The frontier is no longer a single line.

Within the last week, four Chinese labs shipped open-weights coding models — Z.ai's GLM-5.1, MiniMax M2.7, Moonshot's Kimi K2.6, and DeepSeek V4 — and the benchmarks land them at roughly the same agentic-engineering capability ceiling as the Western frontier, at meaningfully lower inference cost.

The story isn't "China caught up." That happened in 2025. The story is: there are now five or six independent labs that can ship a near-state-of-the-art coding model, and the marginal cost of doing so is dropping faster than the gains.

§ 01 What it changes for builders

If you're a security engineer like me, "open-weights coding agent at one-third the cost" is the part to underline. It means more autonomous agents running in more places, which means more agent traffic in your logs, more shell calls per ticket, and more reasons to have a sandbox in your dev stack today, not next year.

The router pattern wins again: pick the cheapest model that clears the eval, default-deny on outbound network from agent processes, and stop treating "which lab made the weights" as the interesting question.