ByteDance Doubao Pricing 2026: How Cheap Is China's Mass-Market AI Really?
Published 19 September 2026 · BytePlus ModelArk pricing checked 2 November 2026
If you're pricing a Doubao API project, start with one distinction that saves a lot of confusion later: Doubao, Volcano Engine Ark, and BytePlus ModelArk are separate products and billing contexts. A price on BytePlus's international ModelArk service doesn't tell you what you'll pay on mainland China's Volcano Engine Ark, and neither one is the price of the consumer Doubao app.
This guide focuses on the current public BytePlus ModelArk online inference price sheet. On that sheet, dola-seed-2-1-turbo is listed at $0.50 per million input tokens and $2.50 per million output tokens for prompts up to 256K tokens. The same row lists cache-hit input at $0.10/M and cache storage at $0.0083/M tokens per hour. These are usage rates, not a promise that every account or region can call every model ID. Before you budget or build around it, check the exact model ID and tariff available in your own ModelArk workspace.
First choose the platform you can actually use
Doubao is ByteDance's consumer assistant. Volcano Engine Ark serves developers in mainland China. BytePlus ModelArk is ByteDance's international developer service, with its own model list, regions, account setup, and billing. So "Doubao price" on its own is too broad to be useful.
The ModelArk model list says all models listed there are supported in ap-southeast-1, while only certain models are also supported in eu-west-1. Its model and feature tables are version-specific. The pricing page helps you estimate a bill, but it doesn't answer a different question: is the model available to you, in that region, on that service? For that, look at the model list and at what your account can actually activate.
I didn't verify a current public mainland Volcano Engine rate card during this review. You may see RMB figures repeated on comparison sites, but without an official current tariff for your own product and account, they shouldn't be treated as a confirmed price. If you'll use Volcano Engine Ark in mainland China, check its live console and billing terms directly. And please don't take a BytePlus USD rate, convert it, and call it the Volcano Engine rate.
BytePlus ModelArk online rates
BytePlus's pricing page was last updated 23 October 2026. Its standard online inference table lists these rates for dola-seed-2-1-turbo:
| Item | Listed rate | What is billed |
|---|---|---|
| Input | $0.50 per million tokens | Input tokens not counted as cache hits |
| Cache-hit input | $0.10 per million tokens | Input tokens that qualify as cache hits |
| Cache storage | $0.0083 per million tokens per hour | Cached tokens while stored |
| Output | $2.50 per million tokens | Generated output tokens |
The listed prompt tier is up to 256K tokens, so treat that as the scope of these rates. If your requests go beyond it, verify supported request sizes and pricing in your account rather than assuming the same price carries over. The table describes an online service rate. A batch job or another deployment offering may have different prices, and a regional or account-specific agreement may change what you pay too.
Caching can lower the charge for repeated input, but it doesn't make all input free. ModelArk's context-cache documentation describes separate cache-hit input and storage charges for explicit caching. Storage is billed by the natural hour, a fraction of an hour counts as an hour, and storage charges stop when the cache expires or is deleted. For implicit caching, the docs say cache-hit input is charged at the hit rate, while storage is free. Which behavior applies depends on the model and API feature you use, so read that documentation before adding a cache discount to a forecast.
BytePlus also lists a Flex option. Its pricing notes put input and output rates at approximately half the standard online rates, keep the cached-input rate the same, and say Flex supports implicit caching only, with no additional cache-storage fee. The word "approximately" is doing real work there, so use the live price sheet and your account terms for a final estimate. Pick Flex only if the service mode fits your application, and don't read a latency or service-level guarantee into the price discount alone.
