News

China is shipping frontier-class open-weight models faster than the US, and they cost a fraction to run

Inkling from the US, Kimi K3 and Qwen3.8-Max from China, plus GLM-5.2. Four frontier-class model releases in eight days, three of them open-weight and two of them priced below Claude. What changed for a UK service business.

2026-07-22 · 9 min read

A cinematic editorial shot of physical release cards fanned out on a dark slate desk under warm-cool ambient light. The cards evoke a stack of recent frontier model announcements laid out for comparison, with the wider mood of new releases arriving faster than the eye can track them. Real-world photograph, shallow depth of field.

The 14 July post covered the previous week's wave: GPT-5.6, a new Claude, Gemini, DeepSeek V4 and the rest. Six days later, four more frontier-class models shipped. This post covers those four, and the bigger shift they describe together: the open-weight race is now being written mostly in Beijing and Hangzhou, and the cheap layer of the AI stack just got a lot cheaper.

What actually shipped between 15 and 19 July

Inkling from Thinking Machines Lab (US, 15 July) is the first open-weight model from Mira Murati's lab. Mixture-of-Experts, 975B total parameters with 41B active per forward pass, multimodal, and a feature the others do not have: controllable reasoning effort. You can dial up the reasoning for a hard task and dial it back for a cheap one, instead of paying for the maximum every time. It debuted at 41 on the Artificial Analysis Intelligence Index, which makes it the new leading US open-weights model. The weights are out, you can fine-tune it on Thinking Machines' own Tinker platform. Pricing for hosted inference is not in the public coverage I have seen yet, but it is positioned as US frontier-class open weights.

Kimi K3 from Moonshot AI (Beijing, 17 July) is the one Silicon Valley is paying attention to. 2.8 trillion parameters, open weights due 27 July, 1M token context window. The benchmark numbers are real: it ranks fourth on Artificial Analysis and currently leads Arena.ai's Code WebDev leaderboard. The price is the part that will move markets: $3 per million cache-miss input tokens, $0.30 cached, $15 per million output, flat across the entire 1M window rather than stepped by context tier. That is roughly the Claude Sonnet tier, and it is open weight, which means you can also self-host if your volume justifies the GPU bill.

Qwen3.8-Max from Alibaba (Hangzhou, announced 19 July at the World AI Conference in Shanghai) is 2.4 trillion parameters, multimodal, and pre-announced days after the Kimi K3 release, which is not an accident. The headline number is that Alibaba's own evaluation puts it just behind Claude Fable 5. The honest caveat: this is a pre-announce, and Qwen3.8 is currently API-only. Alibaba has historically open-weighted after a closed preview, but no 3.8 timeline has been published yet. Until the weights land, this is a watch-this-space story rather than a switch-today story.

GLM-5.2 from Z.ai (China, mid-July) is the one that has been quietly running in production at several shops I read. Mixture-of-Experts, 1M token context, and per three separate leaderboards (BenchLM, Thunder Compute, acecloud) it is the strongest all-round open-weight LLM right now for agentic engineering, software development, and long-horizon reasoning. If you have been following the cheap layer of your stack for the last six months, you may already be running a GLM variant. The GLM 5.2 release moves that ceiling up another notch.

Why Chinese labs are shipping faster, and cheaper

There are three honest reasons and one cynical one. The honest reasons first.

They are mostly competing with each other. The domestic Chinese AI market has priced inference as a commodity. DeepSeek, Moonshot, Alibaba's Qwen team, Z.ai, Zhipu, Stepfun, MiniMax, Tencent, Xiaomi, ByteDance, and the rest are all shipping into the same buyer base, and that buyer base expects frontier-class output at near-zero inference cost. The Chinese frontier is competitive because the Chinese market forces it to be.

The training cost per frontier model has fallen sharply across 2025 and 2026. The infrastructure efficiency story (better kernels, smaller active parameter counts, better data, better distillation) is real, and it benefits every lab equally. A model that would have cost $100M to train in 2024 trains for a fraction of that today. The savings show up in the API price.

Open-weight is a strategy, not a giveaway. When you release open weights, you win developer mindshare without having to run a closed API business at scale. The Chinese labs that have gone open-weight first (DeepSeek, Moonshot, Qwen) have done so in part because the developer base they wanted to capture was the global one, not the domestic one. Open weight is a market-entry strategy.

The cynical reason: there is a strategic subsidy angle to some of this, where the headline pricing is supported by state-level infrastructure investment. That is real, and it is the kind of thing that may or may not last. Either way, the price a UK small business pays is the price a UK small business pays. Kimi K3 is $3 per million input tokens. DeepSeek V4 Flash is $0.14 input / $0.28 output per million tokens, which is twenty to forty times cheaper than the equivalent frontier-class closed model. That gap is the point.

