9月1日至3日，Anthropic、OpenAI、Meta、谷歌四家美国AI龙头企业密集发布新款大模型，核心能力提升集中在Agent方向，从让AI回答问题转向让AI完成工作1。其中，Anthropic推出的Fable 5.1能连续数小时处理编程、研究和知识工作1；谷歌发布Gemini 3.8 Flash，在DeepSWE v1.1基准测试中得分达73.7%，相比前代3.7 Flash版本提升约8个百分点1；Meta推出Muse Spark 1.3，通过降低工具调用次数约20%和Token使用量约25%来优化编码任务表现1；OpenAI发布GPT-6 Astra，在ARC-AGI-3测试中的成绩从GPT-5.6 Sol的7.8%飙升至99.9%1。
对于这波新模型浪潮的意义，观点存在分歧。OpenAI总裁格雷格·布洛克曼宣称"AGI时代已来"1。然而专业机构指出，GPT-6 Astra的高分存在明显局限性1：其使用Provider Adapter框架后才达到99.9%的最高得分，单次运行成本约1.88万美元，高分严重依赖豪华外挂和昂贵算力1。这些成绩无法代表真实世界的复杂性1。
Between September 1 and 3, 2026, four leading American AI companies unveiled new large language models in rapid succession, signaling a strategic shift toward agent-based capabilities.1 Anthropic introduced Fable 5.1 on September 1, capable of sustaining work on programming, research, and knowledge tasks for consecutive hours.1 Google released Gemini 3.8 Flash on September 2, achieving a 73.7% score on DeepSWE v1.1, representing approximately an 8 percentage point improvement over its predecessor Gemini 3.7 Flash.1 Meta launched Muse Spark 1.3 the same day, reducing tool invocations by roughly 20% and token consumption by about 25% on coding tasks.1 OpenAI completed the wave of announcements on September 3 with GPT-6 Astra, which surged from GPT-5.6 Sol's 7.8% performance to 99.9% on the ARC-AGI-3 benchmark.1
These developments reflect a broader industry transition from question-answering systems toward AI capable of executing complex work autonomously.1 OpenAI President Greg Brockman declared that "the AGI era has arrived," reflecting the industry's confidence in these advances.1 However, the remarkable performance gains come with significant caveats. Astra's top score of 99.9% was achieved using the Provider Adapter framework at a cost of approximately $18,800 per evaluation, and these results rely heavily on expensive auxiliary systems and substantial computational resources.1 Professional institutions have pointed out that these achievements contain notable limitations and cannot adequately represent real-world complexity.1
评论
还没有评论，欢迎留下第一条。