What those rates mean for a monthly workload
The examples below use the standard online rates above. They assume every request is within the listed prompt tier, no input qualifies for cache-hit pricing, and no additional product or infrastructure charges apply. The token counts are sample workloads, not a minimum commitment or a claim about what a typical customer uses.
| Example workload | Calculation | Token charge |
|---|---|---|
| Small prototype: 1M input, 250K output | (1 × $0.50) + (0.25 × $2.50) | $1.125 |
| Regular app: 10M input, 3M output | (10 × $0.50) + (3 × $2.50) | $12.50 |
| Larger pipeline: 400M input, 100M output | (400 × $0.50) + (100 × $2.50) | $450 |
At these rates, output costs five times as much per token as uncached input. That's why the second example spends $7.50 on 3M output tokens, more than the $5.00 for 10M input tokens. If your app produces long answers, summaries, or several agent steps, measure output tokens as carefully as input. They're the part that sneaks up on you.
Here's a simple cache illustration. Suppose 1M tokens of input qualify for the listed cache-hit rate instead of the normal input rate. The input charge would be $0.10 rather than $0.50 for that million tokens, before any applicable storage charge. Under explicit caching, holding 1M cached tokens for one billed hour adds $0.0083, and the duration and exact cache mechanics matter. That's why a forecast works best when it separates new input, cache-hit input, cache storage, and output, rather than applying one discount to the whole prompt.
These calculations cover model token charges only. They don't include tools, agent runtime, hosting, taxes, data transfer, or other services your application may need. For background processing, compare the current Flex rate against standard online inference using your real mix of input, output, and cache hits. The potential saving is a pricing difference, not proof that the modes behave identically for your use case.
A practical check before you build around a model
Before you estimate a production bill, confirm the exact name and version shown in your account, the ModelArk region, and whether the endpoint is enabled for your workspace. BytePlus documentation has both model-list and parameter-support pages, and those pages can distinguish versions and mark a model as unavailable for BytePlus currently. A family name like "Seed 2.1" isn't enough to confirm that a specific endpoint is enabled.
Then record your own request mix, counting normal input, cached input, output, and cache duration separately. If you already have a workload, your request logs are the best source. If you're still prototyping, label the volumes as assumptions and rerun the estimate once you have real usage. A per-million price only becomes useful when you multiply it by the tokens your application actually sends and receives.
Finally, check the current pricing and model pages again before launch, since the pricing table and model availability can change independently. Keep a copy of the exact model ID, region, date checked, and applicable billing mode alongside your estimate. Future you will be grateful when a bill needs explaining or the forecast needs updating.
Common questions
Is Doubao API pricing the same as Volcano Engine Ark pricing?
This page doesn't establish that. The figures above are the public BytePlus ModelArk online rates in USD. Volcano Engine Ark is a separate mainland developer platform, so verify its current tariff in the relevant account and region.
Is Seed 2.1 Turbo available to every international account?
The public ModelArk pricing and model documentation list model IDs, but that doesn't prove activation for every account or region. Confirm the exact ID in your workspace before depending on it.
Does cached input always cost $0.10 per million tokens?
That's the listed cache-hit rate for the Dola Seed 2.1 Turbo row in the standard online table. You have to actually qualify for a cache hit, and explicit cache storage can add a separate hourly charge. The context-cache documentation explains how hit and storage billing work, and implicit caching treats storage differently.
Does the Flex rate cut the whole bill in half?
No. BytePlus describes input and output rates as approximately half the standard online prices. Its notes keep the cache-hit input rate the same and say Flex has implicit caching only, with no separate storage fee. Apply each applicable rate to its own token category.
Is this the price of the Doubao consumer app?
No. This is a developer API guide for BytePlus ModelArk online inference. Consumer app subscriptions and mainland Volcano Engine billing are separate questions.
Sources checked on 2 November 2026
- BytePlus ModelArk pricing — standard and Flex rates; page last updated 23 October 2026.
- BytePlus ModelArk model list — listed model IDs and region support; page last updated 23 October 2026.
- BytePlus context cache documentation — cache-hit, storage, expiration, and implicit-cache billing rules.
- BytePlus model parameter support — model-version-specific parameter and availability notes.