Rank 30 games by the score of each game's best manually reviewed screenshot, not by source-quality score. Review composition, scene detail, visual coherence, and readable gameplay. Reward polished stylized art as well as realism; discount title cards, menus, concept art, and frames that do not show play. Treat this as a visual impression, not a runtime playtest or proof of AAA production quality. Select a thumbnail or title to inspect every image's score, rating method, and reason on the game's Markdown page. Secondary images use relative frame adjustments, not independent manual reviews.

Turbo Kart Rally · 📸 9.8/10	Neural Sight · 📸 9.7/10	Kart Royale · 📸 9.5/10
Nerd of Duty · 📸 9.4/10	VECTOR RUSH · 📸 9.3/10	Turbo Kart Grand Prix · 📸 9.2/10
Iridium Reach · 📸 9.0/10	Silent Meridian · 📸 9.0/10	World of ClaudeCraft · 📸 9.0/10
KINDLE · 📸 8.9/10	SUNBREAK — Downhill Club · 📸 8.9/10	Fishslop · 📸 8.8/10
OVERRUN: Dockyard Nine · 📸 8.8/10	Fall Line · 📸 8.7/10	Little Flock · 小羊慢慢 · 📸 8.7/10
The Hourglass City · 📸 8.7/10	Turbo Kart Rush · 📸 8.7/10	Turbo Kart Rush · 📸 8.7/10
Rally Navigator · 📸 8.6/10	Rogue Squadron — The Battle of Yavin · 📸 8.6/10	Clawd Pop 3D · 📸 8.5/10
The Black Sedan · 📸 8.5/10	LUMA — The Garden of Light · 📸 8.3/10	Barista Shift · 📸 8.2/10
Nova Lancer · 📸 8.2/10	Operation Ironhold · 📸 8.2/10	Saber / Descent · 📸 8.2/10
Tater's Flight Sim · 📸 8.2/10	BuildYourTown · 📸 8.1/10	DRIFTLANDS · 📸 8.0/10

Select a model to browse its ranked games and screenshots. Count each multi-model game on every matching model page; do not sum these counts for a collection total. Keep broad and unspecified model labels separate from versioned labels.

Rank the highest-rated repositories verified in this curation run on 2026-09-26.

Repository creation on 2026-09-25–2026-09-26 is not proof of game publication. Inspect the source and model-evidence links before using an entry.

Rank the highest-rated games with publication or qualifying gameplay evidence from 2026-09-20 through 2026-09-26.

Rank the highest-rated games with publication or qualifying gameplay evidence from 2026-09-01 through 2026-09-26.

Start here. These projects have the strongest combined evidence, scope, and source quality. Ratings do not replace evidence grades.

This is a curated index of playable game units with public GitHub source and evidence that connects them to GPT-6 Astra, Claude Opus, or Claude Fable. “Curated” does not mean every attribution has the same strength: the per-entry evidence grade states whether the model claim is direct, creator-reported, repository-level, or inferred.

The source of truth is games.json. It was last verified on 2026-09-26.

Browse the awesome-list index for verified game catalogs with their counted entry totals.

Browse latest game additions by date to compare each source repository's creation time with the date this collection first recorded it.

Browse AI game generators and engines for tools that build or edit playable games from prompts.

Browse Claude Opus 5.5 release-week games for direct source and model-evidence links.

Browse publication-date rankings for daily, weekly, and monthly reports. These reports exclude games without a reliable publication or qualifying evidence date. The audit keeps repository creation dates separate from publication dates.

The dataset stores source-derived reconstruction prompts without presenting them as original transcripts. Newly expanded Opus 5.5 entries link to their per-game prompt notes.

The list contains 794 game units from 521 qualifying repositories. The dataset also retains 80 related or excluded records for audit history. Each row uses one of these model-evidence grades.

• ⚠️ Bad games — 28 low-quality game units excluded from the curated library
• 📦 Other — 52 related, derivative, forked, or non-game records

• Select a game title to open its local per-game note in games/.
• Select the repository link inside the note to open the canonical GitHub source.
• Select files to inspect linked gameplay source. Select play for a published demo.
• Read the evidence grade before relying on a model-attribution claim. The rating is a curation aid, not a benchmark.

Each compact row shows source-quality rating, model, technology, model-evidence grade, and relevant source links. Rows with usable screenshots also show a separate 📸 screenshot rating. FLOPS estimates remain in the dataset but are omitted here because they are static, low-confidence estimates.

Proof over promises. Repository evidence decides what gets counted.

• Verify the canonical GitHub repository URL. Record its numeric ID when public API metadata is available.
• Inspect the README, entry point, and gameplay source. Confirm input, rules or objectives, and game state.
• Place native-engine games in the dedicated non-browser section.
• Record model evidence as direct, creator-reported, repository-level, directory/topic trail, or inferred. Link every grade to its public evidence URL. Do not present weaker evidence as direct confirmation.
• Treat source inspection as proof that the code exists. Treat it separately from runtime playtesting.
• Do not count catalogs, skills, screenshots, visual-only scenes, empty repositories, unchanged forks, or prompt-only projects.
• Keep source records in games.json. Keep rejected candidates in research/candidates.json. Run the validator and regenerate the README and model pages after each dataset edit.

The 0–10 rating uses source completeness, playable mechanics, scope, tests or deployment, and model-evidence strength. It is a curation estimate, not a review score or performance benchmark. The dataset keeps a FLOPS field as a low-confidence static FP32-work estimate at 60 FPS; it is not measured device performance or model-training compute.

Open the games/ directory or the per-game notes index to read a short source-backed note for every named game unit. Each note links model evidence, gameplay source, and available screenshot assets. Use the placeholder only when the repository and related-source scan found no screenshot.

Open rankings/ for generated top games this week, top games this month, and daily rankings for the last seven dates. Read each row's date basis before treating it as a publication date.

Follow the contribution guide. Provide the canonical repository, model-evidence URL, playable-source path, and evidence notes. Validate the dataset and regenerate the README and model pages before opening a pull request.

Primary GitHub repository evidence is preferred. Creator reports and repository trails are useful discovery evidence but are not equivalent to direct model attribution. A source inspection confirms that code exists; it does not claim that a current live demo was playtested unless the record states it. This project follows the practical principle in Google’s AI-search guidance: publish clear, original, useful, crawlable information instead of special markup or artificial content tricks.

If source-backed AI game history should stay searchable, star the repository.