DeepSeek, Qwen, Kimi, GLM: What Chinese Open-Weight Models Cost, and the 3 Ways to Pay
- il y a 11 heures
- 4 min de lecture
The 4 model families coming out of China that matter for enterprise workloads in 2026 are DeepSeek, Qwen (Alibaba), Kimi (Moonshot) and GLM (Z.ai, formerly Zhipu). All 4 publish their weights, most under MIT or Apache licenses, so anyone can download and run them. That word “open” produces a recurring budgeting error: open weight is not free. It means you choose where to pay. There are 3 ways: the vendor's own API, a third-party host, or your own GPUs. This post prices all 3, following the same structure as our Claude and OpenAI pricing posts. All figures are as of August 23, 2026.
Way 1: the vendor's hosted API
Each vendor sells its models per token through its own platform, the same billing model as OpenAI or Anthropic, and in most cases at a fraction of the price. One exception sits in the table: Kimi K3, at $3/$15, costs more than Claude Sonnet 5.
Model | Input ($/MTok) | Output ($/MTok) | Notes |
|---|---|---|---|
DeepSeek V4 Flash | $0.22 off-peak / $0.44 peak | $0.66 off-peak / $1.32 peak | 1M context; time-based pricing since August 16 |
Qwen3.5 Plus | $0.40 | $2.40 | Steps to $0.50/$3.00 above 256K input |
DeepSeek V4 Pro | $0.66 off-peak / $1.32 peak | $1.98 off-peak / $3.96 peak | 1M context; same peak windows |
GLM 5.2 | $1.40 | $4.40 | Cached input at $0.26 |
Kimi K3 | $3.00 | $15.00 | 1M context; cache hits at $0.30 |
gpt-5.6-terra (closed source, for reference) | $1.00 | $6.00 | From our OpenAI post |
Claude Sonnet 5 (closed source, for reference) | $2.00 | $10.00 | Introductory pricing made permanent in August 2026 |
A calibration note: we verified these prices on the vendors' own platforms on August 16 and again on August 23 (api-docs.deepseek.com, platform.kimi.ai, docs.z.ai; Qwen from a tracker dated August 10). 2 numbers had moved between drafting and verification: DeepSeek switched to time-based pricing on August 16, and Moonshot's current flagship is Kimi K3 at $3/$15, not the cheaper K2.5 that Western trackers still list. Treat every number in this table as a same-week snapshot, not a contract.
The pricing mechanics differ from the Western vendors in 3 ways worth knowing.
DeepSeek now prices by time of day, a first among major providers: peak windows (01:00-04:00 and 06:00-10:00 UTC) bill at 2x the off-peak rate, weekends bill off-peak all day, and cache hits stay at about 3% of the input price against 10% in the West.
Alibaba publishes no batch discount and steps Qwen prices up above 256K input tokens.
And Z.ai sells a GLM Coding Plan from $18/month, a Claude Code-style subscription metered in prompts per 5-hour window: the seat model has crossed over.
Way 2: the same weights on a third-party host
Because the weights are public, US and EU inference providers (Together AI, Fireworks, AWS Bedrock for selected models) host the same models at their own per-token prices. The same DeepSeek or Qwen checkpoint can carry 3 different prices depending on who serves it. What you buy from the host is not the model, it is the serving: SLA, data processing location, enterprise procurement and consolidated billing.
That last point is the one European clients ask about first. The vendors' own APIs process requests on infrastructure in China; whether that is acceptable is a data-governance question your compliance team answers, not a FinOps one. A Western host running the same open weights removes that constraint at a higher per-token price. State the requirement first, then price it.
Way 3: your own GPUs
Self-hosting replaces per-token billing with per-hour billing. An H100 rents for $2.50 to $4.00 per hour on demand; a production GLM 5.2 deployment needs on the order of 8x H200, roughly $300K to own. The GPU meter runs 24/7 whether you send tokens or not, which is the opposite of the variable-cost model that makes cloud AI attractive at low volume.
The per-token arithmetic only starts to favor self-hosting at high, predictable volume: around 1B tokens per month for a GLM-class model by one published estimate, 200–500M tokens/day in our own reference framework. Below that, utilization collapses and the rented GPU costs more per token than DeepSeek's API.
And the arithmetic omits the operational tax: a credible production stack takes 0.5 to 2 dedicated ML-Ops FTE, at $150K to $250K loaded cost each in Western Europe, before the first incident at 2am. A 30% paper saving that requires 2 hires is not a saving.
Our rule, from the FinOps skill this series draws on: self-host only when 5 conditions hold at once
A dedicated platform team, prior GPU workloads in production
Prior GPU workloads in production
ML observability already running
90 days of usage data showing predictable volume
And a TCO model still 30% ahead after the operational tax
Anything less, stay on hosted APIs and revisit in 12 months.
What this changes in your AI bill
Route the simple tasks.
Classification, extraction and summarization on DeepSeek V4 Flash at $0.22/$0.66 off-peak cost 9x to 15x less than Claude Sonnet 5. A gateway (LiteLLM, Portkey) makes the routing a config file, not a project.
Read the license before Legal does.
DeepSeek and GLM ship under MIT, most of Qwen under Apache 2.0, Kimi under a modified MIT. Permissive, but “the model is open” is a claim your counsel verifies per model, not a family trait.
Put the residency question in writing.
Vendor API in China, Western host, or self-host: each answers a different compliance requirement at a different price. Decide on the requirement before comparing per-token rates.
Do not self-host on enthusiasm or hype.
The 5-condition rubric above exists because the failure mode is common: GPUs provisioned on projected volume, running at 20% utilization 6 months later.
Where to start
Same inventory as the 2 previous posts, one column wider: for each workload, note which model it calls, through which of the 3 channels, and what the same tokens would cost on the cheapest channel that meets its compliance requirement.
The gap between those 2 numbers is your routing opportunity, and it is usually the largest single line in an AI cost review.
The FinOps skill OptimNow uses for this analysis, including the self-hosted-vs-managed decision rubric, is open source under CC BY-SA: github.com/OptimNow/cloud-finops-skills.
Download it, load it into your own Claude and run the same analysis on your invoice.
Sources
DeepSeek API pricing, DeepSeek, August 23, 2026
Kimi K3 pricing, Moonshot AI, August 23, 2026
GLM API pricing, Z.ai, August 23, 2026
Qwen API pricing tracker, BenchLM, August 10, 2026
GLM 5.2 pricing guide, Layer3 Labs, July 17, 2026
Sonnet 5 introductory pricing made permanent, Anthropic on X, August 2026
.png)