This is a price snapshot, not a model benchmark. I compared four Chinese model endpoints available through the same API catalog so that the billing units and date are consistent. The result is less about finding one winner than identifying three distinct workload tiers:
- Flagship capability and very long context: Kimi K3.
- A middle tier for capability, context, and price: GLM-5.2 during the catalog’s discount window.
- Low-cost generation and high-volume processing: DeepSeek V4-Pro and DeepSeek V4-Flash-0731.
All prices below are the USD rates displayed by AIHubMix on August 3, 2026, per one million input or output tokens. These are aggregator rates rather than the model vendors’ direct list prices. Routing, upstream providers, and promotions may change them.
Scope note: AIHubMix is the data source because this article compares what one customer would pay through one catalog. This is not an endorsement of the platform, and the article does not claim that token price predicts model quality. Recheck live prices before making a production budget.
| Model | AIHubMix input / output | Cached input | Workload to test |
|---|---|---|---|
| Kimi K3 | $3.00 / $15.00 | $0.30 | 1M context, long-horizon coding, and complex knowledge work |
| GLM-5.2, 30% discounted | $0.79 / $2.76 | $0.20 | Long-context coding, agents, and overall value |
| DeepSeek V4-Pro | $0.46 / $0.93 | About $0.00385, route-specific | Output-heavy reasoning, coding, and content generation |
| DeepSeek V4-Flash-0731 | $0.099 / $0.198 | $0.0198 | Classification, extraction, batch jobs, and high concurrency |
Note: The DeepSeek V4-Pro summary rounds cached input to $0.00. Its DeepSeek upstream route shows a cache-read price of approximately $0.0038512 per million tokens. Check the route used by your request when reconciling a bill.
For comparison, AIHubMix currently lists Claude Opus 5 at $5 / $25, GPT-5.6 Sol at $5 / $30, GPT-5.6 Terra at $2 / $12, and GPT-5.6 Luna at $0.20 / $1.20.
On output price alone, discounted GLM-5.2 costs about one-ninth as much as Claude Opus 5. DeepSeek V4-Pro is about one twenty-seventh, while DeepSeek V4-Flash-0731 is about one one-hundred-and-twenty-sixth. Those gaps are substantial, but price alone does not capture task success rate, context requirements, output volume, or latency.
How to Pay From Outside China
The token prices are only half the problem. Developers outside mainland China can run into account and payment friction when a provider expects a local phone number, Alipay, or WeChat Pay. There are three practical approaches.
1. Self-Host Open Weights
GLM-5.2 is published on Hugging Face under the MIT license. The weights do not carry an API fee, but inference is not free: GPUs, serving infrastructure, monitoring, upgrades, and engineering time all move onto your side of the ledger. Self-hosting becomes economically interesting only when utilization and control requirements justify that operational cost.
2. Use an Aggregator Available in Your Region
Several model aggregators expose at least some Chinese models and accept payment methods that are easier to use internationally. The trade-off is inconsistent coverage. A service may list one DeepSeek release but not the newest GLM or Kimi endpoint, and model IDs, context limits, caching, and tool support can differ.
3. Use an OpenAI-Compatible Gateway With Card Billing
AIHubMix is one example of this approach. Its console currently offers both Alipay and Credit/Debit Card top-ups, while the API uses an OpenAI-compatible base URL. Card acceptance can still depend on the issuer and region, so verify it before designing a production billing process.
The engineering benefit is modest but useful: one account, one balance, and one SDK integration for the model IDs in this comparison. The trade-off is an additional dependency between your application and the original model provider. Check the gateway’s status history, data policy, route behavior, and effective price rather than treating API compatibility as equivalence.
What These Prices Actually Buy
GLM-5.2: How the Discount Changes the Middle Tier
AIHubMix displays regular GLM-5.2 prices of $1.13 for input, $3.94 for output, and $0.28 for cached input. The current 30% discounted rates are $0.79, $2.76, and $0.20 respectively. The page labels the discount window as 09:00–08:59 GMT+9; individual upstream routes may show different windows, so production budgets should use the live rate.
GLM-5.2 offers a 1M-token context window, up to 128K output, reasoning, tool use, function calling, and structured outputs. The listed output rate stays below $3 per million tokens while the endpoint supports repository-scale code, long documents, and multi-step agent workflows.
If your application repeatedly reads an entire codebase, a large policy document, or a long conversation history, GLM-5.2 is one candidate to include in the default-model evaluation.
DeepSeek V4-Pro: Sub-$1 Output for Heavier Work
AIHubMix lists DeepSeek V4-Pro at $0.46 for input and $0.93 for output. It also provides a 1M-token context window, up to 384K output, reasoning, tools, function calling, and structured outputs.
That price structure is relevant when output is much larger than input: code generation, bulk drafting, long answers, and first-pass reports. V4-Pro can be tested against Flash for tasks that need heavier reasoning and longer generation. Its listed output rate is lower than GLM-5.2's.
DeepSeek V4-Flash-0731: The Volume Tier
DeepSeek V4-Flash-0731 costs $0.099 for input, $0.198 for output, and $0.0198 for cached input on AIHubMix. It is not the obvious choice for the hardest final decision, but it is well suited to large quantities of cheap, standardized work:
- Text classification and tagging
- Field extraction and format conversion
- First-pass search-result filtering
- Log summarization and ticket routing
- High-volume short-form generation
Sending every request directly to a flagship model means paying a premium for tasks that may not need flagship capability. Flash can be tested at the first layer of a model-routing system.
