Cheapest Chinese LLM APIs in 2026
The cheapest Chinese LLM APIs in August 2026: Qwen3.5 Flash at $0.029/1M input tokens (lowest input bar none), DeepSeek V4 Flash at $0.22/$0.66 off-peak (cheapest all-round — it undercuts even the cheapest US model), and GLM-5.3 at $1.40/$4.40 for frontier coding at a quarter of US flagship prices.
Verified price table (August 2026)
All prices are official list prices per 1M tokens (input/output), cross-checked on August 23, 2026 against provider docs and aggregators. Sorted by input price.
| Model | Provider | Input $/1M | Output $/1M | Best for |
|---|---|---|---|---|
| DeepSeek V4 Flash | DeepSeek | 0.22* | 0.66* | Cheapest all-round: classification, extraction, high-volume chat (*off-peak; peak 0.44/1.32) |
| GLM-5.3-Flash | Zhipu / Z.ai | 0.15 | 0.50 | Cheapest frontier-family tier — promo $0.075/$0.25 to Sep 9 |
| GPT-5.6 Luna (US baseline) | OpenAI | 0.20 | 1.20 | Cheapest US option — for comparison |
| Kimi K2.5 | Moonshot AI | 0.45 | 2.25 | Agent loops, visual coding, 262K context |
| MiniMax M3 | MiniMax | 0.60 | 2.40 | Cheapest >80% SWE-bench scorer — best coding value |
| GLM-5.3 / GLM-5.2 | Zhipu / Z.ai | 1.40 | 4.40 | Frontier coding + 1M context, cheapest in its tier |
| Qwen3.8-Max | Alibaba | 2.00 | 6.00 | Flagship "cowork" tasks, multilingual |
| Kimi K3 | Moonshot AI | 3.00 | 15.00 | Premium 2.8T multimodal flagship, 1M context |
Sources: Z.ai pricing docs (GLM) · Alibaba Model Studio (Qwen3.8-Max) · OpenRouter (Kimi K2.5) · MorphLLM 12-API comparison (V4 Flash, MiniMax M3) · Apidog price-war analysis (V4-Pro, MiMo) · LLM Abacus (Qwen3.5 Flash) · BenchLM (Kimi K3) · Live JSON dataset on GitHub. Prices move weekly — re-verify before large commits.
Single-number champions
- Cheapest input of any LLM API: Qwen3.5 Flash — $0.029/1M input tokens (¥0.20), per LLM Abacus. Ideal for classification and extraction pipelines.
- Cheapest frontier output: DeepSeek V4-Pro — $0.87/1M output tokens, the lowest among frontier-class models per Apidog's price-war analysis.
- Flat-rate long context: Xiaomi MiMo V2.5 Pro — $3/1M output flat at 1M-token context, when you need huge windows without tiered pricing.
Cost levers that matter more than sticker price
- Prompt caching / shorter system prompts — GLM-5.3 cached input drops to $0.26/1M (from $1.40); Kimi K3 cache hits to $0.30 (from $3.00). Often saves more than hopping vendors.
- Route hard tasks only — use Kimi K3 / R1-class for hard steps; V4 Flash / Qwen Flash for 90% of calls.
- Chunk + retrieve — smaller contexts beat larger models for many RAG jobs.
- Batch non-interactive jobs — offline summarization is more elastic; DeepSeek weekend off-peak rates apply on some platforms.
Practical stack (recommended)
High-volume chat/RAG: deepseek-v4-flash or qwen-flash tiers — sub-GPT-Luna cost with better multilingual quality
Agents/coding: kimi-k2.5 or minimax-m3, escalate to glm-5.2/5.3 for frontier steps
Video gen (per-second pricing): seedance-2.5 (~$0.10/s 480p) / happyhorse-1.1 ($0.0988/s 720p)
Billing: USDT top-up, single OpenAI-compatible key via ChinaModelAPI
FAQ
What is the cheapest Chinese LLM API in 2026?
By input price, Alibaba's Qwen3.5 Flash is the cheapest Chinese LLM API at $0.029 per million input tokens. For balanced pricing, DeepSeek V4 Flash ($0.22 in / $0.66 out per 1M tokens off-peak; peak $0.44/$1.32; restructured 2026-08-23) is the cheapest all-round option — it undercuts even the cheapest US model, GPT-5.6 Luna ($0.20/$1.20), on both axes.
How much cheaper are Chinese LLM APIs than GPT?
Compared with the cheapest US baseline (GPT-5.6 Luna at $0.20/$1.20 per 1M tokens), DeepSeek V4 Flash is ~30% cheaper on input and ~77% cheaper on output. Against full US flagships the gap widens to 5-10x: GLM-5.3 delivers frontier-tier coding at $1.40/$4.40.
Which Chinese model is cheapest for coding and agents?
Moonshot Kimi K2.5 at $0.45/$2.25 per 1M tokens is the value pick for agent loops. MiniMax M3 ($0.60/$2.40) is the cheapest model scoring above 80% on SWE-bench. For heavy frontier coding, DeepSeek V4-Pro offers the cheapest output tokens ($0.87/1M) among frontier models.
Is cheaper always worse quality?
No. For many RAG and support workloads, Chinese frontier open-weight models match or beat mid-tier Western models at a fraction of the cost. Always validate on a 20-50 prompt eval set in your own domain before switching.
Are these official prices or relay prices?
The table lists official list prices from providers (Z.ai for GLM-5.3, Alibaba Model Studio for Qwen3.8-Max, Moonshot for Kimi) cross-checked with aggregators (OpenRouter, MorphLLM, LLM Abacus) as of August 23, 2026. Via a relay like ChinaModelAPI you get one OpenAI-compatible key for all of them, with USDT top-up.
How do I migrate from OpenAI to cheaper Chinese models?
Keep your OpenAI SDK. Change base_url to https://api.chinamodelapi.com/v1, swap the API key, and use model ids like glm-5.2, deepseek-v4-pro or qwen-plus. No code rewrite needed. See the qwen-api-guide and deepseek-v3-api for step-by-step code.
Related Guides
One OpenAI-compatible key. Chinese models. USDT top-up.
Get Early Access