四家实验室 72 小时内齐发旗舰模型对比
盘点 GPT-6 Astra 与 Claude、Meta、Google 新模型的发布与定价，重点提 Astra 内置 computer use、Terminal-Bench 4.0 得 58.18%、token 省 45%，并提醒别盲目升级。
Four AI labs just dropped flagship models in 72 hours. The AI model race is getting ridiculous. OpenAI’s GPT-6 Astra launched Sept. 3 alongside major releases from Anthropic, Meta and Google. What stands out: • Astra: built-in computer use • 58.18% on Terminal-Bench 4.0 • 45% fewer tokens than Claude Fable 5.1 on that benchmark • First OpenAI model rated “Critical” for cybersecurity capability Meanwhile: • Claude Fable 5.1 — $10/$50 per 1M tokens • GPT-6 Astra — $10/$50 • Muse Spark 1.3 — $1.25/$4.25 • Gemini 3.8 Flash — $0.75/$3.75 The bigger story isn’t the benchmark scores. It’s that AI models are becoming agents. They don’t just generate text anymore. They can increasingly operate browsers, apps, files and development environments. If you build with AI, don’t blindly upgrade models. Test your existing workflows first. A new model can improve performance — or quietly break the way your agent behaves.