The capability bar moved, again

This is the third capability-bar shift in six months, after the DeepSeek V3 moment in late 2024 and the GPT-5 / Claude 4 moment in early 2025. Each time the cheap layer gets better, the work you used to escalate moves down. Each time the frontier layer gets better at one thing, that one thing becomes a default rather than an exception.

For a UK service business, the practical version is short. The cheap layer of your stack (drafting, summarising, classification, triage, basic extraction, first-pass reply drafting, lead scoring) is now good enough that you should not be paying frontier money for ninety percent of it. The frontier layer (final-draft proposals, contract language, security review, customer-facing responses where brand voice has to be exact) is still worth paying for, because the cost of getting it wrong on those tasks is real.

The old worry, that you were locked into one provider because the cheap model could not do enough, is mostly gone. You can route cheaply across Kimi K3, GLM-5.2, DeepSeek V4 and a half-dozen smaller open-weight models, and pick the one that does the specific job best. The orchestration layer (Hermes-style, or whatever else you use) gets more valuable, not less, because the routing decision is the one that pays back the most attention.

What this means if you are running AI yourself

The boring, correct answer applies. Run your top ten prompts against the new models that matter to your work, the same prompts you already use, not a curated benchmark set. Pick the one that does your job better. Move on. The story of this week is not a single breakthrough. It is that the open-weight race has a clear leader right now and it is not based in California. That is good for the economics of running AI in a small business, and it is the kind of structural shift you only have to understand once, because it does not reverse.

If you are a UK service business owner, the takeaway is short and good. The AI you are paying for (whether that is ChatGPT, Claude, Gemini, or the cheaper open-weight layer running underneath) just got better and cheaper in the same week. The vendors that are on top of their game will already be switching the cheap layer of their stack to one of these new models. If you do not see a change in the next two to four weeks, that is a signal to ask your vendor.

Frequently asked questions

Is the cheap layer of an AI stack now genuinely as capable as the frontier layer was a year ago?

For most of the work a UK service business asks an AI to do, yes. Drafting replies, summarising documents, classifying enquiries, drafting social content, extracting structured data from a PDF, scoring inbound leads. Open-weight models at the Kimi K3, Qwen3.8, and GLM-5.2 tier can do all of that at frontier-class quality, at a small fraction of the per-token cost. The frontier layer still wins on the work where the cost of being wrong is high and the workflow is bounded: the proposal sent to a £100k client, the security review, the contract language, the response that has to be right the first time. The cheap layer is now good enough that the question for a small business is no longer which model to use. It is which task to route where, and how to keep the routing honest.

Why are Chinese labs pricing frontier-class models so much lower than the US frontier?

Three honest reasons and one cynical one. The honest reasons: the Chinese labs are mostly competing with each other in a domestic market that has priced inference as a commodity, the training cost per frontier model has fallen sharply across 2025 and 2026, and open-weight releases let them win developer mindshare without running a closed API business. The cynical reason: there is a strategic subsidy angle to some of it. Either way, the price a UK small business pays is the price a UK small business pays. Kimi K3 is $3 input / $15 output per million tokens, roughly the Claude Sonnet tier. DeepSeek V4 Flash is $0.14 input / $0.28 output per million tokens, which is twenty to forty times cheaper than the equivalent frontier-class closed model. That gap is the point.

Does an open-weight model mean my data is private?

Only if you run it yourself. Open-weight means anyone can download the model, including you. If you then run it on your own hardware or on infrastructure you control, your prompts and your data do not leave that boundary. If you use a hosted endpoint (OpenRouter, Fireworks, Together, the lab's own API), the same rules apply as for any other cloud API: the provider can see your traffic unless they explicitly publish a no-retention policy. Read the provider's data policy before assuming privacy. The honest version: open-weight gives you the option of self-hosting. It does not give you privacy by default.

Should a UK service business switch to a Chinese open-weight model right now?

Not blindly. The cheap layer of most UK small business stacks is already open-weight, and most of it is Chinese-lab origin already (DeepSeek, Qwen, GLM, Yi). The decision is about which specific model to run, on which task, with which routing. For drafting, summarising, classification and triage, Kimi K3, GLM-5.2 and the DeepSeek V4 family are genuinely strong picks at frontier-class price points. For final-draft proposals, contract language and anything customer-facing where the brand voice has to be exact, stick with the frontier model you already trust, and route to it. Test the new model against your top ten prompts before assuming it is a drop-in replacement for the model you currently use on that task.