Multimodal deep search: breaking the boundary between text and vision
In early October 2026, multiple AI search engine vendors launched multimodal deep search capabilities. Users no longer rely solely on keywords or image tags for retrieval; they can directly upload an audio clip, a complex chart, or even a short video, and the AI model will understand it end-to-end and return answers or related content. For example, a user uploads a blurry street photo and asks, "Which city and street is this?" The system automatically reasons by combining location information, architectural style, and historical databases. This update relies on the mature deployment of sparse attention architectures and multimodal alignment models, making search response times near real-time. Previously, similar functions were limited to lab environments; now, through optimized inference engines, they can support hundreds of millions of queries per day.
It is worth noting that these new features place higher demands on API call volumes. Many small and medium developers, using the aggregation services provided by AI API relay stations, can connect to multiple search model interfaces at once and enjoy unit price discounts from token wholesale. Some practitioners revealed that through token relay stations, they obtain the cheapest tokens at only 60% of the official direct purchase price, greatly reducing experiment costs.
Personalized memory: search starts to 'remember' users
Another highly anticipated feature update is the personalized memory capability of search engines. After user authorization, the AI search can maintain a dynamic memory repository over the long term, including common search preferences, professional domain interests, and even summaries of historical conversations. For example, users who frequently focus on healthcare content will, when searching for "side effects" in the future, have the system prioritize authoritative sources in that field and filter out irrelevant entertainment results. This feature relies on the collaboration between on-device small models and cloud large models, protecting privacy while improving efficiency. Multiple vendors emphasize that users can view, delete, or modify memory content at any time, with data anonymization processing complying with current privacy regulations.
The implementation of personalized memory has also spawned new business cooperation models. Some AI hub platforms have begun offering "memory as a service," allowing developers to embed customized memory logic into their applications, while the underlying model calls are completed through AI channel agents. In this model, developers do not need to build complex inference clusters themselves; they only need to use AI large model agents or token agents to allocate memory modules from different models, with pay-per-use being very flexible.
Cost revolution: token wholesale and AI relay stations reshape the API ecosystem
With the launch of high-consumption features such as multimodal deep search and personalized memory, API call costs have become a focus of industry attention. In mid-October, an obvious trend is the full popularization of the "token wholesale" model. In the past, enterprises buying AI API quotas typically required fixed prepayments or were charged at standard unit prices. Now, through third-party AI relay stations (such as AI API relay stations) for bulk purchases, significant discounts can be obtained during off-peak or non-peak hours. Some aggregation platforms have even launched "night-time all-day" packages, allowing companies that process non-real-time requests in batches to obtain similar service quality at lower prices.
At the same time, AI channel agents and AI large model agents have also adjusted their strategies: they not only offer the cheapest token options but also bundle value-added services such as model fine-tuning and cache strategy optimization. An industry report estimates that after adopting the token wholesale model, the monthly API spending of small and medium-sized companies has dropped by an average of about 40%, freeing up more budget for polishing the user experience of search products. An AI search entrepreneur said: "In the past, we dared not easily launch high-cost multimodal search; now, through the aggregated bargaining of AI hub, we finally have the conditions to let every user enjoy deep search without worrying about cost runaway." This change will push more AI applications from "usable" to "easy to use."