China National Radio reported (Tencent News, September 16) that Q3 2026 brought a new high-intensity round of the global large-model contest, with makers clustering releases in September. On September 1, Anthropic launched Claude Fable 5.1 and the shared-architecture Mythos 5.1, emphasizing doubled performance, far lower cost and enterprise data-compliance. On September 3, OpenAI released GPT-6 Astra, positioned as an AGI-level model with autonomous execution. On September 10, DeepSeek V4.1-Flash went live with a new Causal Encoder–Decoder architecture boosting inference efficiency, and on September 12 xAI's Grok 4.7 reached 2.1 trillion parameters (up about 40% from 4.6).
The Scaling Law still holds: OpenAI's GPT-6 Astra—its largest training project ever—used over 100,000 GPUs, showing frontier capability still rests on larger clusters and interconnects. Guosen Securities analyst Bao Enxue said no Scaling-Law bottleneck is yet visible, while Shanghai Jiao Tong University's Tian Feng noted that each additional one-percent gain now costs exponentially more.
The密集 releases highlight a shift from pure capability competition to efficiency and cost optimization as inference-scale deployment accelerates.
Source: China National Radio / Tencent News
