Independent notes Updated: 2026-08-13

Chinese frontier LLMs are powerful — and surprisingly inexpensive

Outside mainland China, it can be hard to track which Chinese models are worth testing, where to find official prices, and how they compare with familiar regional cloud AI services. This independent page maps DeepSeek, GLM and Kimi, with official links, pricing snapshots, and a practical referral access option for a broader model catalog.

Explore the models View access option

Independent personal page. Not an official provider website.

Model Market Snapshot
DeepSeek GLM Kimi
Official price snapshot
DeepSeek V4-Flash-0731~$0.14 / 1M in
GLM-5.2$1.40 / 1M in
Kimi K3$3.00 / 1M in
Official links included
Pricing snapshots
Developer-focused model map
Transparent referral disclosure

The Chinese LLM landscape at a glance

Chinese LLMs have become highly competitive in reasoning, coding, long-context work, and agentic workflows. Their official per-token pricing can be unusually aggressive compared with many familiar cloud AI services.

For overseas developers, the issue is not only model quality; it is discovering, comparing, and accessing the right services. This page is a map, not a ranking.

Use this page as a starting point, not as financial or technical advice. All models, prices, and availability should be verified directly with the official providers and the referral service.

Three Chinese LLM families worth watching

DeepSeek

DeepSeek

DeepSeek V4-Pro-0813 · DeepSeek V4-Flash-0731

Strong reasoning and coding models with excellent long-context capabilities and agentic workflow support. Official token pricing is among the most aggressive in the industry.

  • Advanced reasoning
  • Coding & agentic workflows
  • Long-context experiments
  • Very low official token pricing
Official API pricing Official DeepSeek
Z.AI GLM

GLM by Z.AI

GLM-5.2

Positioned for long-horizon tasks and project-scale engineering. Officially described as designed for reasoning, agentic engineering, and coding workflows.

  • Long-horizon tasks
  • Project-scale engineering context
  • Agentic engineering
  • Coding workflows
Official pricing GLM-5.2 docs
Kimi

Kimi by Moonshot AI

Kimi K3

Kimi's flagship model for long-horizon coding and end-to-end knowledge work. Native multimodal understanding with a 1M-token context window and industry-leading intelligence.

  • Flagship for long-horizon coding
  • End-to-end knowledge work
  • Native multimodal understanding
  • 1M-token context · always-reasoning
Official pricing Kimi K3

Practical capability map

Reasoning

All three model families show strong reasoning performance. DeepSeek is widely discussed for chain-of-thought tasks; GLM-5.2 is positioned for deep analytical work; Kimi K3 always reasons, with configurable reasoning effort for long-horizon work.

Coding

DeepSeek V4-Pro-0813, GLM-5.2, and Kimi K3 are all positioned for coding and software engineering. Kimi K3 is a flagship model for long-horizon coding and agentic engineering workflows. Use cases include code generation, debugging, and multi-file refactoring.

Long context

DeepSeek models support long-context experiments. Kimi K3 features a 1M-token context window, suitable for long-horizon software engineering and knowledge work. GLM-5.2 is positioned for project-scale engineering context.

Agent workflows

DeepSeek and Kimi both emphasize agentic workflow support. Kimi K3 is built for long-horizon agent tasks with tool-use and reasoning-effort controls. GLM-5.2 is described as designed for agentic engineering. These models can be used as part of tool-use, planning, and multi-step automation pipelines.

Cost-sensitive prototyping

DeepSeek V4-Flash-0731 at $0.14/1M input (cache miss) enables cost-sensitive prototyping at scale. Kimi K3 with cache hit at $0.30/1M offers another path for iterative development. These price points can change how developers budget AI experiments.

Multilingual readiness

Chinese frontier models are generally strong in Chinese and English. For Russian-language use cases, developers should test specific Russian-language performance with their own prompts and tasks.

Model capabilities and benchmarks change quickly. Use this page as a map, not as a final ranking.

Official price snapshots: Chinese LLMs can be dramatically inexpensive

For developers used to regional cloud AI pricing, the public token prices of some Chinese frontier models can look unusually aggressive.

