If your team pipes documents and screens through a model every day, the buying question usually narrows to one fork: call an expensive flagship for every page, or bolt eyes onto a cheaper model and run it in bulk. Ling-3.2-flash, released by Ant Group's Ling team on November 5, plants itself squarely on the second branch. But the first thing to read is not its scorecard. The model is free for its first two weeks — and the official pricing table still has no row that says what it costs after that.
The short version
Ling-3.2-flash is a 560B-total, 25B-active MoE with a 1M context and native image, text, and video input — the same body as the text-only 3.1, with eyes added (official changelog and model docs, checked directly)
The only price the official pricing page confirms is "free for the first 2 weeks." The reported list price of $0.30 in / $0.90 out per million tokens matches the previous generation's official list price, but the 3.2 row itself is still missing from the table (re-checked Nov 8, 2026)
The headline scores are Ant's own picks, measured by Ant: 92.21 on OmniDocBench 1.5, 43.71 on ZEROBench (Sub), and 85.86 on Vision2Web, each ahead of Claude Opus 5 (max) — with no independent rerun yet
Ling-3.2-flash is an MoE model with 560B total parameters and 25B active per token. Its body is exactly the size of the previous Ling-3.1-flash and the context window is the same 1M; the real change is on the input side. Where 3.1 was text-only, 3.2 takes images and video natively at the architecture level — not a bolted-on vision module, but visual information as a first-class input to reasoning, per the official model docs.
The company aims it at four things: complex visual understanding; multimodal long-horizon agents that keep a goal across multi-step tasks using visual and text input; 3D understanding and interaction; and medical multimodality, reading medical images and reports. Trying it is deliberately easy: it is live in the chat interface, and the API accepts both the OpenAI and the Anthropic protocols, so code written for Claude or GPT can be pointed at it by changing the model name.
The fine print matters, though. The official multimodal guide caps video at 30 seconds, one clip per request, sampled at a fixed 2 fps for at most 32 frames, and images at 40 per request inside a 32 MiB body. This is not a model for feeding whole meeting recordings; it is sized for short screen captures and documents split into pages.
On price, exactly one thing is confirmed by the official pricing page: the first two weeks after launch are free, and billing reverts to the original price afterwards. The problem is the original price. The same page's tables list Ling-3.1-flash and Ling-3.0-flash-VL — but there is still no Ling-3.2-flash row. The reported list price, $0.30 per million input tokens and $0.90 per million output tokens, is exactly the previous generation's official list price. Plausible, but until the row exists, no budget should treat it as settled.
| Model (API, per 1M tokens) | Input | Output | Cache read | Notes |
|---|---|---|---|---|
| Ling-3.2-flash | $0.30 (reported) | $0.90 (reported) | Not confirmed | Free for first 2 weeks (official pricing page, checked directly); list-price row not yet published |
| Ling-3.1-flash | $0.30 list → $0.075 discounted | $0.90 list → $0.225 discounted | $0.06 list → $0.015 discounted | Text-only, 75% off (official pricing page, checked directly) |
| Claude Opus 5 | $5 | $25 | $0.50 | Ant's chosen comparison target (Anthropic official pricing, checked directly) |
| Claude Opus 5.5 | $4 | $20 | $0.20 | Current Anthropic flagship |
If the reported price holds, Ling-3.2-flash would list at roughly one-seventeenth of Opus 5 on input and one twenty-eighth on output — an unusually wide gap between a flagship document model and a Flash-tier one. The other direction matters too: 3.1 currently sells at a quarter of its list price, yet no document says what discount structure, if any, follows 3.2's free window. And the payment terms are a practical gate for teams in Korea: the official pricing page requires a personal Alipay account before the first call, enterprise Alipay is not yet supported, and enterprise invoicing is still in development. A cheap sticker and a payable invoice are still two different sentences on this platform.
The performance claims should be read as three vendor-picked events: document understanding on OmniDocBench 1.5 (92.21 vs Opus 5 max's 90.56), visual reasoning on ZEROBench (Sub) (43.71 vs 39.52), and image-to-web generation on Vision2Web (85.86 vs 80.20). The margins run 1.65 to 5.66 points — "stands in the same ring as a flagship on document and screen tasks" is the accurate reading, not a rout. The announcement claims no lead in general intelligence, coding, or long agentic work overall, and that silence is itself information about where this model sits.
Three days after launch, the one signal visible outside the company is an arena. On DesignArena, which turns one image into frontend code, Ling-3.2-flash placed in the top ten at Elo 1238, in the same band as Meta, DeepSeek, and Anthropic models, per the announcement; the leaderboard above shows the spot. Arenas differ from vendor test tracks because real users bring real prompts and pick winners — but the underlying ranking data was not independently re-checked for this article.
The use list itself is the official docs' four target areas: document processing that extracts fields from scans and charts and reasons over them; GUI and office agents that work across screens over multiple steps; 3D assistance that reads drawings and spatial layouts; and medical image and report interpretation support. Protocol compatibility is the practical bonus: because both the OpenAI and Anthropic formats are accepted, a document pipeline running on a flagship today can be A/B-tested against this model with the same prompts and the same eval set, without a rewrite. The flip side: three days in, there are still no confirmed outside production deployments and no independent performance measurements in the checked sources. The case section is thin not because the model is weak, but because only three days of verification time exist.
First, who designed the test. All three events were chosen and measured by Ant, against Opus 5 in its max configuration. Winning at document and visual events is a different sentence from winning overall, and no independent lab has rerun the same document sets under its own conditions yet. Medical multimodality, notably, is a stated target with no medical score in this announcement at all.
Second, the blank in the price list. The free fortnight is confirmed and the same-as-3.1 list price is reported, but until the pricing table grows a 3.2 row, cache pricing, discount structure, and the real post-promo rate are all unsettled. Wire a whole pipeline in during the free window and a surprise price structure arrives after the switching cost is already paid. The free window is best used not as free usage, but as the period to measure per-task token counts and success rates on your own documents.
Third, the conditions of access. Payment runs through personal Alipay, enterprise accounts and enterprise invoices are not there yet, and the data path runs through a Chinese platform — all items a Korean team's data policy has to clear first. There are also no official numbers for Korean-language documents or screens: a high document-understanding score says nothing about Korean invoices until you measure them yourself.
The conclusion splits cleanly. Confirmed: a 560B/25B native multimodal model with a 1M context and dual-protocol compatibility has shipped; the two free weeks are written on the official pricing page; and 3.1's official list price is $0.30 in and $0.90 out. Not yet confirmed: whether 3.2's list price lands at that reported number, whether the document and visual performance holds outside Ant's three chosen events, and how it handles Korean documents in practice. The test fits inside the free window: run ten document and screen tasks you currently send to a flagship through both models with identical prompts, and compare success rate, total tokens per task, and the share of outputs a human throws away. Only after those numbers exist can anyone state this model's real price.
Puzz, in one line
Judge this model not by its scorecard but by the empty row in its price table. Ling-3.2-flash is a model whose real price is the one you measure on your own documents during the free two weeks, ribbit.