This guide uses source checks from Sep 4, 2026. Provider and gateway prices can change; preserve the checked date with every forecast.
Why This Topic Matters Now
Chinese AI API price tables are easy to misread because the columns are not equivalent. DeepSeek separates peak and off-peak cache-hit, cache-miss, and output rows. QwenCloud documents Batch API, context caching, thinking tokens, failed-call policy, and separate tool fees. Kimi K3 uses cache-hit, cache-miss, and output rows with a 1M-token context window. Z.AI lists cached input, tool fees, media rows, and a GLM-5.3-Flash promotional window that ends on September 9, 2026 UTC+8.
This article gives Tier 1 and Tier 2 API buyers a normalization worksheet. It is deliberately practical: keep row owner visible, store source date, split token classes, mark promotions, and separate AIWave gateway rows from direct provider rows. A finance forecast built this way is slower to write, but much easier to defend when the first invoice or usage ledger arrives.
Source Facts Checked Today
DeepSeek official docs crawled for Sep 2026 show prices per 1M tokens and warn that product prices may vary. V4 Flash and V4 Pro each have cache-hit input, cache-miss input, output, peak, and off-peak rows. The page says peak hours are 01:00-04:00 and 06:00-10:00 UTC Monday through Friday, with all other hours off-peak. It also lists 1M context, 384K maximum output, and route concurrency limits.
QwenCloud pricing read through Jina on Sep 4, 2026 says text models bill input and output per million tokens, some models use single-request input tiering, Batch API uses 50% of real-time input and output rates for eligible async workloads, context caching discounts cached input, thinking tokens count as output tokens, and Batch plus cache discounts cannot be combined on the same request. It also lists Web Search at $10 per 1K calls and Image Search at $8 per 1K calls.
Kimi official pricing pages checked on Sep 4, 2026 say Kimi API bills input and output tokens separately, Web Search adds $0.004 per invocation, Kimi K3 has a 1M-token context with no context-length tiering, and Kimi K3 prices are $0.30 cache-hit input, $3.00 cache-miss input, and $15.00 output per 1M tokens. Z.AI pricing checked the same day lists GLM-5.3-Flash at $0.075 input, $0.015 cached input, and $0.25 output during a promotion ending September 9, 2026 UTC+8, with GLM-5.3 and GLM-5.2 at $1.40 input, $0.26 cached input, and $4.40 output.
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.
| Worksheet field | Why it matters | Example source signal |
|---|---|---|
| row_owner | Direct and gateway terms differ | DeepSeek, QwenCloud, Kimi, Z.AI, or AIWave |
| source_date | Rows change | Sep 4, 2026 check date |
| unit | Per-token, per-call, image, second, or character | QwenCloud media and speech sections |
| cache_status | Hit and miss rows are not the same | DeepSeek, Kimi, Z.AI cached-input columns |
| output_policy | Reasoning and generated text can dominate | Kimi and Qwen output billing notes |
| tool_fee | Search can add a separate line | Kimi Web Search and QwenCloud tool rows |
| promo_end | Temporary rates expire | Z.AI GLM-5.3-Flash through Sep 9, 2026 |
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 decimal import Decimal
api_key = "YOUR_API_KEY_HERE"
row = {
"row_owner": "aiwave",
"model": "deepseek-v4-pro",
"source_checked_at": "2026-09-04",
"input_per_1m": Decimal("1.914"),
"cache_hit_per_1m": Decimal("0.063736"),
"output_per_1m": Decimal("5.742"),
"account_group": "vip",
"pricing_version": "a42d372ccf0b5dd13ecf71203521f9d2",
}
print({"placeholder": api_key, "normalized_row": row})
Start With Row Owner
The first worksheet column should name who owns the row. A direct DeepSeek row is not an AIWave gateway row. A Kimi membership plan is not a Kimi API token row. A Z.AI promotional row is not a permanent internal budget. When row owner is explicit, reviewers can ask the right follow-up: account terms, gateway multiplier, source date, currency, feature family, and whether the row is eligible for the workload.
Normalize Units Before Ranking
Do not rank prices until the units match. Text tokens, cached input, output tokens, thinking tokens, search invocations, generated images, video seconds, speech characters, and gateway quota units do not answer the same budget question. Put every row on its native unit first, then create a separate calculated view if a blended workload estimate is needed. The native row should remain visible for audit.
Split Cache Hit and Cache Miss
Cache rows deserve separate columns because a repeated prompt is not guaranteed. DeepSeek, Kimi, and Z.AI all expose cache-related pricing in different forms. A long-context workload should record stable prefix length, volatile suffix length, prompt template version, measured cache share, and source date. If a prompt template changes, the next forecast should start a new assumption rather than silently inheriting the old cache share.
Treat Output as Its Own Budget
Output is often the line that surprises agent teams. Kimi K3 output is $15.00 per 1M tokens on its current pricing page. DeepSeek and Z.AI also price output separately from input. QwenCloud says thinking tokens count as output tokens. A practical ledger should set max tokens by task class and record generated output separately from prompt input. This makes verbose reasoning, retry loops, and long explanations visible.
Mark Promotions and Expiry Dates
Z.AI lists GLM-5.3-Flash promotional pricing that ends on September 9, 2026 UTC+8. That is useful for a trial started today, but it should not become a silent annual forecast. Store normal price, current promotional price, expiry date, and reviewer note. When the date passes, the worksheet should force a recheck instead of letting a stale promotional assumption survive in procurement copy.
Internal Links for Price Review
Readers doing price review should move through Pricing, Models docs, Chat Completions, Trust, the add-on billing ledger, and the pricing-version drift monitor. Those links connect normalization rules to live gateway evidence.
Procurement Review
Procurement should require source URL, checked date, row owner, currency, unit, input, cached input, output, tool calls, media calls, failed-call policy, promotional expiry, gateway pricing_version, and account group. Engineering should attach one request-level usage object and one calculated workload estimate. Security should confirm that the estimate was produced from redacted or approved data, not copied from a production transcript.
Final Checklist
A Chinese AI API price table is ready for comparison when row owner, source date, unit, cache status, output policy, tool fees, promo expiry, failed-call policy, and gateway group are separate. Recheck direct provider pages and AIWave /api/pricing on publication day. If a number cannot be traced back to a URL and checked date, keep it out of the article and the finance workbook.