Argon 写码 + Claude 审校的双模型流水线

alpha404 @alpha404ai原推发布于 2026.10.16

在 X 上打开原推

Google's Gemini 4 Argon + Anthropic's Claude is the craziest AI stack of 2026. Google's Argon scored 19.6 on its hardest eval. Claude Opus 5.5 scored 3.8. Everyone's picking a side. Wrong game. You run both. Argon does the grind. 1M tokens out in one run. It plans a small step, writes the code, runs the tool, and loops until the tests pass. A monitor watches every move and kills the run if it goes off the rails. Claude does the judging. It reads what Argon built, checks the proof, and signs the commit or sends it back. One model that never gets tired. One model that says no. On screen: attempt #2, 425,250 tokens. The first try failed, the error went back into context, and the loop kept going. Nobody touched it. Google's Gemini 4 Argon + Anthropic's Claude is the craziest AI stack of 2026. Google's Argon scored 19.6 on its hardest eval. Claude Opus 5.5 scored 3.8. Everyone's picking a side. Wrong game. You run both. Argon does the grind. 1M tokens out in one run. It plans a small step, writes the code, runs the tool, and loops until the tests pass. A monitor watches every move and kills the run if it goes off the rails. Claude does the judging. It reads what Argon built, checks the proof, and signs the commit or sends it back. One model that never gets tired. One model that says no. On screen: attempt #2, 425,250 tokens. The first try failed, the error went back into context, and the loop kept going. Nobody touched it. The setup: Argon writes and retries. Monitor can halt any step. Tests decide what passes. Claude signs the commit. Argon still returns 404 for most of us. The day it opens, this stack eats a week of dev work overnight. The full illustrated setup is below. Almost nobody has it saved. Save it before the 404 turns into 200. The setup: Argon writes and retries. Monitor can halt any step. Tests decide what passes. Claude signs the commit. Argon still returns 404 for most of us. The day it opens, this stack eats a week of dev work overnight. The full illustrated setup is below. Almost nobody has it saved. Save it before the 404 turns into 200.