Billing and Pricing

XSI bills in real time from the actual usage reported by the upstream: tokens, cache, images, video seconds, audio duration, and search and tool calls can each be priced separately. Balances and charges are always in Chinese yuan (CNY).

Where prices come from

SourceDescription
api/pricing.jsonDefault system price table, shipped with the image and loaded at startup; changes require a restart
Tenant PRICINGTenants override or add model prices through /x-config; each model's entry is replaced as a whole and takes effect immediately

GET /x-pricing shows the system list price, the price after tenant overrides, and the account's final CNY price. See Pricing catalog API.

Currency and multipliers

The price table records the provider's list price in its original currency, which multipliers then convert to CNY:

Charge = list price Γ— usage Γ— model Rates Γ— (fast tier multiplier, tokens only) Γ— account Rates Γ— account Factor
MultiplierSet byDescription
Model RatesPrice tableConversion and markup factor. In the default price table, it is 7 for models priced in USD and 1 for models priced in CNY
Account RatesParent accountA sub-account's value is not lower than its parent's; the difference between parent and child is the parent's resale margin
Account FactorTenant ownerPrice coefficient, useful for discounts and markups for different customers or channels

Charges are deducted only from the calling account's own balance, never level by level up the ancestor chain. The parent's margin is realized when it transfers credit to a sub-account, converted at child.Rates Γ· parent.Rates.

Price table structure

{
  "ChatPricing": {
    "gpt-6.2-sol": {
      "InputText": 2, "CachedText": 0.1, "CacheWrite": 2.5, "OutputText": 10,
      "FastTokenMultiplier": 2, "Rates": 7,
      "ContextTiers": [{"InputThreshold": 272000, "InputText": 4, "CachedText": 0.2, "CacheWrite": 5, "OutputText": 15}]
    }
  },
  "ImgPricing": {"grok-imagine-image-quality": {"Sizes": {"1k": 0.05, "2k": 0.07}, "InputImage": 0.01, "Rates": 7}},
  "VideoPricing": {"grok-imagine-video": {"Resolutions": {"480p": 0.05, "720p": 0.07}, "InputImage": 0.002, "Rates": 7}},
  "AudioPricing": {"gpt-live-1": {"Unit": "per_minute", "Input": 0.05, "Rates": 7}}
}

Chat models ChatPricing

FieldUnitBilled for
InputText, InputAudio, InputImagePer million tokensInput that misses the cache
CachedText, CachedAudio, CachedImagePer million tokensCache reads (not counted again as input)
CacheWritePer million tokensCache writes
OutputText, OutputAudio, OutputImagePer million tokensOutput
ReasonTextPer million tokensPrices reasoning tokens separately; if unset, reasoning tokens are included in output and billed at OutputText
QueryPer callTool calls such as web search, x_search, and code execution; DeepSeek search and fetch
FindPer callFile search
PagePer pageDocument pages
SizeHigh, SizeMedium, SizeLowPer callTiered tool calls in Responses
CallPer requestFixed fee in CNY, not multiplied by the model Rates
FastTokenMultiplierMultiplierToken multiplier for the fast tier (applies when greater than 1)
ContextTiersβ€”Long-context tiers, see below

Other categories

CategoryBilling
ImgPricing(Sizes[size] Γ— image count + InputImage Γ— input image count + Call) Γ— rate multiplier; sizes not configured use Call as the per-image price, and the grok-imagine series uses the highest configured price
VideoPricing(Resolutions[resolution] Γ— seconds + InputImage Γ— input image count + Call) Γ— rate multiplier; duration 1–15 seconds, default 6 seconds; resolution defaults to 480p, and unknown resolutions are billed at the highest price
AudioPricingBy duration (per-second billing) when Unit: "per_minute"; otherwise by tokens or characters
RerankPricingBy token
CallPricingFixed price per request
FineTuningPricingBy token, for ft: models

Image models that also appear in ChatPricing (such as gpt-image-2.6-*) are billed by token.

Caching

  • Cache reads are billed separately at the Cached* rates and are no longer counted as input.
  • Anthropic cache writes (cache_creation_input_tokens, both 5-minute and 1-hour) are billed at CacheWrite.
  • OpenAI cache writes (input_tokens_details.cache_write_tokens) are included in input: 30-minute writes are billed at CacheWrite and 12-hour writes at 2 Γ— CacheWrite; if the model has no CacheWrite, the input price applies.

Fast tiers

Value in request or responseTierToken multiplier
speed: fast, service_tier: priority or fastfastThe model's FastTokenMultiplier
speed: ultrafast, service_tier: ultrafastultrafast6
default, standard, flexStandard1

The tier reported in the upstream response takes precedence; if the response does not report one, the request value is used. Multipliers apply only to token charges.

Long-context tiers

When a request's total input tokens (including cache reads and writes) exceed a tier's InputThreshold, all tokens in that request are billed at that tier's prices; fields not set in the tier fall back to the base prices. For example, when input to gpt-6.2-sol exceeds 272000 tokens, the input and output prices become 4 and 15.

What gets billed

CaseBilled
POST request where the upstream returns 200 (any 2xx for video creation)By actual usage
Streaming request disconnected midwayBy usage generated so far
Upstream response missing usageEstimated from output content and the conversation
GET, DELETE (fetching results, polling video, deleting files, etc.)Not billed
Upstream error (non-200)Not billed
POST rejected by the resource allowlist, POST rejected by the model allowlist or a model-level limit (403/429), or a request body that cannot be parsedXAPI_FEE (default Β₯0.012); a 429 from account-level rate limits is not charged
Successful, but the model is not in the price tableXAPI_FEE, and the operator is notified
Some management queries (see each endpoint)XAPI_FEE

XAPI_FEE appears on the bill under the model name XAPI and is capped at Β₯0.24.

Credit and expiry

  • A balance is made up of multiple credits, each with its own amount, remaining balance, source note, and expiry time; the available balance is the sum of unexpired credits.
  • Validity: admin top-ups follow Days (default and maximum 365 days); initial credit granted when a tenant owner creates a sub-account lasts 1 year (100 years above 100 million); credit transferred by a regular manager expires no later than its source credit; signup bonuses last 1 month; subscription credit follows the plan period.
  • Credits come in two fund types: pay-as-you-go and subscription. Role 3 (subscription) accounts use subscription credit first and can fall back to the pay-as-you-go balance once it runs out.
  • A charge can push the balance briefly negative; the debt is recorded as negative credit and is offset first by subsequent top-ups.
  • When the balance drops below USER_MIN_BALANCE βˆ’ 0.01, the account is frozen (429) and recovers automatically after a top-up.