At 100 million output tokens, the page rates imply approximately $19.80 for DeepSeek V4-Flash-0731, $93 for DeepSeek V4-Pro, $276 for discounted GLM-5.2, $1,500 for Kimi K3, and $2,500 for Claude Opus 5. This excludes caching, failed retries, and tokens consumed around tool calls, but it illustrates why batch work deserves a dedicated low-cost tier.
Kimi K3: The Higher-Cost Long-Context Option
Kimi K3 is the most expensive Chinese model in this comparison: $3 for input, $15 for output, and $0.30 for cached input. AIHubMix shows a 1M context window and up to 1M output, with text, image, and video input. Its positioning is clear: long-horizon coding and end-to-end knowledge work.
What Actually Determines the Bill
1. Price GLM-5.2 Using the Discount Window
GLM-5.2’s regular and discounted output rates are $3.94 and $2.76. At 100 million output tokens, the difference is about $118. At meaningful scale, whether requests hit the discounted route changes the monthly budget. Record the route and effective unit price in monitoring rather than treating a promotional rate as permanent.
2. Caching Only Helps Repeated, Stable Context
Caching becomes valuable when an agent resends the same system prompt, tool definitions, project rules, or reference material on every turn. DeepSeek V4-Flash-0731 lists cached input at $0.0198, discounted GLM-5.2 at $0.20, and Kimi K3 at $0.30. If every prompt is different, low cache prices will not automatically reduce the bill.
3. Output-Heavy Workloads Need Output-First Pricing
Summarization and classification often have large inputs and short outputs. Code generation, long-form writing, and deep reasoning tend to produce more output. Optimize the former around input and cache rates; optimize the latter around output rates. Adding the two headline prices together is rarely a useful ranking method.
4. The Cheapest System Usually Uses More Than One Model
A practical three-layer route looks like this:
- Send classification, extraction, and simple batch jobs to DeepSeek V4-Flash-0731.
- Use GLM-5.2 or DeepSeek V4-Pro for most production work.
- Reserve Kimi K3 or another high-priced frontier model for the small set of high-risk, very long, or unusually difficult tasks.
This routing policy can be implemented in the application based on task type, context length, and retry count. Measure fallback frequency because a cheap first attempt can become expensive when it repeatedly fails.
Calling the Models Through One Gateway
AIHubMix exposes an OpenAI-compatible API. An application already using the OpenAI SDK typically needs only a different base_url, API key, and model ID:
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["AIHUBMIX_API_KEY"],
base_url="https://aihubmix.com/v1",
)
response = client.chat.completions.create(
model="glm-5.2",
messages=[
{
"role": "user",
"content": "Review this project's architecture and propose a migration plan.",
}
],
)
print(response.choices[0].message.content)
Change model to deepseek-v4-pro, deepseek-v4-flash-0731, or kimi-k3 to switch models. Test streaming, timeouts, tool calls, and structured outputs separately before production use. API compatibility does not mean every model behaves identically.
A Reasonable Evaluation Setup
For a simple starting configuration:
- Default-model candidate: GLM-5.2. Discounted pricing and a 1M context window make it worth testing for code and long tasks.
- Output-heavy production model: DeepSeek V4-Pro. Output costs less than $1 per million tokens.
- Batch model: DeepSeek V4-Flash-0731. Push low-risk, high-concurrency work to the lowest-cost tier.
- Long-horizon candidate: Kimi K3. Test it when 1M context, multimodal input, and unusually long output are genuinely useful.
Do not choose from a few subjective chat responses. Build an evaluation set of 50–200 real tasks and measure success rate, human rework, average input and output tokens, P95 latency, and retry count. “Cost per successful task” is usually more informative than price per million tokens.
FAQ
What Is the Cheapest Model in This Comparison?
DeepSeek V4-Flash-0731 has the lowest input and output rates at $0.099 / $0.198. It is appropriate for batch and standardized tasks. Test DeepSeek V4-Pro, GLM-5.2, or Kimi K3 for complex reasoning and long-horizon coding.
How Much Does GLM-5.2 Cost on AIHubMix?
The page shows regular rates of $1.13 input, $3.94 output, and $0.28 cached input. The current 30% discounted rates are $0.79 input, $2.76 output, and $0.20 cached input, all per million tokens. Discount windows and routes may change.
Why Does the DeepSeek V4-Pro Price Differ From Other Sites?
This article uses the current price of AIHubMix’s aggregated deepseek-v4-pro model ID. AIHubMix exposes several upstream routes at different prices, and its model page says the aggregated ID uses the official discounted channel. Always compare the same model ID and actual route.
Are Chinese Models Always Better Value Than OpenAI or Anthropic Models?
Their token prices are often much lower, but the real metric is cost per completed task. Extra retries, manual corrections, or unnecessarily long output can reduce the apparent savings. Evaluate with your own workload.
Will These Prices Stay the Same?
No. Platforms change promotions, providers, routes, and caching rates. This article is a snapshot from August 3, 2026. Use the live AIHubMix model pricing page for production budgets and customer quotes.
Top comments (0)