GameFeel Benchmark
GameFeel Benchmark is a community-maintained AI benchmark for developing games with AI. Maintained in the kodasdream/gamefeel-benchmark repository, it reports coding, game balance, creativity and evaluation scores alongside task costs. Its public materials describe tests on 12 unpublished game projects, with three projects each in roguelike, MMORPG, single-player RPG and survival or sandbox genres.[1]
The project publishes aggregate results and visualization code. These allow inspection of the displayed scores and ranking calculations; they are distinct from a release of the underlying game projects, individual task outputs or human ratings.[1][2][3]
Tasks and grading
The maintainer describes seven subcategories across the 12 projects, approximately 60,000 tasks in total. Coding creation, debugging and reverse-evaluation use answer keys checked by Claude Opus 5.5. Gamefeel, ideas, visuals and audio use human ratings. For creative changes, models are asked to propose or implement 30-100 additions and balance adjustments.[1]
Reverse-evaluation asks a model to judge other models' work and commercial-game changelogs. The data distinguishes blind, assisted and released-game conditions. Human scores are averaged and scaled to 100. These are the maintainer's reported procedures, rather than an independently reproduced evaluation.[1]
Reading the results
The published CSV separates the model name, effort setting, billing basis and scores. Its metadata identified the dataset as version 1.5.1, with 52 entries and an update label of November 16, 2026 at 23:47. The label does not specify a timezone.[2][4]
| Field | Interpretation |
|---|---|
| Model and effort | A named model configuration, rather than a model family regardless of settings |
| Task price | A reported US-dollar cost with a separate API, subscription or free billing label |
| Missing value | An unavailable measurement, distinct from a score or price of zero |
| Reverse-evaluation fields | Separate conditions and a combined result |
| Overall score | Available only when all seven component scores are present |
In version 1.5.1, the dashboard calculates the overall result as a weighted mean: the six coding, gamefeel and creativity scores each have weight 1, while combined reverse-evaluation has weight 1.5. The sum is divided by 7.5 and rounded to one decimal place. Individual-category charts can include a model that lacks the other scores.[3]
The Gamefeel Capabilities Index, or GCI, is a separate version-aligned series. The dashboard explains that it maps overlapping models onto the preceding version's scale and excludes unreleased or stealth entries from its progress-over-time chart. It is not the Artificial Analysis Intelligence Index, which appears in a different data field.[2][3]
Early Qwen labels
A November 8 social post circulated an early Qwen comparison.[5] The later public data labels two entries Qwen-4-27B-Early? (qoDuck) and Qwen-4-Flash-Early? (qoHawk). The visualization code explicitly attaches warning markers to them. The November 16 snapshot also includes Qwen4-Max and preview configurations with incomplete overall results.[2][3]
Those rows should retain their uncertainty labels. A benchmark nickname or tentative identification does not establish a model's official name, parameter count, release status or relationship to the released Qwen4-Max. An old social ranking also cannot substitute for a dated comparison using the same benchmark version and effort settings.[2][3][5]
Evidence and limitations
A cost per benchmark task is not an API price per million tokens. The dataset mixes subscription, API and free access, so its cost field should not be read as a universal quote for another workload.[2][3]
The README describes unpublished games as protection against prior exposure. That is a claim about the test material, rather than proof that every tested model is free of contamination. Aggregate CSV values alone also do not disclose the individual judgments needed to reproduce the human-rated results.[1][2]
Research on LLM-as-a-judge has documented position, verbosity and self-enhancement biases. Those findings provide context for evaluating a model-assisted grading process; they do not demonstrate that a particular GameFeel result is biased.[6]
GameFeel is distinct from GameDevBench, an ICML 2026 benchmark built from 333 tutorial-derived game-development tasks. The two projects use different tasks and evaluation procedures, and their numerical scores are not interchangeable.[7]
Continue reading
- Benchmark (AI)In artificial intelligence and machine learning, a benchmark is a specified evaluation used to compare systems under common conditions.Mentioned 7 times above
- Developing Games with AIDeveloping games with AI means using generative AI during game design and production, or incorporating a model into the game itself. These are different engineering tasks.Linked in the introduction
- Artificial AnalysisArtificial Analysis is an independent benchmarking and analytics platform that evaluates artificial intelligence models and API providers across intelligence, speed, price, and latency, and it is best known…Also in AI Benchmarks
- LLM-as-a-judgeLLM-as-a-judge is the practice of using a strong large language model to evaluate the outputs of other models, or of itself, in place of a human annotator.Also in AI Benchmarks
- HumanEvalHumanEval is a benchmark for measuring whether a code-generating large language model can complete short Python functions so that they pass unit tests. It contains 164 hand-written tasks.Also in AI Code Generation
- SWE-bench VerifiedSWE-bench Verified is a 500-problem, human-validated subset of the SWE-bench software engineering benchmark, released on August 13Also in AI Code Generation
More to read
References
- ^1 ^2 ^3 ^4 ^5kodasdream. The GameFeel Benchmark: README and methodology. Accessed November 17, 2026.
- ^1 ^2 ^3 ^4 ^5 ^6 ^7kodasdream. GameFeel benchmark results CSV. Snapshot inspected November 17, 2026.
- ^1 ^2 ^3 ^4 ^5 ^6kodasdream. GameFeel dashboard source, `app.js`. Scoring, warning labels and GCI display logic. Accessed November 17, 2026.
- ^kodasdream. Benchmark metadata, `data/meta.json`. Version 1.5.1, update label November 16, 2026.
- ^1 ^2Layton Gott. Post discussing early Qwen GameFeel results. November 8, 2026.
- ^Lianmin Zheng et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS 2023 Datasets and Benchmarks Track, revised December 24, 2023.
- ^Wayne Chi et al. GameDevBench: Evaluating Agentic Capabilities Through Game Development. Proceedings of the 43rd International Conference on Machine Learning, PMLR 306, pages 19395-19423, 2026.
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
1 revision · v1 · 776 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Cite this page: AI Wiki. "GameFeel Benchmark." aiwiki.ai, updated 17 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/gamefeel_benchmark
