3 October 2026 · Agentic game · isometric 3D
Digger 3D · GPT-6 Luna
Codex completed a Digger 3D browser game build. The retained session notes report 28/28 simulation self-tests passing and a small interactive smoke check; manual acceptance remained partial.
What was checked
- The Codex session notes record a successful TypeScript check and Vite production build.
- The session notes report 28 of 28 simulation self-tests passing.
- The recorded interactive check covered starting the game, moving, collecting a gem, and firing with cooldown.
- The session notes do not record completing a level or manually checking bags and the bonus mechanic.
Limits of this evidence
The source metrics were selectively transcribed. The original prompt, Codex transcript, private session log path, session identifier, local source paths, and tool arguments/results are excluded. The input count is cumulative across calls and includes repeated conversation context; cached input is a subset. Manual task acceptance was partial. An early self-test left a browser-local high score of 20,005; subsequent test changes avoided modifying the score, but the stored browser value was not cleared.
Configuration
The model, serving engine and agent used for this execution.
| Field | Value | Detail |
|---|---|---|
| Model · cloud_baseline | GPT-6 Luna · cloud_baseline · reasoning xhigh | — |
| Runtime | OpenAI hosted inference Provider-managed service; exact build not exposed | inference engine and exact build |
| Harness | OpenAI Codex | the agent or client that drove the run |
| Hardware | Provider-managed cloud inference; physical host not disclosed | hardware recorded for this execution |
| Workload | Digger 3D · revision 1 | One-shot coding challenge. The prompt file matches the local Qwen Digger 3D run; project scaffolds and harness behavior may differ. |
| Repetitions | 1 | executions recorded for this workload; one means an indicative result, not an estimate |
Agent session counters
Agent-wide usage totals. These include reasoning, tool calls and waiting, and are not model decode speed.
| Metric | Value | Definition |
|---|---|---|
| Task elapsed | 80.4 min | user request received to Codex task completion; includes reasoning, tool calls, and waiting · n=1 |
| End-to-end output rate | 20.32 tok/s | output_tokens divided by task_wall_time; includes reasoning, tool calls, and waiting; not model decode speed · n=1 |
Context and token counters
Token totals as reported by the harness or engine. Cached and cumulative figures are not additive with fresh input.
| Metric | Value | Definition |
|---|---|---|
| Cumulative input | 12,264,781 tokens | cumulative across model calls; repeated conversation context can be counted again · n=1 · Cached input is a subset of this total and is not additive. |
| Cached input | 12,028,672 tokens | cumulative cached input across model calls; included in cumulative_input_tokens · n=1 |
| Output tokens | 98,008 tokens | cumulative model output across calls; includes reasoning output · n=1 |
| of which reasoning output | 44,152 tokens | subset of output_tokens · n=1 |
Request records & provenance
The retained Codex metrics do not include token emission timestamps or a reliable TTFT measurement. The end-to-end output rate must not be compared to local engine decode or prefill rates.
Run ID: 2026-10-03-gpt-6-luna-codex-digger-3d-baseline
One execution, with checks transcribed from the retained record. A completed session is not a quality score. Measurement method · All runs