This guide uses source checks from Sep 8, 2026. Provider and gateway prices can change; preserve the checked date with every forecast.
Why This Topic Matters Now
A SaaS team adopting Chinese AI APIs rarely needs one more model list. It needs cost governance that survives real product traffic. DeepSeek, Qwen, GLM, and Kimi all expose different combinations of input tokens, cached input, output tokens, context behavior, add-on tools, media modes, failed-call policies, or peak-window caveats. A gateway can simplify endpoint and billing operations, but it does not remove the team's responsibility to label the source row, route owner, account group, and workload class.
The Sep 7 keyword report points to this problem from several directions: Chinese AI API queries, DeepSeek pricing context, QwenCloud billing details, GLM pricing context, Kimi K3 long-context interest, and brand searches that need stronger docs and pricing paths. For Tier 1 and Tier 2 engineering leaders, the practical response is a governance ledger. It should make every forecast reproducible and every variance explainable without exposing customer data or private business metrics.
Source Facts Checked Today
AIWave /api/pricing checked on Sep 8, 2026 returned success=true, 63 records, pricing_version a42d372ccf0b5dd13ecf71203521f9d2, default group ratio 3, and VIP group ratio 1. Computed public base examples per 1M text-token units were DeepSeek V4 Flash at $0.638 input, $0.020288 cache-hit input, and $1.914 output; DeepSeek V4 Pro at $1.914 input, $0.063736 cache-hit input, and $5.742 output; GLM-5 at $1.55 input, $0.400001 cache-hit input, and $4.96 output; GLM-5.1 at $2.10 input, $0.680001 cache-hit input, and $6.60 output; Kimi K3 at $4.50 input, $0.90 cache-hit input, and $22.50 output; qwen3.6-flash at $0.267805 input and $1.606832 output.
QwenCloud pricing docs checked on Sep 8, 2026 returned HTTP 200 and documented pay-as-you-go billing, per-million-token text pricing, context-tiered request billing, failed-call billing behavior, Batch API discounts, context caching, thinking-token billing, built-in tool fees, and separate modes for image, video, speech, embedding, reranking, and tools. A Qwen forecast should therefore record feature family and failure class instead of only prompt and completion tokens.
Z.AI and Kimi pricing pages checked on Sep 8, 2026 also returned HTTP 200. The Sep 7 keyword report records Z.AI rows with input, cached-input, cached-input-storage, output, and built-in tool sections, and Kimi docs with token billing, context caching, 1M-token Kimi K3 context, and a web-search add-on. Those official-provider details are external context. AIWave gateway rows should remain a separate tab with the current AIWave source date and pricing_version.
Planning Matrix
A source-dated planning matrix keeps the page useful for engineers and procurement reviewers. It turns a search query into an auditable route decision instead of a loose model preference.
| Ledger field | Why it matters | Example |
|---|---|---|
| Row owner | Avoids blended provider and gateway claims | AIWave, DeepSeek, Z.AI, QwenCloud, or Kimi |
| Checked date | Prices and policies move | 2026-09-08 |
| Account group | Explains effective gateway multiplier | default, vip, or svip |
| Token class | Separates input, cache, and output | input, cache-hit input, output |
| Feature family | Tools and media bill differently | text, search, image, rerank, speech |
| Failure class | Some providers document failed-call rules | timeout, 4xx, 5xx, blocked, retried |
| Route owner | Keeps changes auditable | team, service, queue, or customer segment |
Implementation Pattern
The implementation pattern keeps credentials as placeholders, pins the AIWave base URL, records the model, and leaves room for route-specific controls. Production applications should move credentials into environment or secret storage.
from dataclasses import dataclass
from datetime import date
@dataclass
class RouteCostPolicy:
model: str
row_owner: str
source_checked_at: str
account_group: str
input_usd_per_1m: float
cache_hit_usd_per_1m: float | None
output_usd_per_1m: float
max_output_tokens: int
deepseek_flash = RouteCostPolicy(
model="deepseek-v4-flash",
row_owner="AIWave",
source_checked_at="2026-09-08",
account_group="vip",
input_usd_per_1m=0.638,
cache_hit_usd_per_1m=0.020288,
output_usd_per_1m=1.914,
max_output_tokens=800,
)
print({"api_key": "YOUR_API_KEY_HERE", "policy": deepseek_flash})
Start With Row Ownership
The first governance rule is simple: never merge rows with different owners. AIWave rows describe gateway pricing and account-group behavior. DeepSeek rows describe direct DeepSeek context. Z.AI rows describe direct GLM context. QwenCloud rows describe Qwen-specific billing rules. Kimi rows describe Kimi-specific context and add-on policy. Put each owner in a separate source column so a later reviewer can update one source without rewriting the whole forecast.
Use Three Token Classes
At minimum, keep fresh input, cache-hit input, and output separate. Many teams create a single blended cost column too early and lose the reason a workload moved. Long prompt templates, repeated reference packs, and agent memory can change cache-hit behavior. Long answers can move output faster than input. A governance ledger that keeps token classes separate can show whether the issue is prompt size, repeated context, answer length, or model choice.
Add Feature Family Columns
QwenCloud, Z.AI, and Kimi pages all remind teams that modern model APIs are not only text completion. Search, tools, image, video, speech, embeddings, reranking, and agent features can have different billing behavior. Add feature_family, tool_name, tool_call_count, and media_units columns before the first production rollout. Even if a route starts as text-only, the schema is ready when product managers ask for a search-enabled or multimodal workflow.
Track Failure and Retry Policy
Cost governance is not only successful calls. Failed-call behavior, retries, and timeout handling can create budget movement and user-visible incidents. Store status class, retry count, stop reason, timeout, and fallback route. If the provider documents billing behavior for failed calls, link the source. If the gateway receipt includes usage after an error, preserve it. If it does not, record that absence as part of the variance investigation.
Respect Account Groups
AIWave public pricing exposes group ratio context. A forecast for a funded VIP account should not be compared directly with a default-group test unless the multiplier is explicit. The account group belongs in every receipt, and a team should recheck it before a monthly forecast. Paid VIP or SVIP behavior is owned by the server-side automation described in current decisions; blog guidance should never tell customers to set token groups manually.
Internal Links for Governance
A governance reader should visit Pricing, Docs, Models docs, Trust, Status, the price-table guide, and the add-on billing ledger. Those links connect this article to live evidence and earlier field notes.
Procurement Review
Procurement should ask engineering for one ledger row per approved route and feature family. The row should include owner, source URL, checked date, account group, model, input, cache-hit input, output, feature family, tool count, failure class, retry rule, route owner, and budget owner. That is enough to compare DeepSeek, Qwen, GLM, and Kimi without flattening them into a misleading average.
Final Checklist
A Chinese AI API cost-governance system is ready when every route has a source owner, checked date, account group, token classes, feature family, retry rule, fallback rule, and review owner. Recheck official provider pages and AIWave pricing before purchase decisions. Keep public proof narrow: dated rate cards, route receipts, model catalog, trust boundaries, and machine-readable status evidence.