Gemini 4 Argon 无预告发布：18 项基准中 13 项领先
帖中称 Argon 支持 100 万输入输出 token、Harvey Legal 19.6% 远超 Opus 5.5 与 Astra，并已在 Google 内部重写 libgav1 SIMD 代码、释放数百 TiB 内存。
GEMINI 4 ARGON JUST BEAT CLAUDE BY 5X AND GPT BY 4X Google shipped Gemini 4 Argon with no warning. It leads on 13 of 18 major benchmarks and already runs inside Google itself. Runs on Google + 1M input tokens + 1M output tokens. One prompt returns a reply the size of 8 Harry Potter books. Devs, this hits your workflow on day one. The numbers are insane: - Harvey Legal (legal agents): Argon 19.6% vs Claude Opus 5.5 at 3.8% and GPT-6 Astra at 5.4% - DeepSWE v1.1 (long-horizon coding): 77.9%, #1 - CWE-bench (vulnerability discovery): 68%, tied for top - Promo pricing: $2 / $10 per 1M tokens - Standard: $4 / $20 (same as Opus 5.5, 2.5x cheaper than GPT-6 Astra) - Cache hit: 95% off. Repeat calls cost $0.10 per 1M tokens The opportunity is even wilder: Google runs Argon internally and the numbers are brutal. It rewrote 32,000 lines of SIMD code inside libgav1, the new decoder ships 2.7x faster. It freed 300 to 1,000 TiB of memory across Google's data centers. Wiz has already put it on live production security duty. One person + Argon + $10 = a SWAT team of senior engineers, while competitors wait for the public rollout. The mass release is coming.
这条是引用着另一条写的：x.com/google/status/2105388143902175529