Playgama launches game-based ranking for AI coding tools

Playgama has launched the AI Game Index, a daily ranking of AI coding agents and language models based on how players respond to browser games published in its Sandbox. The index, available on Playgama’s website, covered more than 5,400 games and 1.4 million launches as of Oct. 29.

Rather than test models against prescribed programming tasks, the index measures the time players spend in games and the likes they give them. Playgama says the rankings use games with at least 20 plays, calculate median results across an agent’s or model’s eligible games, and count only sessions lasting at least 30 seconds. A model needs 15 games that meet that play threshold before receiving a rating.

Developers declare the agent and model used when submitting a game, but those labels are self-reported and unverified. The form records a model’s name, not its version, an important limitation when releases can change rapidly. TechBullion reported that Playgama founder Dmitry Kachmar said OpenAI Sol had replaced Astra within two weeks, arguing for the daily updates.

Claude Code was used for 50% of games in the database. Codex accounted for 19%, followed by the Claude application at 11%. At the model level, Claude Opus appeared in 56% of projects, compared with 10% for OpenAI Sol and 9.1% for Claude Sonnet.

Prevalence did not necessarily translate into the strongest engagement results. Among development tools, games made with OpenCode received 64 likes per 1,000 plays, ahead of Codex at 55 and ChatGPT at 50. At the model level, DeepSeek games had the highest cited likes rate, at 72 per 1,000 plays, while OpenAI Astra had 65 and Claude Opus 38.

The play-time results also differed from the likes rankings. Games attributed to Gemini Pro had an average session of four minutes and 30 seconds, followed by OpenAI Astra at four minutes and 20 seconds. Playgama describes the project as a behavioural evaluation of published games, unlike coding benchmarks such as SWE-bench or answer-voting arenas such as LMArena.

The results apply only to Playgama Sandbox, where developers can publish HTML5 games, use the Playgama Bridge SDK and collect player analytics. Genre, marketing and a project’s overall quality can all affect engagement independently of an AI tool’s capabilities. Kachmar characterised the index as “a thermometer, not a judge.”

The index is in beta. Playgama says it plans to add agents, models and measures as more games are released.