bytedance API Pricing

Short answer. ByteDance (Volcano Engine) prices in CNY only, which makes it the one provider in this database where the dollar cost is a currency decision as much as a pricing one. Its structure also includes two features no competitor has: a low-priority tier at roughly half price, and hourly billing for cache storage.

The ladder (CNY per 1M tokens)

Model	Input	Output	Out/in	Cache	Low-priority
Doubao Seed 2.1 Pro	¥6.00	¥30.00	5.00	¥1.20 (20%)	~¥3.00 / ¥15.00
Doubao Seed Evolving	¥6.00	¥30.00	5.00	¥1.20	—
Doubao Seed 2.1 Turbo	¥3.00	¥15.00	5.00	¥0.60 (20%)	—

At an indicative 7.1 CNY/USD, Pro works out to roughly $0.85 input / $4.23 output — competitive with the mid-tier flagships, and cheaper than Kimi K3 by a factor of three on input.

The 5:1 output ratio is the steepest here

Every ByteDance model prices output at 5x input. For generation-heavy workloads — long reports, code synthesis, multi-turn agents — that ratio dominates the bill far more than the input rate. A model that looks cheap on input can be the most expensive option per completed task.

Cache storage is billed hourly

¥0.017 per 1M tokens per hour is a small number that compounds: a 1M-token cache held for a month costs about ¥12, roughly two full re-reads of the same context. For long-lived system prompts this is fine; for caches created per session and forgotten, it is pure waste.

This is the only provider here that charges for cache residency rather than cache reads, and it changes the discipline: you must expire caches, not just create them.

Low-priority is batch by another name

The ~50% low-priority tier is the closest thing to a batch discount in this database, but it is a scheduling class rather than an API mode: work is served when capacity allows, with no latency guarantee. For offline evaluation and data processing it is the cheapest option here; for anything user-facing it is unusable.

Our view

CNY-only pricing is a real procurement cost. It creates FX exposure, complicates budgeting, and puts the invoice in a currency your finance system may not handle. The discount versus international providers has to cover that friction.

The 5:1 output ratio is the fact that decides workloads. Pro is competitive for input-heavy tasks (RAG, classification, extraction) and expensive for output-heavy ones. Most buyers compare input prices and get this backwards.

Hourly cache storage is the most under-appreciated charge in Chinese AI pricing. It is small per token-hour and large per forgotten cache. Any team using long system prompts should model it explicitly.

Low-priority is the best batch substitute in the market — if your pipeline can wait. Half price for offline work is a better deal than a 50% batch discount with a fixed turnaround, provided you can tolerate indefinite queueing.