China Securities Co., Ltd.: Rubin has entered mass production, which is expected to further unlock the training and inference limits of large models.

China Securities Co., Ltd. has published a research report stating that since July this year, prices for both open-source and closed-source models have converged, and the overall token prices have declined, favoring the implementation of AI applications. With Rubin entering large-scale production, it is expected to further expand the limits of large model training and inference. Greater cluster scale, memory bandwidth, and interconnect efficiency are driving continuous scaling of model training, accelerating the evolution of capabilities such as long context, MoE, and reinforcement learning, and the potential intelligence ceiling of the models remains promising. Concerns regarding the slowdown of Anthropic's ARR may be alleviated by Rubin's shipments, which are expected to drive further scaling of models. The Ox Alpha model has sparked discussions. We remain optimistic about open-source models and AI applications. The overall token cost continues a downward trend. On August 22, OpenAI announced a temporary price reduction for developers of GPT-5.6 Sol for three months, with input and output prices reduced by 20% and 33%, respectively. Since July, the prices of both open-source and closed-source models have converged, with overall token prices declining, favoring the deployment of AI applications. According to Bloomberg, Anthropic recently informed investors that its ARR in July was about $65 billion, while Yipits estimate is around $80 billion, which is 20% higher than the former. China Securities Co., Ltd. believes that several reasons may account for Yipits higher estimate: 1) The Yipit sample includes credit cards, which may involve some credit prepayments; 2) The core sample of 1,300 companies are early adopters of AI, which skews average levels higher. However, the overall trend is clear: Yipit had previously overestimated Company A while underestimating OpenAI. China Securities Co., Ltd. believes that the global trend of rapidly growing ARR for large models remains unchanged. With Rubin entering large-scale production, it is expected to further expand the limits of large model training and inference. Reflecting on the launch of the Blackwell architecture, new generation models such as GPT-5.5, GPT-5.3-Codex, and Grok 4.5 have begun to clearly utilize GB200/GB300 for training. The performance improvements of these models are mainly evident in complex programming tasks, specialized knowledge jobs, and long-cycle intelligent agent tasks. Higher cluster scale, memory bandwidth, and interconnect efficiency are driving continuous scaling in model training, accelerating the evolution of capabilities like long context, MoE, and reinforcement learning. Rubin has achieved generational upgrades in areas like HBM4, NVLink 6, and per watt token throughput, supporting larger parameters, longer thought chains, and more complex intelligent agent training, making the potential for intelligent model capabilities very promising. Competition among models remains fierce, with price reductions for the GPT Sol series and discussions sparked by the Ox Alpha model. China Securities Co., Ltd. observed that last week OpenAI reduced the prices for the GPT-5.6 Sol short-context API from $5 for input and $30 for output per million tokens to $4 and $20, representing declines of 20% and 33%, respectively, with discounts at least lasting until November 21. Combined with the previous 80% price cut for Luna and a 20% reduction for Terra, the competition among top models is shifting from a pure capabilities contest to pricing, speed, and agent availability. Meanwhile, OpenRouter launched the anonymous model Ox Alpha providing around 1.05 million tokens of context for free, supporting text, images, videos, and tool calls, targeting long-term programming and intelligent agent tasks. In just a few days after its launch, the cumulative call volume approached 10 billion tokens, and its programming performance and real-world source rapidly prompted discussions among developers. Commercialization: The developer price for GPT-5.6 Sol has been temporarily reduced for three months. China Securities Co., Ltd. believes that the overall token cost remains on a downward trend, with a temporary price reduction for developers of GPT-5.6 Sol for three months. According to OpenAI's official X account, while continuously expanding capabilities and improving efficiency, they will reduce the API and quota prices for GPT-5.6 Sol by more than 20% in the next three months, with the latest input/output price set at $4/$20 per million tokens, reflecting declines of 20% and 33.3%, respectively. According to SiliconData, the overall LLM Token index is stable with a downward trend, particularly evident since June of this year; the closed-source model index has declined steeply since July; and the open-source model index has risen from the bottom but currently has relatively low growth. Token: Last week, OpenRouter anticipated that token call volume would grow by 19.6% week-over-week. In the week of August 17-23, the projected token calls reached 902 trillion, reflecting a 19.6% increase, and setting a new weekly high. This acceleration in growth is expected to be driven by the launch of the new model Ox Alpha, which is currently free, along with significant discounts for MiMo-V2.5 and Nemotron 3 Ultra, rapidly releasing token demand. Last week, the official version of DeepSeek V4 Flash maintained the top rank in token calls, while the anonymous model Ox Alpha, launched on August 21, ranked fourth that week, likely being a domestic model based on technical fingerprints; four of the top five ranks last week were domestic open-source models, namely DeepSeek V4 Flash 0731, MiMo-V2.5, Hy 3, and DeepSeek V4 Flash 0423. In the weekly token calls on OpenRouter, open-source models accounted for 77.8%, with Chinese open-source models accounting for 70.0% (this data reflects the top 9 token call volumes with the week not fully covered at the cut-off time and is not a complete statistic). The decline in open-source model proportion over the past month is mainly due to the rapid growth of token calls for the closed-source model GPT 5.6 Luna after its significant price cut; part of last week's decline can be attributed to the recently launched model Ox Alpha not being formally released and not clearly categorized as either open-source or closed-source, which ranked fourth in token calls. Last week, in terms of token market share (ZenMux perspective), DeepSeek's share fell back to 37.5%, still holding the top position. According to ZenMux (a large model aggregation platform), last weeks top five in token market shares were DeepSeek, Anthropic, OpenAI, Google, and Z.AI, with DeepSeek and Z.AI combined accounting for 43.4%; Dots Studio, a large model under Xiaohongshu, made the list, with the official release of Dots3-Note Preview on August 14 and is currently available for free on ZenMux platform.