Chinese AI Price War 2026: Real API Cost Across Ten Providers

Verification date: September 20, 2026. Token prices, promotions and peak/off-peak schedules change often.

1. The ten providers and their current models

Budget-tier models from the same providers: DeepSeek deepseek-flash (V4.1-Flash): $0.15 / $0.003 / $0.60 off-peak, $0.30 / $0.006 / $1.20 peak. GLM-5.3-Flash: $0.15 / $0.03 / $0.50. MiMo-V2.5: $0.14 / $0.0028 / $0.28. Step 3.5 Flash: $0.10 / $0.02 / $0.30.

Pricing caveats:

• Kimi K3. The Kimi Open Platform's mainland list is ¥20 cache-miss input, ¥2 cache-hit input and ¥100 output per 1M tokens, with a 1,048,576-token context. The international dollar rate above is a separate listing. At an assumed ¥7.2 per US$1 it is about 8% above the mainland equivalent; keep the two price lists separate.
• Qwen. qwen3.8-max is listed at $2.00 input and $6.00 output. Published reports agree, and one reports a flat rate across the full 1M context. Alibaba's cache documentation says the cache-hit price is not a fixed percentage of input and points to the console, so the $0.25 cache figure should be treated as a reported rate rather than a published fixed formula.
• Tencent. Hy4 preview's identity and specs (770B total, 49B active, open weights) are documented. Public pricing evidence is less direct than for the other models in this table.
• Baidu. ERNIE 5.1's dollar prices are not confirmed. The mainland yuan list applies to mainland accounts, and one review notes Baidu's international catalog lists ERNIE 5.0, not 5.1. We do not convert it into a USD price.
• MiniMax. The page shows M3 with a "permanent 50% off" tag. The $0.30 / $1.20 above is the discounted rate; the struck-through list rate is $0.60 / $2.40.

The two unresolved cells:

• qwen3.8-flash: Alibaba's model page confirms it is current, has a 1M-token context and supports context caching. Its exact input, output and cache rates are not published in the pricing material cited here. Third-party trackers disagree, so it is excluded from the comparison.
• BytePlus international Doubao/Seed: Official documentation confirms the international service, its Singapore and EU regions, and the current catalogue, but the numeric international pay-as-you-go table is not available in the cited material. Volcano Engine's mainland rates are a different price list and are not used here.

3. The caching war

Every provider here discounts cached input, but by very different amounts (cache-hit price as a share of the normal input price, CALCULATION):

Three consequences:

1. A high hit rate rewards a small ratio. In workload D, caching cuts Xiaomi's MiMo-V2.5-Pro bill by 57% ($60.90 to $26.39) and DeepSeek V4-Pro's by 48%. It cuts Step 3.7 Flash's by 30%, because its 20% ratio saves less and most of its bill is output.
2. Rankings flip. On workload D, MiMo-V2.5-Pro ($26.39) beats Step 3.7 Flash ($30.20), DeepSeek's peak-priced Flash ($30.48) and MiniMax-M3 ($34.80), though it costs more on uncached input than three of them.
3. Output still dominates. In workload D on MiMo-V2.5-Pro, $17.40 of the $26.39 is output, and caching cannot discount it. Kimi K3's output rate ($15) is 17× Xiaomi's flagship rate ($0.87).

Cache hits are not guaranteed anywhere. StepFun, for instance, says it cannot force every request to hit, and Alibaba's cache price is not published as a simple percentage. Measure your own hit rate before budgeting on a discount.

4. The clock, the context and the region

DeepSeek's peak and off-peak. Its Flash and Pro rates are doubled at peak. Peak hours are 01:00–04:00 and 06:00–10:00 UTC (09:00–12:00 and 14:00–18:00 Beijing time), Monday to Friday, excluding Chinese public holidays; everything else is off-peak. That is 7 peak hours on each weekday, or 35 of the 168 hours in a week (about 20.8%). Off-peak is therefore 133 of 168 weekly hours (about 79.2%): 17 hours on each weekday and all 24 hours on weekends and Chinese public holidays, so weeks with a public holiday have fewer peak hours. If your traffic were spread evenly across the week (a MODELING ASSUMPTION), your blended price would be about 1.21× the off-peak rate. Workload C on V4.1-Flash, uncached: $120 off-peak, $240 all-peak, about $145 evenly spread. On V4-Pro: $462, $924 and about $558. Scheduling batch work off-peak halves the bill.

MiniMax's long-context tier. Above 512K input tokens, MiniMax-M3 charges $0.60 input, $0.12 cache read and $2.40 output, twice the shorter tier. Workload C would cost $480, not $240. Independent reporting describes Alibaba's Qwen3.8-Max rate as flat across its 1M context; check the Alibaba console for the current account-specific price.

Region matters. Prices differ by route, and we keep them separate:

• Kimi's mainland list of ¥20 / ¥2 / ¥100 is not a global USD price. At an illustrative ¥7.2 per US$1 (MODELING ASSUMPTION), it is about $2.78 / $0.28 / $13.89, versus $3.00 / $0.30 / $15.00 internationally.
• Alibaba's international (Singapore) endpoint is priced differently from Beijing, which one secondary source reports as 60–70% cheaper for some models. Mainland endpoints can carry different account, data-residency and access conditions.
• Baidu's ERNIE 5.1 is listed only in yuan for mainland accounts. We do not turn it into an official dollar price.

5. Cheap per token is not cheap per task

Token prices compare inputs and outputs. They do not compare how many tokens each model needs to do the same work. We make no claim about which model is better, but four things can move your real bill:

• Reasoning tokens. Reasoning models produce thinking tokens before the answer. Z.ai states no separate reasoning surcharge (a secondary report of its pricing page), which means thinking is billed at the output rate. Check each provider's rule.
• Effort settings. GLM-5.3 is reported as reasoning-always-on with a default of maximum effort. StepFun's Step 3.7 Flash offers low, medium and high; DeepSeek's V4 models support thinking and non-thinking modes.
• Vendor token-efficiency claims. Xiaomi says its flagship uses roughly 40–60% fewer tokens per agent task than some peers, and StepFun's own model card says Step 3.5 Flash currently needs longer trajectories than Gemini 3.0 Pro. Both are vendor statements. Test on your own tasks.
• Host prices are not vendor prices. Third-party hosts can quote rates far from the vendor's list. OpenRouter's headline price for GLM 5.2 (about $0.49 input) was roughly a third of Z.ai's own $1.40, reflecting a cheaper third-party host, and host serving configurations can differ. We use vendor lists except where labeled.