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

Pac-Bench evaluates how well models can one-shot a Pac-Man game, with model labels using short names without provider prefixes. Run data includes full requested and actual IDs, wall time, and token counts from harness transcripts. Phase 2 entries use Claude Code via OpenRouter, while Phase 3 entries use Antigravity with Gemini. The Grok Bot entry was written in-chat, and gpt-6 Codex cards are API-key reruns. Costs are based on OpenAI's published Standard rate card or Cursor Cloud's chargedCents sum, with HTML size also recorded.