CITIC Construction Investment: Rubin's entry into large-scale mass production is expected to further open the upper limits of large-scale model training and inference

The Zhitong Finance App learned that CITIC Construction Investment Securities released a research report saying that since July this year, prices for open source and closed source models have converged, and comprehensive token prices have declined, which is beneficial to the implementation of AI applications. As Rubin enters large-scale mass production, it is expected that the upper limit of large-scale model training and inference will be further opened. Higher cluster size, video memory bandwidth, and interconnection efficiency drive continuous scaling of model training, and accelerate the evolution of capabilities such as long context, MoE, and reinforcement learning. The intelligent upper limit of the model is still worth looking forward to.

Anthropic ARR is slowing down, Rubin shipments are expected to drive the model to continue scaling, and the Ox Alpha model sparks discussions

Continue to be optimistic about open source models and AI applications. The comprehensive cost of tokens is still on a downward trend. On August 22, OpenAI announced that GPT-5.6 Sol developer prices will be reduced for a period of three months, and input/output prices will be reduced by 20%/33% respectively. Since July this year, prices for open source and closed source models have converged, and comprehensive token prices have declined, which is beneficial to the implementation of AI applications.

1/According to Bloomberg, Anthropic recently told investors that its July ARR was about 65 billion US dollars, while Yipit's caliber was about 80 billion US dollars, which is 20% higher than the former. CITIC Construction Investment Securities believes that several reasons may lead to Yipit's high caliber: 1) the Yipit sample included credit cards, which may have a prepaid portion of Creidt; 2) the 1,300 companies in the core sample were early adopters of AI, and the average level is high. Overall, however, there is a clear trend that Yipit previously overestimated Company A in terms of caliber but underestimated OpenAI. CITIC Construction Investment Securities believes that the overall rapid growth trend of the world's big model ARR has not changed.

2/ CITIC Construction Investment Securities believes that as Rubin enters large-scale mass production, it is expected that the upper limit of large-scale model training and reasoning will be further opened. Looking back at the launch of the Blackwell architecture, next-generation models such as GPT‑5.5, GPT‑5.3‑Codex, and GroK 4.5 began to clearly use GB200/GB300 training. The improvement in model capability was mainly reflected in complex programming, professional knowledge work, and long-term smart tasks. Higher cluster size, video memory bandwidth, and connectivity efficiency promote continuous scaling of model training and accelerate the evolution of capabilities such as long context, MoE, and reinforcement learning. Rubin has once again achieved generational improvements in terms of HBM4, NVLink 6, and token throughput per watt. It is expected to support larger parameters, longer thought chains, and more complex agent training. The maximum intelligent space of the model is still worth looking forward to.

3/ Model competition is still fierce, GPT sol series price cut, OxAlpha model sparks discussion

CITIC Construction Investment Securities observed that last week, OpenAI lowered the GPT 5.6 sol short context API price from an input of 5 US dollars and an output of 30 US dollars/million tokens to 4 US dollars and 20 US dollars, respectively. The discount will last until at least November 21. Combined with the previous 80% price reduction for Luna and 20% price reduction for Terra, the competition for the head model is shifting further from a simple competition of ability to price, speed, and agent availability. At the same time, OpenRouter launched the anonymous model OX Alpha, which provides about 1.05 million token contexts free of charge, supports text, images, videos, and tool calls, and targets long-term programming and agent tasks. The cumulative number of calls within a few days of launch is close to 10 billion tokens, and its programming performance and actual source quickly sparked discussion among developers.

Commercialization: GPT-5.6 Sol developer price cut for a limited period of three months

CITIC Construction Investment Securities believes that the comprehensive cost of tokens is still on a downward trend, and the GPT-5.6 Sol developer price will be lowered for a limited period of three months. According to OpenAI's official X account, while continuously expanding the boundaries of capabilities and improving efficiency, the GPT-5.6 Sol API and quota price will be reduced by more than 20% within the next three months. The latest input/output price is 4/20 US dollars per million tokens, with reductions of 20% and 33.3%, respectively. According to SiliconData, the overall LLM Token index has been declining steadily, especially since June of this year, and the downward trend is quite obvious; the closed source model index has declined since July this year, and the trend is steep; the open source model index has climbed from the bottom and the trend is upward, but the increase is currently relatively low.

Token: Last week, OpenRouter expected a 19.6% month-on-month increase in token calls

Last week (8.17-8.23), token calls were expected to reach 90.2 trillion dollars, +19.6% month-on-month, and weekly calls reached a new high. Accelerated growth last week is expected to launch a new model, Ox Alpha. Currently, it is free, and MiMO-v2.5 and Nemotron 3 Ultra are all heavily discounted, driving the rapid release of token demand. Last week, the official version of DeepSeek V4 Flash's token call remained at number one. The anonymous model Ox Alpha 8.21 launched on OpenRouter, which ranked 4th last week. Judging from technical fingerprints, the most likely domestic model was a domestic open source model; 4 of the top 5 last week were domestic open source models, namely DeepSeek V4 Flash 0731, MIMO-v2.5, Hy 3, and DeepSeek V4 Flash 0423.

Of OpenRouter's weekly token calls, the open source model accounted for 77.8% last week, and the Chinese open source model accounted for 70.0% (the data statistics are in the top 9. At the time of data capture, the statistical time window for the previous week was not fully covered; currently, the statistics are incomplete). The share of open source models has declined in the past month, mainly due to the rapid increase in token calls after the closed source model GPT 5.6 LUNA was drastically reduced. Part of the reason for the decline last week was that the newly launched model Ox Alpha was not officially released. It is not clear that it is an open source/closed source model. This model ranked 4th in token calls last week.

Last week's token market share (ZenMux caliber), DeepSeek's share of calls fell back to 37.5%, still ranking first. According to ZenMux (Big Model Aggregation Platform) statistics, the top 5 tokens in market share last week were DeepSeek, Anthropic, OpenAI, Google, and Smart Spectrum, with DeepSeek and Smart Spectrum accounting for 43.4%; Xiaohongshu's big model Dots Studio was on the list. On August 14, Xiaohongshu officially released Dots3-Note Preview and open source, and is currently available free of charge on the ZenMux platform.