GameFeel Benchmark 1.5 lands on Priors

Game development is the ultimate stress test for agentic reasoning because code changes immediately impact fragile mechanics and balance. GameFeel v1.5 shows frontier models still face steep hurdles with creative coherence, with even the top model scoring barely above 50%.

• –Evaluations span 12 unpublished games across four genres, ensuring models cannot rely on memorized GitHub code.
• –Version 1.5 reworks the saturated Reverse-Evaluation section into blind and assisted tracks with 39 in-context examples.
• –Anthropic's Fable 5.2 High leads the 19-model leaderboard at 52.4, edging out Google's Gemini 4 Carbon High at 51.1 and Opus 5.5 XHigh at 46.0.
• –Priors pairs benchmark accuracy with cost-per-task metrics, helping developers weigh whether high-effort reasoning modes justify their subscription and API costs.