I tested LongCat-2.5 and Qwen3.8-Max on the same browser game, and LongCat got it done for $0.28 versus $0.82. That's nearly 3x cheaper for the same task.
I regularly test cheaper models on full app builds, and LongCat-2.5 from @Meituan_LongCat looks like a strong option.
With 1.6T total parameters, around 48B active per token, and native multimodal input, it’s built for long-horizon tasks. Its training includes sequences of up to 1M tokens, helping it stay on track through complex, multi-step projects.
I gave both models the same one-line prompt in my own harness:
"Build a playable Angry Birds-style browser game."
Both games worked. LongCat-2.5’s version included three levels, three bird types, breakable wood, stone, and glass blocks, plus sound effects and star ratings.
LongCat-2.5 completed the task for $0.28 in about 34 minutes, compared with $0.82 for Qwen3.8-Max. It also wrote its own physics tests along the way.
About 4.7M of its 4.8M input tokens came from cache, keeping costs low across 53 agent turns. Cached input is just $0.006 per 1M tokens, while uncached input costs $0.30 and output costs $1.20 per 1M tokens.
Try LongCat-2.5: http://longcat.ai/platform/product?utm_source=e56