On November 5, Ant Ling released Ling-3.2-flash, a model with 560 billion total parameters and roughly 25 billion activated per inference step. API input and output prices start at about RMB 2 per million tokens. We believe its real significance lies not in the vendor’s claim that it “outperforms Claude Opus 5,” but in the substantial price reduction for image, video, and long-document processing.
What this is
Ling-3.2-flash uses a Mixture-of-Experts (MoE) architecture, activating only a subset of parameters for each request to reduce computing costs, and supports a 1-million-token context window. Images, video, and text can be fed directly into the model without first converting visual inputs into text through a separate vision model.
It is also compatible with the OpenAI and Anthropic API protocols, so most existing applications only need to change the endpoint and model name. The service is free for the first two weeks after launch, after which input will cost about US$0.30 per million tokens and output US$0.30–$0.90 per million tokens—approximately RMB 2–6 at the exchange rate used in the article.
Ant says the model surpassed Claude Opus 5 (max) on OmniDocBench, ZEROBench, and Vision2Web, and targets use cases including visual understanding, long-horizon agents, 3D interaction, and medical imaging. For now, however, all these results come from tests selected by the vendor and cannot be treated as independent conclusions.
Industry view
We believe the model’s clearest competitive advantages are its pricing, 1-million-token context, and API compatibility—not its ranking on leaderboards. One of the largest cost barriers to multimodal applications has been the repeated billing of images and video after they are segmented and converted. Combining native input with low-cost APIs could help applications such as video understanding and bulk document analysis become affordable more quickly.
The risks are equally clear. First, the benchmarks behind the official claim that it “outperforms Claude Opus 5” have not been independently retested, and a few-point advantage on extremely difficult tasks may not translate into consistent production performance. Second, Ant is not releasing the model weights, preventing enterprises from deploying it privately and requiring additional assessment of data boundaries and long-term version control. Over the same period, Baidu released the open weights for ERNIE 5.5 Flash Base, indicating that Chinese models are diverging into two tracks: “open weights” and “closed API.”
Impact on regular people
For enterprise IT: Calling costs for multimodal applications may continue to fall, making image, video, and long-document processing easier to incorporate into routine projects. Production purchases should still be tested against real business samples to measure accuracy, latency, and total cost.
For professionals: Complex presentation analysis, meeting-recording summaries, and cross-document retrieval will gradually become standard features. In the short term, however, they are better suited to boosting efficiency than to making unreviewed business decisions.
For the consumer market: Low-cost video understanding and content generation will spread more quickly into office software, e-commerce customer service, and short-video services. Consumers will gain more convenience, while also facing more frequent AI-generated content and authorization issues.