GO
| HSI1 | 24,642.51 | +132.42 | 177.48B |
| HSCEI1 | 8,216.84 | +51.06 | 51.28B |
| Back Zoom + Zoom - Block Traded | |
|
2026-09-28 12:52:41 Citi published a research report stating that the latest frontier models showed improvements in both performance and cost per task, highlighting enhanced model efficiency and inference economics, which may help strengthen the longer-term economic advantages of more efficient models over less efficient open-source weight alternatives. The bank noted that the top proprietary models released this week lifted Artificial Analysis (AA) intelligence frontier scores by 9% while reducing AA task costs by 55% compared with their predecessor models. The accelerated rollout of frontier models continues to widen the gap with open-source models. The bank said open-source model cost per task fell 35% WoW to USD0.8, while the discount versus closed models widened from 40% to 60%. The intelligence lead of proprietary models expanded from 9 points to 12 points. Cache-related operations accounted for 61% of task costs, indicating that the next phase of inference efficiency may come from providing more context at lower marginal costs, benefiting vendor margins and making more agent workloads viable. US-Europe blended token prices fell 25% WoW, but the Silicon Data token spending index rose 3%, reflecting demand elasticity at lower price levels. Proprietary model releases reached 20 last week, about 67% higher than the previous five-week peak of 12. The bank also said XIAOMI-W (01810.HK)'s MiMo-V2.6-Pro demonstrated that achieving capability improvements through agentic RL has become inexpensive. A roughly USD2.6 million RL run lasting six days on an existing base model produced a top-tier open-source weight model. Frontier laboratories are expected to increasingly replicate similar processes across their own model series. Consumer-side agent applications are gaining rapid adoption, with Grok Bot recording 935,000 iOS downloads within 44 days and Muse surpassing 2 million downloads within 15 days. However, computing power constraints and the zero-sum reality of token economics may force vendors to balance consumer adoption against higher-value enterprise and wholesale revenue streams. Personalized, always-on autonomous agents may boost short-term CPU and network demand, while long-term GPU demand may also rise as agents accumulate context, become more personalized, and undertake longer tasks. OpenAI DevDay on Tuesday (29 Sep) is expected to provide more information on personalized agent architecture and the acceleration of its enterprise business. (da/ad)~ AASTOCKS Financial News Website: www.aastocks.com This article was automatically translated by AI, the Chinese version should be considered the authoritative version. AASTOCKS.com Limited does not guarantee its accuracy or completeness and accepts no liability for any damages or losses arising from the use of this translation. | |