The same game, $0.28 and $0.82: LongCat-2.5's low cost isn't only about the model
With the same prompt, both models produced a playable Angry Birds-style browser game. LongCat-2.5's bill was $0.28, and Qwen3.8-Max's was $0.82. Behind the price gap are cache hits, activated parameters, and serving methods, and the comments also brought retry counts and finished-task cost into the discussion.
The same prompt: Build a playable Angry Birds-style browser game. Both models delivered playable browser games; LongCat-2.5's bill was $0.28, Qwen3.8-Max's was $0.82, nearly a 3x difference. The blogger ran the test in his own harness, not an official benchmark.
LongCat-2.5 comes from Meituan's LongCat. According to public information, it has 1.6T total parameters, about 48B activated per token, native multimodal input, a training sequence of up to 1M tokens, and is aimed at long-horizon tasks. In this game task, the blogger says it completed 3 levels, 3 types of birds, destructible wood, stone, and glass blocks, plus sound effects and a star rating. The Qwen3.8-Max version is also playable, but the original post did not provide a feature-by-feature comparison on the same dimensions.
Cost details are more worth looking at. LongCat-2.5 ran 53 agent turns this time, and about 4.7M of the 4.8M input tokens came from cache. At its pricing, cached input is $0.006 per 1M tokens, uncached input is $0.30, and output is $1.20. Roughly, cached input was about 2.8 cents, uncached input about 3 cents, and the bulk of the rest was output. The blogger also mentioned that it wrote its own physics tests during the process.
The comments didn't focus only on the 3x price gap. Some asked whether the Qwen run had cache hits, and some wondered whether the cheaper model required more retries and fixes. That reminder is very practical: a low per-token price does not mean a cheap task; finished-task cost, retry count, and human intervention count need to be looked at together. Another commenter's calculation was more direct: if LongCat reliably delivers once, running it three times at $0.28 and keeping the best version would total about $0.84, close to a single Qwen run at $0.82. The premise for this conclusion is that at least one of the three runs is deliverable and the quality does not collapse.
Others added that LongCat-2.5 activates only about 48B parameters per token, and Meituan's blog mentions w4a4 on the serving side; fewer activated weights and lower numerical precision accumulate into a cost advantage in full application builds. Someone also pointed out that the 4.7M/4.8M cache hit rate is itself part of the story: if the harness frequently rewrites early turns, the cache gets invalidated.
So this test is more like a cost sample, not an overall model ranking. It shows that in long-horizon agent builds, cache pricing, activated parameters, and serving quantization all directly affect the bill. For developers, LongCat-2.5 is worth putting into your own harness and running the same prompt; first record completion time, retry count, cache hits, and final playability, then talk about cheapness. The demo entry is at LongCat platform.
发布时间: 2026-11-04 05:41