China Securities has released a research report noting that since July this year, prices for open-source and closed-source models have converged, with a decline in overall token costs, which is favorable for the commercialization of AI applications. As Rubin enters mass production, it is expected to further push the limits of large model training and inference capabilities.

Higher cluster scale, memory bandwidth, and interconnect efficiency are driving continuous scaling in model training, while accelerating the evolution of capabilities such as long-context processing, Mixture of Experts (MoE), and reinforcement learning. The potential for further advancements in model intelligence remains promising.

Despite concerns over Anthropic's annualized revenue run rate (ARR) slowdown, Rubin's shipments are anticipated to drive continued model scaling. The discussion sparked by the Ox Alpha model also keeps the outlook positive for open-source models and AI applications. The overall token cost continues its downward trajectory; on August 22, OpenAI announced a limited-time 20% price cut on input tokens and a 33% reduction on output tokens for GPT-5.6 Sol developer pricing over three months. Since July, the price gap between open-source and closed-source models has narrowed, and lower token prices are boosting AI application adoption.

According to Bloomberg, Anthropic recently informed investors that its July ARR was approximately $65 billion, while Yipit's estimate was around $80 billion, 20% higher than the former. China Securities believes several factors may explain the higher Yipit figure: 1) Yipit's sample includes credit card data, which may capture prepaid amounts; 2) The core sample of 1,300 companies consists of early AI adopters with generally higher spending levels. However, the clear trend now is that Yipit previously overestimated Anthropic while underestimating OpenAI. China Securities maintains that the rapid overall growth trajectory of global large model ARR remains unchanged.

China Securities also believes that as Rubin enters mass production, it is expected to further expand the ceiling for large model training and inference. Looking back at the launch of the Blackwell architecture, next-generation models like GPT-5.5, GPT-5.3-Codex, and Grok 4.5 have begun explicitly using GB200/GB300 for training, with capability improvements primarily seen in complex coding, specialized professional tasks, and long-horizon agentic tasks. Higher cluster scale, memory bandwidth, and interconnect efficiency are driving continuous scaling in model training and accelerating the evolution of long-context, MoE, and reinforcement learning capabilities. Rubin's generational advancements in HBM4, NVLink 6, and token throughput per watt are expected to support larger parameter counts, longer chain-of-thought reasoning, and more complex agent training, leaving ample room for future improvements in model intelligence.

Competition among models remains intense. GPT Sol series price cuts and the Ox Alpha model have sparked discussions. China Securities observed that last week OpenAI reduced the short-context API price for GPT-5.6 Sol from $5 input and $30 output per million tokens to $4 and $20, respectively, representing cuts of 20% and 33%, with the discount lasting at least until November 21. Combined with the earlier 80% price cut for Luna and 20% for Terra, competition among leading models is shifting from pure capability comparisons to pricing, speed, and agent usability. Meanwhile, the anonymous model Ox Alpha appeared on OpenRouter, offering roughly 1.05 million tokens of free context with support for text, images, video, and tool calling, targeting long-horizon coding and agentic tasks. Within days of its launch, cumulative usage approached 10 billion tokens, and its coding performance and true origin have quickly become a topic of debate among developers.

Regarding commercialization, China Securities believes the comprehensive token cost continues to decline, with the GPT-5.6 Sol developer price temporarily reduced for three months. According to OpenAI's official X account, while continuously expanding capability boundaries and improving efficiency, the company will cut GPT-5.6 Sol API and quota prices by over 20% in the next three months, with new input/output prices at $4 and $20 per million tokens, representing reductions of 20% and 33.3%, respectively. According to SiliconData, the overall LLM token index has remained stable with a slight decline, particularly showing a notable downward trend since June this year; the closed-source model index has been falling since July, with a steep trajectory, while the open-source model index has climbed from its bottom, trending upward but with relatively modest gains so far.

On token usage, last week (August 17-23) OpenRouter recorded an estimated 90.2 trillion tokens, up 19.6% week-over-week, marking a new weekly high. The accelerated growth is likely due to the launch of the new model Ox Alpha, which is currently free, along with significant discounts on MiMo-V2.5 and Nemotron 3 Ultra, quickly driving token demand. DeepSeek V4 Flash's official version maintained its top ranking in token usage last week. The anonymous model Ox Alpha, which launched on OpenRouter on August 21, ranked fourth for the week; based on technical fingerprints, it is most likely a domestic (Chinese) model. Among the top five last week, four were domestic open-source models: DeepSeek V4 Flash 0731, MiMo-V2.5, Hy 3, and DeepSeek V4 Flash 0423. In OpenRouter's weekly token usage, open-source models accounted for 77.8%, with Chinese open-source models representing 70.0% (based on the top 9 models by usage; note the previous week's statistics window was not fully covered at the time of data capture, so these are incomplete figures).

The share of open-source models has declined over the past month, mainly due to the rapid token usage growth of the closed-source model GPT 5.6 Luna following its significant price cut. Part of last week's decline stems from the fact that the newly launched Ox Alpha has not been officially categorized as open-source or closed-source, despite ranking fourth in token usage for the week. In terms of weekly token market share (based on ZenMux data), DeepSeek's share fell to 37.5% but remained in first place. According to ZenMux (a large model aggregation platform), the top five by token market share last week were DeepSeek, Anthropic, OpenAI, Google, and Zhipu, with DeepSeek and Zhipu together accounting for 43.4%. Xiaohongshu's Dots Studio also made the list; on August 14, Xiaohongshu officially released and open-sourced Dots3-Note Preview, which is currently available for free on the ZenMux platform.