Gemini 4 Argon 19 项基准拿下 13 项，但 agentic 编码垫底

aipulsedaily @aipulseda1ly原推发布于 2026.10.01

在 X 上打开原推

Gemini 4 Argon just dropped and it tops 13 of 19 benchmarks in Google's own table. $2 in and $10 out intro pricing, and a 1M token output limit, up from 64K. But it comes in dead last on Terminal-Bench 4.0 and FrontierSWE, the agentic coding tests, behind GPT-6 Astra, Fable 5.1 and Opus 5.5. We all seen this before where they destroy a benchmark but it doesnt perform in agentic or real world tasks. Its not out yet. Trusted cyber defenders get it first, then paid API and Google AI Ultra. Well see what happens on my NYC test.