Volcano Engine Coding Plan Exposed for Double-layer Billing: Nominal on a Per-Use Basis, but Actually Deducted based on Token Consumption Equivalent to Usage

According to 1M AI News monitoring, users discovered an undisclosed hidden mechanism in the billing method of the Volcano Engine Ark Coding Plan. Some developers reported on V2EX that the consumption rate of quotas was much faster compared to similar packages on other platforms. After contacting customer service, they received a reply stating, "If the tokens consumed in a single model call are much higher than the average call consumption, the estimated number of calls that can be made within the cycle will also be much less than 6,000 times."

Using the Pro package (200 RMB/month) as an example, the stated quota is 6,000 requests every 5 hours, 45,000 every week, and 90,000 every month. However, the actual deduction is not a one-time deduction per call but is converted into multiple deductions based on the consumption of a single token. The conversion formula calculated by users is `usage = max(round(use_token/token_limit), 1)`, and the hidden multiplier varies for different models: approximately 2 times for DeepSeek-V3.2, approximately 4 times for Doubao-Seed-2.0-Code, and approximately 6 times for Doubao-Seed-2.0-Pro. This means that a single call using Doubao-Seed-2.0-Pro may be counted as consuming 6 quotas.

As an example, a user mentioned that their single call consumed 510,000 tokens, which would be counted as only one call on platforms like Ali Bailian, but on the Volcano Engine, it may be counted as approximately 20 calls. In AI programming tasks executed by the Agent, it is common for a single call to consume hundreds of thousands or even millions of tokens. This billing method causes the package quota to be quickly depleted.

Currently, domestic Coding Plan packages in China are generally billed based on the number of calls, with platforms like Ali Bailian and Xiaomi MiMo deducting one call per call without token conversion. The double-layered billing system of the Volcano Engine, which is "stated by number of times, converted by token," is relatively rare in the industry and is not clearly explained in a prominent position on the package page. Users only become aware of this mechanism after contacting customer service due to abnormal consumption. Recently, users have also reported that ByteDance's AI programming tool Trae has switched from pure per-call billing to a similar token conversion-based billing.