Wink Pings

Same Prompt, Four Flagship Models: Qwen 4 Max Ran for Nearly 5 Hours and Still Didn't Rank Near the Top

Tony ran Qwen 4 Max, Fable 5.2, Gemini 4 Argon, and GPT-6 Astra with the same Alpine Gold prompt. The slowest, Qwen, took 4 hours 56 minutes and cost about the same as Gemini, but its result was far behind; Fable was the best but nearly 12 times more expensive; Astra was the fastest. When choosing a model, you have to go by the cost to complete the task.

Same Alpine Gold prompt, each of the four models ran once, with no retries. Tony posted the results and the bill together.

ModelCostTime
Qwen 4 Max$12.034 hours 56 minutes 36 seconds
Fable 5.2$209.503 hours 2 minutes 54 seconds
Gemini 4 Argon$11.612 hours 1 minute 35 seconds
GPT-6 Astra$17.8251 minutes 10 seconds

Qwen 4 Max took 4 hours 56 minutes 36 seconds and cost $12.03. Gemini 4 Argon cost $11.61 and took 2 hours 1 minute 35 seconds. The cost difference between them was only $0.42, but the time difference was nearly 3 hours. GPT-6 Astra cost $17.82 and took 51 minutes 10 seconds. Fable 5.2 cost $209.50 and took 3 hours 2 minutes 54 seconds.

Tony's conclusion was direct: Qwen 4 Max's result was far from the frontier. What it produced in nearly 5 hours was not worth the wait. Fable 5.2 made the best airplane, but it cost nearly 12 times as much as Astra and took more than 3 times as long. Argon was the cheapest, and Astra was the fastest.

Someone in the comments asked whether Qwen was overhyped, or whether raw performance matters more than speed and cost. Tony replied: In this round it lost on all three. It took nearly 5 hours, cost about the same as Argon, and the result was not at the frontier. When the quality is there, he is willing to pay more and wait longer; this time it wasn't worth it.

Another reply mentioned that if the results are comparable, paying $6.21 more to save 70 minutes is an easy choice before a deadline. Tony agreed: Astra was 1 hour 10 minutes faster than Argon and only $6.21 more expensive.

Some background. Tony doesn't write code; he uses AI agents to build 3D movies, SaaS tools, and apps. Starting in May 2026, he built his first app, Laeli, in under three months. This test comes from the same set of tasks he ran more than 70 times across 12+ models. His core judgment is: the best model and the right model are rarely the same. Price should be judged by the cost to complete the task, not the token price.

He gave another example: for the same movie, the High setting cost $26 and the Ultra setting cost $682. Ultra was better, but the extra cost and wait weren't worth it. So for everyday use he uses Extra High. The same logic applies to this comparison: Fable 5.2 had the best quality, but at $209.50 a run, it is only suitable for a very small number of special tasks. Qwen 4 Max's problem is simpler: slow is tolerable; slow and bad is not.

When choosing a model, first ask what the task needs. For the fastest usable result, Astra; for the lowest cost and acceptable quality, Argon; for the best result and no concern about the bill, Fable 5.2. Qwen 4 Max gave no reason to choose it in this round.

Original tweet: https://x.com/EnvolDev/status/2122432152340209964

Long article: https://x.com/i/article/2120246086157881344

发布时间: 2026-11-17 17:27