Chinese LLM API Pricing Research — 2026-09-12
Token-based API pricing for Chinese frontier models, in USD per million tokens, updated from official docs on September 12, 2026.
Token-based API pricing for Chinese frontier models. All prices in USD per million tokens unless noted. Updated from official provider documentation on September 12, 2026.
Unified pricing comparison
All prices are per 1M tokens. USD is shown first; CNY equivalents use $1 ≈ ¥7.15 for comparison. “Uncached input” means cache miss. DeepSeek and Grok show peak or long-context tiers where applicable. A dash means that the source did not list a separate rate.
| Model | Uncached input | Cached input | Cache write | Output | Context and limits | Source |
|---|---|---|---|---|---|---|
| DeepSeek V4 Pro | $0.66 (¥4.72) off-peak / $1.32 peak | $0.022 (¥0.16) off-peak / $0.044 peak | — | $1.98 (¥14.16) off-peak / $3.96 peak | 1M; max output 384K; concurrency 500 | official pricing |
| DeepSeek V4.1 Flash (V4 Flash alias) | $0.15 (¥1.07) off-peak / $0.30 peak | $0.003 (¥0.02) off-peak / $0.006 peak | — | $0.60 (¥4.29) off-peak / $1.20 peak | 1M; max output 384K; concurrency 2,500; vision | official pricing |
| Hy4 preview (Tencent Hunyuan) | $0.834 (≈¥5.96) | $0.042 (≈¥0.30) | — | $2.501 (≈¥17.88) | 1,048,576; max output 64K | official announcement and pricing |
| GLM-5.3-Flash (Zhipu AI) | $0.15 (¥1.07) | $0.03 (¥0.21); storage limited-time free | — | $0.50 (¥3.58) | 1M; native multimodal input | official pricing |
| Muse Spark 1.3 (Meta) | $1.25 (≈¥8.94) | $0.15–$1.00 (≈¥1.07–¥7.15) | — | $4.25 (≈¥30.39) | 1M; standard tier; $20 credits for new accounts | Meta Model API |
| GPT-5.6 Luna (OpenAI) | $1.00 (≈¥7.15) | $0.10 (≈¥0.72) | $1.25 (≈¥8.94) | $6.00 (≈¥42.90) | 256K | official pricing |
| Grok 4.6 (xAI) | $2.00 (<200K); $4.00 (≥200K) | $0.50 (<200K); $1.00 (≥200K) | — | $6.00 (<200K); $12.00 (≥200K) | 500K; ≥200K prompts billed at 2x | official pricing |
| Gemini 3.8 Flash (Google) | $0.75 (≈¥5.36) through 2026-12-31 | $0.075 (≈¥0.54) through 2026-12-31 | — | $3.75 (≈¥26.81) through 2026-12-31 | 1M+; text, image, video, audio input | official pricing |
| Kimi K3 (Moonshot AI) | $2.80 (¥20.00) | $0.28 (¥2.00) | — | $14.00 (¥100.00) | 1,048,576 (1M) | official pricing |
| GPT-5.6 Terra (OpenAI) | $2.50 (≈¥17.88) | $0.25 (≈¥1.79) | $3.125 (≈¥22.34) | $15.00 (≈¥107.25) | 256K | official pricing |
Price competitiveness: GLM-5.3-Flash and DeepSeek V4.1 Flash tie for the lowest off-peak uncached input price at $0.15 per 1M tokens. GLM-5.3-Flash has the lowest output price at $0.50; DeepSeek V4.1 Flash is next at $0.60 off-peak, while its peak rate doubles. DeepSeek’s cache-hit price remains unusually low at $0.003 off-peak.
Model notes
DeepSeek V4 Pro
- Supports thinking and non-thinking modes.
- Compatible with OpenAI and Anthropic API formats.
- Features JSON output, tool calls, chat prefix completion (beta), and FIM completion (beta, non-thinking only).
- Current API version:
DeepSeek-V4-Pro-0813; off-peak pricing is $0.66 input, $0.022 cached input, and $1.98 output per 1M tokens.
DeepSeek V4.1 Flash (V4 Flash alias)
- The current official model version behind the Flash API is
DeepSeek-V4.1-Flash; use model IDdeepseek-flash. - Legacy IDs
deepseek-v4-flashanddeepseek-v4-flash-vision-expremain accepted but are retired aliases routed to V4.1 Flash. - Supports thinking and non-thinking modes, vision, JSON output, tool calls, Responses API, and Anthropic-format requests.
- Off-peak pricing is $0.15 input, $0.003 cached input, and $0.60 output per 1M tokens; weekday peak pricing is double.
Hy4 preview
- Tencent’s early Hunyuan 4 preview is available through Tencent Cloud and OpenRouter.
- Model ID:
hy4-preview.
Kimi K3
- Flagship model, always reasoning with configurable
reasoning_effort(low/high/max, default max). - Supports automatic context caching, tool calls, JSON mode, and structured output.
- Web search (
web_search) is being upgraded and is not recommended at this time. - Model ID:
kimi-k3.
GLM-5.3-Flash
- Current cost-focused model in Z.AI’s catalog; the standard GLM-5.3 model is also listed separately at a higher rate.
- Supports native multimodal input with a 1M-token context; cached-input storage is currently listed as limited-time free.
- Model ID:
glm-5.3-flash.
Grok 4.6
- Batch API discount is not listed for Grok 4.6.
- Priority processing is available at 2× standard rates.
- Model ID:
grok-4.6. - Prompts at or above 200K tokens use the long-context tier; cached input is $0.50 below 200K and $1.00 at or above 200K.
GPT-5.6 Luna and Terra
- Batch pricing is 50% off for both models; Flex processing is available at batch rates.
- Luna is the budget tier; Terra is the mid-tier of the GPT-5.6 family.
- Model IDs:
gpt-5.6-lunaandgpt-5.6-terra.
Gemini 3.8 Flash
- Multimodal input—text, image, video, and audio—is billed at the same input rate.
- Output pricing includes both response and reasoning tokens.
- The introductory $0.75 input / $3.75 output rates are listed through December 31, 2026; the standard rates then double.
- Model ID:
gemini-3.8-flash.
Muse Spark 1.3
- Multiple cache tiers are available for discounted input rates.
- New developer accounts receive $20 in free testing credits.
- The standard tier remains $1.25 input and $4.25 output per 1M tokens; a separate contributor tier is not used for this comparison.
- Model ID:
muse-spark-1.3.
Notes on CNY → USD conversion
Kimi K3 lists pricing in CNY; the other models in this comparison publish USD pricing or are represented using their USD list price. Approximate conversion used: $1 ≈ ¥7.15. Actual USD billing may vary based on each provider’s payment processor. DeepSeek’s peak/off-peak schedule is UTC Monday-Friday: 01:00–04:00 and 06:00–10:00 peak; all other hours are off-peak.