Verified, then logged.
Sun 22 Nov 2026 · 00:23 KST / 15:23 UTC한국어 | ENSearch /

02 / IT · Analysis

Ant's Ling-3.2-flash tops Claude Opus 5 in visual benchmarks

Ling-3.2-flash's lead over Claude Opus 5 is limited to some evaluations, while Ant Group is pitching low usage fees and a context of 1 million tokens.

An analyst seen from behind compares two side-by-side monitors showing the same video stills with objects outlined differently on each screen in a dim office. AI-generated image

Ant Group's Ling team unveiled Ling-3.2-flash, a multimodal model for the enterprise AI agent market, on the 5th, according to Korean AI news outlet AI Times. Its selling points are low usage fees and a context of 1 million tokens. The model also reportedly outperformed Anthropic's Claude Opus 5 in some evaluations.

AI Times reported the performance comparison and the usage fees together. The lead over Opus 5 is limited to some evaluations. That boundary has to be drawn first for the report to be read correctly.

AI Times highlights the model's ability to understand images and video. A win on visual tasks means something different from a ranking on other work. The first question is which evaluations it led.

The lead sits in visual understanding

AI Times named visual understanding evaluations as an area where Ling-3.2-flash beat Claude Opus 5. That category sits close to the core work of a multimodal model that handles images and video. Reading it as an overall performance lead goes beyond what the report says.

The same report said the model placed in the top 10 of an external web-building evaluation. Web building is a different kind of task from visual understanding. What sets this ranking apart is that it came from outside the company.

An enterprise push built on low usage fees

According to AI Times, usage fees and context are what Ant Group is leading with in the enterprise market. Context is the span of input a model can refer to at one time. Ling-3.2-flash can handle up to 1 million tokens of it.

The report also gives the model's size. It has 560 billion parameters in total, but only 25 billion are active during inference. Judging inference scale from total size alone misses that structure.

How low the fees are can be judged only by setting per-token rates beside those of rival models. The lead in some evaluations needs rereading in that same comparison. A verdict on Ling-3.2-flash as an enterprise model can follow from there.