the title\. Stats are wall time and token counts from harness transcripts where available\. Phase 2 entries use Claude Code pointed at OpenRouter\. Phase 3 entries use Antigravity with Gemini\. The Grok Bot entry was written in\-chat from the same short prompt \(no peeking\)\. The gpt\-6 Codex cards are API\-key reruns\. Their cost is OpenAI's published Standard rate card applied to captured response usage\. Cursor Cloud costs, where filled, are the dashboard chargedCents sum\. HTML size is the entry file as served\. A dash means the run did not record that number\. [github\.com/jonclegg/pacman\-bakeoff](https://github.com/jonclegg/pacman-bakeoff)

Similar Articles

ProgramBench (5 minute read)

TLDR AI

ProgramBench is a new benchmark that evaluates AI agents' ability to reconstruct complete software projects from compiled binaries and documentation without access to source code or decompilation tools.

Introducing BenchBench (5 minute read)

TLDR AI

Introduces BenchBench, a benchmark that tests AI models' ability to create effective benchmarks for other models, with GPT 5.2 being the only successful winner so far while frontier models like GPT 5.5 and Opus 4.6 struggled.

PaperBench: Evaluating AI’s Ability to Replicate AI Research

OpenAI Blog

OpenAI introduces PaperBench, a benchmark evaluating AI agents' ability to replicate state-of-the-art AI research by replicating 20 ICML 2024 papers with 8,316 gradable tasks. The best-performing model (Claude 3.5 Sonnet) achieves only 21% replication score, below human PhD-level performance, highlighting current limitations in autonomous research capabilities.