Kimi K3, Qwen 3.8 and GLM 5.3 Are Not Cheap Models

For about eighteen months, "use a Chinese open-weight model" was reliable cost advice. It is now a assumption worth re-testing, because the flagship tiers from Moonshot, Alibaba, and Z.ai have quietly repriced into the same band as the American labs, and in one case above them.

This is not an argument against any of these models. Two of them are excellent. It is an argument against reaching for them reflexively on the belief that they are the budget choice.

Kimi K3 costs more than GPT-5.6 Terra

Kimi K3 went live on 16 July 2026: 2.8 trillion parameters, a 1M-token context window, native visual understanding, built on Kimi Delta Attention. It is a serious frontier model and it is priced like one. $3.00 per million cache-miss input tokens, $0.30 cache-hit, and $15.00 per million output tokens, flat across the full context with no long-context surcharge.

Set that against GPT-5.6 Terra at $2 / $12 after the 30 July price cut. K3 is 50% more expensive on input and 25% more on output than OpenAI's everyday business tier. Against DeepSeek V4-Flash off-peak at $0.22 / $0.66, K3 is 13.6x on input and 22.7x on output.

Open weights is not a price

The strongest source of the cheap-model belief is the phrase "open weights", which reads as free. It describes a licence, not a cost. What it costs is set by how much hardware the model needs to run at a usable speed.

Kimi K2 is a 1 trillion parameter mixture-of-experts model with 32B active parameters routed across 384 experts. The INT4 quantised weights alone are 594GB and need roughly 640GB of aggregate VRAM, which puts 8x H100 80GB at the tight minimum. FP16 inference needs around 2,308GB, meaning a multi-node cluster. Aggressive 2-bit quantisation gets the footprint to 340GB and still wants 350GB of combined RAM and VRAM.

The builds that people actually attempt at home tell the story. A 384GB DDR5 CPU build runs at about 10 tokens per second. A 4x RTX 3090 rig with 256GB of RAM manages roughly 7. For a chat toy that is tolerable. For an automation pipeline that needs to process a queue, it is not a serious option, and K3 at 2.8T parameters is further out of reach again.

Weights you cannot afford to serve are a licence, not a saving. The honest way to price an open-weight flagship is to look at what hosted providers charge for it, which is exactly the table above.

Qwen 3.8: check which model you are actually getting

Alibaba opened Qwen 3.8 API access on 3 August 2026 at $2 input, $6 output, $0.25 cached, down from Qwen 3.7 Max at $2.50 / $7.50. Qwen 3.8-Max is a 2.4 trillion parameter sparse MoE with roughly 95B active per token, a 1M context window, and up to 128k output tokens. Independent aggregation puts it at 79.9 out of 100, ranking sixth of 218 models, which is genuinely strong.

Two caveats matter before you build on it. The flagship's open weights are promised and not released, so the self-hosting escape hatch does not currently exist, and official benchmarks for Max have not been published, leaving only Alibaba's internal numbers. What did ship open, on 13 to 14 August, is Qwen 3.8-27B under Apache 2.0: a dense vision-language model with a 262,144 token context, 65,536 max output, and 61.7 on SWE-bench Pro.

The 27B is the interesting one for automation work, and it carries a trap. At $0.40 input and $3.00 output (OpenRouter lists $0.45 / $3.20 across two providers), output costs 7.5 times input. Most models sit between 3x and 5x. If your workload generates long outputs, that ratio will hurt you in a way the input price never signals, and it makes the 27B a poor default for summarisation or content generation despite the attractive input rate.

GLM 5.3 is not a new model

Z.ai launched GLM 5.3 on 14 August 2026, and the coverage has been enthusiastic about the benchmark gains, particularly a claimed 84.5% on CyberGym that puts it ahead of Claude Mythos 5 and GPT-5.6 Sol.

The detail that changes how you should read those numbers: GLM 5.3 is not a new architecture and not a new pretrain. It runs on the same 744B parameter base as GLM 5.2, and every reported gain comes from extended post-training. That is a legitimate way to improve a model. It also means the gains are concentrated where the post-training was aimed, and the general-purpose behaviour underneath is the model you already knew.

For anyone planning a migration, two practical blockers. Open weights are staged rather than immediate, released after safety evaluations with reporting suggesting the end of August 2026. And Z.ai has not published a confirmed per-token rate for 5.3, so the working reference is GLM 5.2 at $1.40 / $4.40 with cached input at roughly a fifth of uncached. Building a cost model on a price that has not been announced is not a plan.

What is actually cheap

If the goal is low cost per unit of useful work, the answer in August 2026 is a small open model on a hosted endpoint. Gemma 4 26B A4B at $0.070 / $0.300 is 43x cheaper on input and 50x cheaper on output than Kimi K3, and for classification, extraction, routing, and structured output it does the job. We put the numbers side by side in Gemma 4 against DeepSeek V4.

The frontier tiers from Moonshot, Alibaba, and Z.ai are worth paying for when the task genuinely needs frontier reasoning, long-context work over a million tokens, or vision. Reaching for them because they are Chinese and therefore assumed cheap is how teams end up paying $15 per million output tokens for work a 26B model would have completed. Before any of this decides your architecture, read why per-token price predicts your bill poorly, because token consumption varies more between these models than their prices do.