Chinese AI model API pricing in 2026: GLM-5.2, DeepSeek, Kimi and MiniMax compared

Here is the short version. As of June 2026, the official list prices per million tokens (input / output, USD) for the major Chinese models are:

For contrast, the closed frontier: Claude Fable 5 runs $10 / $50, Opus 4.8 $5 / $25, GPT-5.5 $5 / $30. The gap is not subtle. A million output tokens costs $50 on Fable 5 and $0.87 on DeepSeek V4-Pro — a 57x difference.

The rest of this guide explains what you get at each price point, when the cheap option is genuinely enough, and how to actually pay for these models from outside China.

What the prices buy you

Price without capability is noise, so anchor on the independent numbers. On the Artificial Analysis Intelligence Index v4.1, the closed leaders score 60 (Claude Fable 5), 56 (Opus 4.8) and 55 (GPT-5.5). GLM-5.2 scores 51 — fourth overall, first among open weights. Then come MiniMax-M3 and DeepSeek V4-Pro at 44 and Kimi K2.6 at 43.

Read those two lists together and the story writes itself: the fourth-smartest model in the world charges about 9% of what the smartest one does for output tokens.

On coding specifically:

• GLM-5.2: 62.1 on SWE-bench Pro (multi-file GitHub issues), first on Design Arena, ahead of GPT-5.5 on several long-horizon coding benchmarks. Details in our GLM-5.2 deep dive.
• DeepSeek V4-Pro: 93.5 on LiveCodeBench — the highest score of any model, closed or open — and a 3206 Codeforces rating.
• Kimi K2.6: 66.7 on Terminal-Bench 2.0 and the strongest tool-use/MCP scores of the group, with a 256K context.

The cost structures that actually matter

Three things move a real bill more than the headline rate:

Prompt caching. GLM-5.2 charges $0.26 for cached input against $1.40 uncached — about a fifth. Kimi K2.6 goes further: $0.16 cached against $0.95. If your agent resends a big stable system prompt every turn, caching is the difference between a scary bill and a boring one. We cover the mechanics in prompt caching explained.

Output-heavy vs input-heavy workloads. DeepSeek V4-Pro's $0.87 output rate makes it uniquely cheap for generation-heavy work (drafting, code output, long answers). GLM-5.2 and Kimi charge $4.00–4.40 for output — still a fraction of closed rates, but a 5x spread inside the Chinese camp.

Routing. The cheapest setup is never one model. Send bulk classification and extraction to V4-Flash at $0.14, default real work to GLM-5.2 or V4-Pro, and reserve a closed frontier model for the rare task that justifies it. With an OpenAI-compatible gateway the router is three lines of code.

How to pay from outside China

The prices above are the providers' mainland list rates, and this is where overseas developers hit the wall: direct sign-up usually wants a Chinese phone number and Alipay or WeChat Pay. You have three realistic options:

1. Self-host the open weights. GLM-5.2 is MIT-licensed on Hugging Face (zai-org/GLM-5.2). Free weights, but you pay in GPUs and ops time — realistic only above a certain scale.
2. Use a Western aggregator that carries some Chinese models. Coverage and model freshness vary.
3. Use a gateway built for exactly this. Turiloop is an OpenAI-compatible API gateway that provides the current generation of Chinese models — GLM-5.2, DeepSeek V4, Kimi, MiniMax — plus selected frontier models, with pay-as-you-go billing and international card payment, no Chinese phone number required. You change one base_url and keep your OpenAI SDK code. Current per-model rates are on the models page.

FAQ

What is the cheapest capable LLM API in 2026? DeepSeek V4-Flash, at roughly $0.14 / $0.28 per million tokens with cache hits near $0.014. For work that needs more reasoning, DeepSeek V4-Pro at $0.435 / $0.87 is the best price-per-capability on the market. See when Flash is enough.

How much does the GLM-5.2 API cost? Z.ai's official rate is $1.40 per million input tokens and $4.40 per million output, with cached input at $0.26. That's roughly a sixth of GPT-5.5's price for a model that beats it on several coding benchmarks.

Are Chinese AI models really cheaper than OpenAI or Anthropic? Yes, by 5x to 50x on list price. Claude Fable 5 charges $50 per million output tokens; GLM-5.2 charges $4.40; DeepSeek V4-Pro charges $0.87. The quality gap is real but much smaller than the price gap — GLM-5.2 sits within 9 points of the best model in the world on the Artificial Analysis index.

Can I use these models without a Chinese phone number? Yes. Either self-host the open weights, or use an OpenAI-compatible gateway like Turiloop that handles the upstream accounts and takes international cards, pay-as-you-go.

*Prices are official provider list rates as of June 29, 2026, and change often — treat this page as a snapshot and the providers' pages as the source of truth.*