ByteDance put Doubao 3.0 on its international Volcano Engine platform on November 21, priced squarely at the mid-tier API market: $0.35 per million input tokens, $1.40 per million output, and $0.07 on a cache hit. On cache-heavy agent workloads that lands roughly 30% below Qwen4-Max and a little under half of GPT-6.2 Sol.

What the model is

A sparse mixture-of-experts with about 1.1 trillion total parameters and 32 billion active, native vision input, and a 512K context window. ByteDance is not claiming frontier reasoning: the release materials compare against Qwen3.5-Max and Gemini 4 Flash on multilingual and document tasks, and against nobody on hard mathematics.

The interesting part is the tokenizer. ByteDance claims a 22% reduction in tokens per Chinese character against its previous generation, and 8% on English. Since pricing is per token, tokenizer efficiency is a price cut that never appears in the price sheet — and it is the most common reason two models at the same nominal rate cost different amounts in production.

Where it is strong

Document understanding, at least on published numbers. Doubao 3.0 scores 71.2 on internal DocVQA-style benchmarks that ByteDance says were re-run on 4,000 Chinese and English forms, and it handles tables with merged cells far better than the previous generation. For back-office automation in Chinese-speaking markets it is the most direct competitor to Qwen4-Max on price.

It is also fast. Median time to first token in the Singapore region measured under 400ms on a warm cache in our own test, which matters more than benchmark points for anything interactive.

Where to be careful

Three things. First, data residency: the international endpoints run in Singapore and Seattle, and the standard terms permit use of API inputs to improve services unless you opt out on an enterprise agreement. Read that clause before sending customer documents.

Second, English. The model is competent and clearly behind Sol and Astra on long-form English reasoning; the gap shows up on multi-step tasks rather than single-turn quality, which is exactly the pattern that makes it look better than it is in a demo.

Third, latency outside Asia. From Frankfurt we measured 1.9 seconds to first token on the same warm-cache prompt — usable for batch, painful for chat.

How to evaluate it honestly

Run it on your own traffic, not on the release benchmarks, and measure cost per completed task rather than cost per token. The tokenizer difference alone will move your bill by 10% to 20% in either direction depending on language mix, and the cache-hit rate will move it further.

Then decide by workload shape. For document extraction, Chinese-language support and high-volume classification, it is now the default candidate on price. For long-horizon agent work in English, the 32B active footprint shows, and the money you save on tokens you will spend on retries.