Doubao billing: online, Fast, batch, and cache

Beginner: choose a model and then a service

Pro, Turbo, Lite, and Mini are model variants. Online, Fast, and batch are services or billing scenarios, not model names. This comparison explains selection; it does not promise every scenario for every model, region, or channel.

Batch is not a uniform discount on every line item. Check input, output, cache hits, and audio separately. Consult the official sources for upstream prices and capabilities, and the model marketplace for gateway models, endpoints, and current selling prices. This guide deliberately does not maintain a duplicate selling-price table.

Batch does not mean many simultaneous online requests

Concurrent calls to /v1/chat/completions, asyncio, Promise.all, stream: false, or several tasks in one prompt do not turn online requests into upstream batch inference. Client-side asynchrony and response streaming are independent of the service's billing scenario.

The official Volcengine SDK example uses the dedicated client.batch.chat.completions.create method. This is a direct Ark example, not a gateway SDK method. It also does not establish that batch always requires file uploads or guarantees completion within a fixed time.

The current Doubao adaptation includes batch price branches but does not add public batch submission, lookup, or download endpoints. Only a trusted server-side job flow can establish batch identity. Client batch: true, headers, or service_tier: "batch" cannot claim the discount. Consult the endpoint matrix rather than guessing a batch route. Media workflows in the async task guide are not Doubao text batch inference either.

How Fast is selected

The current billing adaptation checks service_tier: "fast" in the effective upstream request after channel filtering and parameter overrides, then freezes the scenario in the request's billing snapshot. A client parameter removed by the channel does not trigger the Fast branch. Field support does not establish model availability, and successful forwarding does not guarantee a fixed latency.

A Turbo model name does not automatically select Fast. Do not apply one model's Fast text-price increase to another model or to audio. Verify unlisted scenarios and items against official information and channel capabilities instead of extrapolating.

Input, audio, and cache without double charging

This breakdown applies when the respective categories are priced separately and upstream usage provides the necessary details. Total input may already include both audio and cache; raw counts must not simply be added together.

Example: 10000 total input tokens include 2000 cache hits and 1000 audio tokens, of which 200 are cached audio. With no other separately priced categories, ordinary input is 7200, cached text 1800, uncached audio 800, and cached audio 200. The total remains 10000. Missing audio-cache details do not justify inventing the intersection.

Copytoken cost = sum(disjoint category tokens * scenario price) / 1000000
storage cost = retained tokens * billable hours * storage price / 1000000

Retaining 100000 tokens for two hours represents 200000 token-hours, not a one-time write of 100000 tokens. Storage requires actual metering; missing storage data must not become a one-hour charge inferred from cache hits. These formulas explain usage units, not the final gateway charge, which also depends on the effective scheme, currency, and applicable group.

Length bands use the complete input

The current Mini form uses len for complete input length. Its 32k and 128k boundaries mean 32000 and 128000 tokens: equality belongs to the preceding band, and a larger input enters the next band. The third displayed band is 128k–256k, not a band starting above 256k or a promise to support requests beyond the model's context limit. Pricing conditions are not request-length validators; check the exact version's input and output limits.

An input of 40000 tokens with 30000 cached tokens still selects the 32k–128k prices, not the lowest band based on the remaining 10000. The band selects rates for the whole request; it is not progressive pricing that splits the first 32000 tokens from the rest. Do not apply one Mini version's bands to the entire Doubao family.

Administrators: fill, save, and read back

With administrator access, open model pricing settings, select the exact model name, and choose per-token pricing. For adapted models, the scenario-pricing form separates online, Fast, and batch items. Mini also has input-length bands. Sharing the Doubao brand does not make every field applicable.

1. Confirm the exact version and displayed currency. Ordinary prices are per million tokens; storage is per million token-hours.
2. Existing base prices populate only existing items, not every scenario. Fill the remaining fields or load reference prices into the draft, then verify official prices. References are not live-synced; their baseline uses 1 USD = 7 CNY, not a current exchange-rate quote or a guarantee that gateway selling prices equal upstream CNY prices.
3. Blank means unconfigured; enter 0 explicitly for free items. Review every band and save. Editing or loading references alone does not change live pricing.
4. Reopen the same model and verify all scenario values, currency, and bands. The form persists one multi-tier billing rule; the multi-tier per-token label does not mean two bills are added together.
5. If a custom expression cannot be represented losslessly, keep reviewing it in the expression editor instead of overwriting it to obtain form fields. If the new fields are absent entirely, check the deployed form version and exact model-name match before investigating stale browser assets.

Public prices and group-specific schemes require separate checks; changing public prices does not necessarily replace every group's scheme. The administrator action to apply one model's prices to multiple models only copies configuration. It is unrelated to batch inference, and configuring a batch rate does not enable upstream service or create jobs.

Fine-tuning is not ordinary inference

Full fine-tuning and LoRA are training charges. Inference with a fine-tuned model or endpoint has separate text and audio pricing. Do not append training rates to each base-model call or replace base-model prices with post-fine-tuning rates. This inference adaptation does not provide training-task APIs or automatic training settlement.

Expert: reconcile requests against their price snapshots

Validate online and supported Fast calls on a small scale against matched tiers, usage, and final consumption logs. Validate batch branches offline or through an explicitly integrated trusted batch flow, not by forging client fields. The following is a suggested business ledger, not a promise that every response contains these fields:

Copyrequest_id, original_model, upstream_model, endpoint, group,
effective_service_tier, matched_tier, full_input_tokens,
text_input_tokens, text_cached_tokens, audio_input_tokens,
audio_cached_tokens, output_tokens, cache_storage_token_hours,
pricing_version, currency, final_charge

Pre-consumption may use estimates; reconcile completed work using actual usage and its frozen pricing rules, not today's rates. See billing, usage, and cost reconciliation for refunds, retries, and the wider cost ledger.