Model Input / 1M Cached input / 1M Output / 1M Source Notes
DeepSeek V4-Flash-0731 $0.14 (cache miss)
$0.0028 (cache hit)		 $0.0028 $0.28 Official DeepSeek API pricing Cache hit/miss pricing model
DeepSeek V4-Pro-0813 $0.435 (cache miss)
$0.003625 (cache hit)		 $0.003625 $0.87 Official DeepSeek API pricing Cache hit/miss pricing model
GLM-5.2 $1.40 $0.26 $4.40 Official Z.AI developer pricing Input / cached input / output model
Kimi K3 $3.00 (cache miss)
$0.30 (cache hit)		 $0.30 $15.00 Official Kimi API pricing Cache hit/miss pricing model
YandexGPT Pro 5.1 Reference ~$6.557 ~$6.557 ~$6.557 Official Yandex Cloud AI Studio pricing Approx. converted from official per-1,000-token synchronous-mode pricing

These are official public pricing snapshots from the model/platform providers. They are not the prices of the referral service. Official pricing can change.

YandexGPT Pro 5.1 is included only as a regional pricing reference point, not as a strict quality-equivalent comparison.

DeepSeek official pricing Z.AI official pricing Kimi official pricing YandexGPT pricing reference

Why access can still be fragmented

Despite low official token pricing, developers outside mainland China may still prefer a third-party token gateway for practical reasons:

Multiple model vendors
Different accounts
Different billing systems
Different documentation styles
Difficult to compare several model families quickly
Some prefer one balance and a broader model catalog
Some developers also want non-Chinese models in the same catalog

A practical access option

If you want to explore a broader model catalog through one third-party token gateway, my referral link points to a service I personally track. According to the provider page, some token prices may be around 4% lower than the corresponding official Chinese API pricing, and the catalog may include additional model families such as Claude Fable, GPT-5.5, Seedance 2.0 and more. Please verify the current model list, pricing, billing rules and terms on the provider page before topping up.

→ More models available in catalog ···

Catalog examples may change. Check the provider page before topping up.

Open referral access page Check current model catalog
Affiliate Disclosure

This page contains referral/affiliate links. I may receive a bonus or commission if you register or top up through them. I am not affiliated with DeepSeek, Z.AI, Moonshot AI, Kimi, Yandex, or the third-party gateway unless explicitly stated. Always verify pricing, model availability, billing rules, data handling terms and service terms on the official/provider pages.

FAQ

No. This is an independent personal page. It is not affiliated with or endorsed by DeepSeek, Z.AI, Moonshot AI, Kimi, Yandex, or the third-party gateway referenced on this page.

No. The prices shown in the pricing section are official public pricing snapshots from the model/platform providers (DeepSeek, Z.AI, Kimi, Yandex). They are not the prices of the referral service. The referral service may have its own pricing, billing rules, rate limits, and terms.

YandexGPT Pro 5.1 is included only as a regional pricing reference point — it provides a familiar price anchor for developers who work with regional cloud AI services. It is not presented as a quality-equivalent comparison with the Chinese models listed.

No. The "may be around 4% lower" claim is based on what the provider page states. This may change at any time. Always verify the current pricing on the provider page before topping up. Do not rely on this page as a guarantee of savings.

No. Claude Fable, GPT-5.5, and Seedance 2.0 are not Chinese models. They are listed as examples of additional model families that may be available through the referral service catalog, alongside the Chinese frontier models. They are presented here only as catalog examples.

This is a practical choice, not a recommendation. Some developers prefer one unified account, one balance, and a broader model catalog that includes both Chinese and non-Chinese models. Others may prefer to go directly to each official API. Both approaches are valid. Evaluate based on your own needs.

Production use of any third-party service requires careful evaluation. You should verify the provider's terms of service, data handling policies, privacy practices, billing rules, rate limits, and reliability guarantees before using it for production workloads. This page is not production advice.

Before topping up through any referral link, verify: (1) the current model list on the provider page, (2) the current pricing for each model, (3) billing rules and minimum top-up amounts, (4) rate limits and quota policies, (5) data handling and privacy terms, (6) the provider's terms of service and refund policy. Do not rely solely on this page.