How we measure
AI vs Games is an independent scoreboard for AI models and agents playing video games. The internet's scoreboard for AI playing video games. Every result backed by a source.
The one rule: one game at a time
We never combine different games into a single “smartest AI” ranking. A chess Elo, a Pokémon badge count and a Minecraft win rate measure different things under different rules. Every table on this site covers one game or one benchmark, using that benchmark's own metric and method.
What every result shows
- Model: the exact version as the source names it (for example “GPT-6 Astra”, plus effort level if given). Source labels are mapped to one canonical name, and effort or budget suffixes (max, high, thinking…) go into the run setup instead of the model name.
- Game, and the game version when known.
- Run setup: harness, inputs, tools, human assistance and budget. When a source doesn't say, we write “not disclosed”. Unknown values stay empty, never zero.
- Achievement, with the source's own metric and units.
- Evidence: one or more source links. A record is never published without at least one.
- Verification label: Verified checked against primary evidence (game telemetry, turn-by-turn logs or save states, an official leaderboard or replays) · Creator-reported stated by the run's creator or authors (a post, paper, self-reported list or the AI's own notes) and not independently checked · Awaiting verification unconfirmed, conflicting or missing key facts.
- Dates: Last verified (when the value was last confirmed against primary evidence), Last checked (when we last looked), plus published (the evidence's date) and discovered (when we first recorded it). If nothing changed, we say so. We never invent news.
Agent Adventures vs Official Benchmarks
Agent Adventures are open-ended runs of whole games, usually streamed. Official Benchmarks are fixed tests run by an organizer. We keep each benchmark's metric, method and version, store immutable snapshots, and flag any change in method. VideoGameBench's real-time and Lite modes stay separate.
Highlights
The homepage “Biggest moment” only uses verified or creator-reported records. We say “surprise” rather than “upset” unless someone documented a prediction before the match. Every highlight explains why it was picked and links its evidence. “Then vs Now” appears only for the same benchmark or run series, the same metric and units, and a compatible method, with any method changes flagged.
How data is collected
Collectors use official APIs and permitted public pages, with timeouts, retries with backoff, ETag caching for GitHub, and automatic shutoff after repeated errors. We never delete a record because a source went away. We don't scrape X, Twitch or YouTube indiscriminately. Kaggle Game Arena data is a reviewed snapshot until bulk access is confirmed. New records and milestones go to a human review queue before they are published. Only low-risk updates, such as “last checked” dates, apply automatically.
Sponsor independence policy
Sponsors can buy visibility: clearly labelled placements in the newsletter or on the site. Sponsors can never buy or influence rankings, verification labels, record verdicts, corrections or editorial highlights. Sponsored items are always marked “Sponsored”. If a sponsor's own model, agent or run appears in our data, it follows exactly the same rules as everyone else, and we disclose the relationship next to it.
Corrections
Every record has a “Submit a correction” link. We fix errors, note the correction in the change log, and never quietly rewrite history.
Data & reuse
Everything is downloadable on the data page (JSON, versioned, CC BY 4.0). Please credit AI vs Games and the original sources.
Data last refreshed: Oct 18, 2026 2:05 PM ET.