Can decision models play games from screenshots?

A benchmark tested six decision models and two LLMs on image- and text-based versions of Tic-tac-toe, Minesweeper, and Flappy Bird, using deterministic game engines as performance baselines. Decision models, designed for fast selection among fixed options, responded in under a second but generally struggled with the multi-step reasoning required: across Tic-tac-toe they achieved 26 draws in 120 games versus 39 in 40 for the LLMs, none cleared a Minesweeper board, and most failed to pass even one Flappy Bird pipe. GPT-6 Astra Responses performed best, drawing every Tic-tac-toe game against perfect play, clearing all Minesweeper boards, and averaging 8.4 pipes in Flappy Bird image mode, while GPT-6 Luna Responses also substantially outperformed its decision-tuned counterpart. Failures among decision models included missing obvious blocks, choosing provable mines, favoring the first text option, and repeating fixed action patterns regardless of game state. Tests comparing GPT-6 Luna with and without reasoning indicated that reasoning substantially improved results, although differences between API routes and image handling may also matter. Across these simple, labeled games, screenshots were not consistently worse than explicit text state descriptions, suggesting that visual perception was not the central limitation; the results emphasize evaluating model choices against both baseline performance and achievable ceilings for specific agent tasks.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.