PacBench records this Codex high-reasoning run at 84/100, taking 5 minutes 25 seconds and costing $0.19097; it is a verifiable single-run result, not a measure of the model's overall capabilities.
Tasks this can help assess: Generating a single-file Pac-Man game from one prompt and checking controls, ghost logic, softlocks, the maze, and sound effects.
Tasks this should not be generalized to: Other coding tasks, consistency across repeated runs, or speed and cost rankings across different harnesses.
Applicable model version: gpt-6.1-terra (requested and actual versions match).
Test environment or client: Codex CLI codex exec --ephemeral --json, authenticated with an API key; a blank git cell.
Reasoning effort and parameters: high; web search disabled, with no sandbox or search parameters specified. The CLI logged missing local model metadata and used fallback metadata; the execution parameters still specified gpt-6.1-terra.
The prompt published in the repository asks the model to independently build a single, self-contained HTML game without network access or external dependencies. It must support keyboard controls, a maze and dots, four ghost types, power pellets, scoring, lives, levels, pause, and start/end states. The public site also shows a shorter task summary. The full original prompt is in the repository prompt file.
Opus 5.5 scores the game on the site by playing it automatically for 90 seconds and reviewing the source code and maze. The base weights are controls 20, ghosts 25, softlocks 20, maze 20, and sound 15. The site says later rules adjust for minor control and sound issues and add arcade visual similarity and extra-feature scores, for a maximum of 100. Terra's record notes minor issues with controls and sound; the other three categories pass. The scoring notes give an evaluation date of 2026-09-28, earlier than the 2026-10-26 date shown in Terra's entry. The site does not explain how that date relates to Terra's score.
| Entry | Harness / reasoning effort | Score | Time | Cost | Output HTML |
|---|---|---|---|---|---|
| GPT-6.1 Terra | Codex / high | 84/100 | 5 min 25 sec | $0.19097 (the site card shows $0.19) | 33.1 KB |
| Grok 4.7 | Grok Build / high | 83/100 | 36 minutes | $1.69 | 52.7 KB |
Terra's inspection record says the arrow keys, WASD, pause, and on-screen directional buttons work. There is no cornering control, swipe, or restart key; after a game over, the focused button may cause Space or Enter to restart the game. Holding a direction key repeats the action. The four-ghost strategy, fixed-step movement, and classic 28×31 maze checks pass, and all 244 dots are reachable. The sound effects include eating dots and power pellets, death, bonus rewards, and level completion, but lack the original opening music, siren, and frightened-ghost loop.
Run identifier: gpt-6.1-terra-api; Phase 1; both requested and actual model are gpt-6.1-terra.
Tokens: 272,880 input, including 237,061 cache reads and 35,786 cache writes; 26,871 output, including 8,327 reasoning tokens; 299,751 total. The site notes that reasoning tokens are included in output tokens.
Costing method: Terra cost is calculated using OpenAI Standard short-context rates: $1 per million uncached input tokens, $0.05 for cache reads, $1.25 for cache writes, and $5 for output tokens. The ephemeral run records aggregate usage for the whole run, with no per-response breakdown. The site says no individual request exceeded the long-context threshold and no regional surcharge was included.
Output inspection: The source provides a runnable single-HTML game and itemized scoring notes; this article summarizes the inspection record and did not independently replay the game.
Comparison row: The Grok 4.7 entry also lists high effort, but uses the Grok Build harness; its costing method is not described in that entry. The scores and measurements should not be treated as a controlled model comparison.
This public record makes Terra's input requirements, execution settings, game artifact, single-run usage, and scoring rationale verifiable. It also shows Terra and Grok 4.7 with adjacent scores on the leaderboard, but their harnesses, actual execution, and cost sources differ. A single run cannot establish general differences between the models in speed, cost, or coding ability. The score also depends on the site's game rules and model-based review.
The repository provides the prompt, scoring notes, Playwright automated-play code, scoring metadata, and game HTML. Reproduction requires rerunning with the same harness, model version, reasoning effort, and rates, then repeating the automated play and source/maze review; API usage and cost may vary between runs. The full per-response transcript for the original run is not public.
GPT-6.1 Terra
