Volcengine Ark Review: Pricing & Comparison
ByteDance's own model gateway: the Doubao Seed family plus DeepSeek/GLM, batch inference at 50% and a daily free allowance
Last verified: 2026-10-15 · Visit official site →
The platform owner’s own gateway
Volcengine Ark is the model-serving product of Volcengine, which is ByteDance’s own cloud. That determines how it differs in category from the other vendors on this page: it is not a relay — the platform owner is also the owner of the models (Doubao/Seed) and the compute. None of the usual relay risks apply, such as an upstream channel being cut off or quota being clawed back, and contracts, invoicing and compliance all have a clear counterparty.
The three things currently promoted on the product page also signal its direction: Seedance 2.5 opening up fully, the launch of the Ark CLI (supporting Claude Code, Cursor and TRAE), and the Ark Agent Plan.
Pricing: tiered by length and time of day, built for cost engineering
Selected pricing (¥ per 1M tokens, input/cache/output):
- doubao-seed-2.1-lite: ¥0.8 / ¥0.16 / ¥2.7 (audio input ¥12)
- doubao-seed-2.1-pro: ¥6 / ¥1.2 / ¥30
- doubao-seed-evolving: ¥6 (0-1024 length) / ¥1.2 / ¥30
- doubao-seed-2.0-mini: [0,32] ¥0.2 in → ¥2.0 out; [32,128] ¥0.4 → ¥4.0; [128,256] ¥0.8 → ¥8.0
- doubao-seed-2.0-pro: [0,32] ¥3.2 in → ¥16.0 out; [32,128] ¥4.8 → ¥24.0; [128,256] ¥9.6 → ¥48.0
- deepseek-v4.1-flash: peak/off-peak tiers (peak = Mon-Fri 09:00-12:00 and 14:00-18:00 Beijing time)
Batch inference: 50% of online inference unit prices; transparent prefix caching cuts cached input by a further 60%Beyond pay-as-you-go, there are low-latency, standard and low-priority online inference tiers, TPM guarantee packs and model units, which makes cost versus latency a tunable parameter rather than a fixed choice. Note that the official documentation states the price pages are reference only and the actual cost is confirmed at checkout.
Promotions and referral vouchers
The current promotions are stated concretely: the Agent Plan from ¥9.9 for a limited time (the docs title it the Agent Plan Small & Medium 2.5x-discount programme), the second phase of the collaboration rewards programme offering up to 5 million free tokens per model per day, and a referral voucher — invite a friend who orders and you receive a voucher worth 5% of their order with no cap, while they get an extra 5% off their first subscription.
One premise you have to accept
There are no overseas models here. The catalogue is the first-party Doubao/Seed family plus a handful of third-party domestic models (GLM 5.3 Flash, GLM 5.2, DeepSeek V4.1 Flash, DeepSeek V4 Pro/Flash) — no GPT, Claude or Gemini, and no Kimi or Qwen entries either. If your stack needs those models, Ark cannot help, which is exactly why relays exist. Conversely, if what you want is domestic models with platform-grade stability and proper invoicing, it is the steadiest option among the vendors added in this round.
Two further caveats: the product page promises production-grade guarantees without publishing any availability percentage, so SLA terms require a sales conversation; and some model IDs (such as deepseek-v4-pro-260425) are already flagged as retiring, while pro tiers default to 500 RPM / 1M TPM, which means high-concurrency workloads need to buy additional guarantee packs.
Information verified 2026-10-15. Pricing, model list and promotions reflect what the volcengine.com product page, official docs and the activity page displayed on that date.
Related reviews
- Zhipu AI: official Zhipu GLM series, multimodal AI, top-tier Chinese-language capability
- Tencent Cloud Hunyuan: Tencent’s official Hunyuan model, deeply integrated with the Tencent Cloud ecosystem
- Baidu Qianfan: official Baidu platform, ERNIE model family, enterprise-grade SLA, supports model fine-tuning
- SiliconFlow: domestic open-source LLM cloud service, 100+ models including DeepSeek/Qwen/GLM, self-built inference engine + domestic chip support
Quick facts
| Pricing model | Pay-as-you-go in ¥ per 1M tokens, tiered by input length and time of day: doubao-seed-2.1-pro ¥6 in / ¥1.2 cached / ¥30 out; doubao-seed-2.1-lite ¥0.8 in / ¥2.7 out; doubao-seed-2.0-mini ¥0.2 in / ¥2.0 out at short lengths; deepseek-v4.1-flash is billed in peak and off-peak tiers; batch inference costs 50% of online inference and transparent prefix caching cuts cached input by a further 60% |
|---|---|
| Model coverage | First-party Doubao/Seed family (doubao-seed-evolving, seed-2.1-pro/lite/turbo, seed-2.0-pro/lite/mini, Seedance 2.5 video, Seedream 5.0 pro image) plus selected third-party domestic models (glm-5-3-flash, glm-5-2, deepseek-v4-1-flash, deepseek-v4-pro-ga, deepseek-v4-flash-ga); no GPT, Claude or Gemini, and no Kimi or Qwen entries in the model list |
| Latency / SLA | The product page promises only production-grade guarantees and elastic-concurrency latency assurance, with no availability percentage published; it also offers low-latency, standard and low-priority inference tiers, TPM guarantee packs and model units; the official docs state that pricing pages are reference-only and the actual cost is confirmed at checkout |
| Mainland direct connect | Mainland direct-connect (ByteDance's own cloud, cn-beijing region) |
| Best for | Developers / Enterprise |
| Referral program | The Agent Plan page publishes a referral voucher: invite a friend who places an order and you receive a voucher worth 5% of that order, with no cap, while the friend gets an extra 5% off their first subscription. |
Pros
- The platform owner is also the owner of the models and the compute (ByteDance): none of the relay risks around upstream channels being cut off apply, and invoicing, contracts and compliance all have a clear counterparty
- A fully published price sheet with real cost engineering levers — batch inference at 50%, transparent prefix caching for a further 60% off cached input, TPM guarantee packs, and low-latency vs low-priority tiers — rather than a single headline number
- Promotions are stated concretely and verifiably: Agent Plan from ¥9.9, up to 5 million free tokens per model per day, and a 5% referral voucher
Cons
- No overseas models at all (no GPT, Claude or Gemini), so developers who need those get no answer here
- The product page promises production-grade guarantees but publishes no availability percentage, and the docs state that pricing pages are reference-only with the real cost confirmed only at checkout
- Some model IDs are already flagged as being retired (for example deepseek-v4-pro-260425), and pro tiers default to modest limits of 500 RPM / 1M TPM, which is tight for high concurrency
Compare more AI API relays
See the full comparison board — filter by price tier, model coverage, and mainland direct-connect status.
Back to the comparison board →