Open-weight Chinese models now hold the top of the open leaderboard, and the same week two US labs cut prices to answer them

The most consequential fact this week is not a single release but a leaderboard position. Artificial Analysis now ranks Kimi K3 from Moonshot AI and Qwen3.8 2.4T A95B from Alibaba as the highest-intelligence open-source models, ahead of DeepSeek V4 Pro and Zhipu's GLM-5.2. All of these are open weights, meaning the weights can be downloaded and run anywhere. On Scale's SWE-bench Pro leaderboard, GLM-5.2 leads the open-weight field at 62.1%, within striking distance of GPT-5.4 (xHigh) at 59.1% on the standardized public set. The gap between what you can license from a hosted US lab and what you can download and run yourself has narrowed to the point where an organization choosing between them is weighing deployment control against a few points of measured capability, not a capability chasm.

What corroborates that reading is what happened on price. OpenAI and Anthropic both cut prices on their models this week, and the reporting attributes the move directly to competition from Chinese rivals. DeepSeek, for its part, went the other way and launched V4-Pro at up to 14 times the price of V4-Flash: $1.32 per million input tokens and $3.96 per million output tokens, against $0.14 and $0.28 for Flash. So the same week produced a segmenting Chinese vendor moving upmarket and two US labs cutting prices to hold ground. Both are responses to the same pressure, which is that capability at a given intelligence level is getting cheaper and more portable faster than any single lab can defend a margin on it.

The cost pressure is not abstract. Canva cut its 2026 revenue growth forecast from 30% to 20% after AI feature costs ran past its expectations, then reported it had reduced the cost of serving an AI task by roughly 90% by moving to in-house models. That is the whole argument in one company: inference cost is now a first-order line item, and the fix Canva reached for was owning the model rather than renting it. Q2 revenue still reached $921.9 million, up 25.2% year on year, so this is a margin story, not a demand story. The same logic that pushed Canva in-house is what makes an open-weight model at 62.1% on SWE-bench Pro a commercial proposition rather than a research curiosity.

Against that, the adoption evidence this week is almost entirely about what is being sold, not what has landed. IBM's partnership to embed OpenAI frontier models into IBM Consulting and its new OpenAI Practice, Anthropic being evaluated at a $2 trillion valuation, OpenAI and Anthropic's price cuts: these are market-structure signals, movements in vendor offerings and capital, not proof of deployment inside enterprises. The one number that points at real adoption is Salesforce reporting that agentic AI deployments more than doubled year on year in its own index, and that is a vendor measuring its own customers. Read carefully, the week shows the supply side reorganizing aggressively around a price and capability squeeze, with deployment evidence still thin and mostly self-reported.

Federal money, meanwhile, is going where it always goes: into research and services rather than models. The Department of Defense led $671.2M across 47 awards in physical, engineering and life sciences R&D, and computer systems design services drew $278.9M. That is procurement of capability-building, not procurement of frontier models, and it is a useful counterweight to the launch noise. The buyers with the largest checkbook are paying for integration and research, which is exactly where the open-weight shift makes the model itself the cheap part.

What we track and how it's verified