A senior brokerage analyst summarises the moment: globally, AI coding has passed its steep technology curve, and after the secondary-market cooldown the industry front line moved to close combat. Overseas, Anthropic and OpenAI fight fiercely for users with free quotas and longer trial time, a quota war. In China the fight runs from mindshare to commercial cards: Alibaba claims the top AI-coding market share, Zhipu passed 1bn dollars in annual recurring revenue, and Kimi released its K3 model with coding ability near the world’s best. Behind the standoff is a harsh fact: AI coding has entered its second half of commercial slaughter and breakout.
The barrier is falling
Until 2024, training a usable coding base model cost tens of millions of dollars in corpus cleaning and pretraining. Today, with Meta’s Llama 3, Kimi and DeepSeek open models, general code ability has been commoditised. A startup no longer reinvents the wheel: take a strong open base, feed cleaned code and error-fix logs, wrap it in an agent framework that runs and debugs itself, and ship a system that scores well on SWE-bench quickly. Code data is everywhere, and models, algorithms and open tools are abundant, so the entry barrier is visibly lower.
Coding is also the scene where large models improve fastest and close the loop easiest: rich public training data, clear syntax, and direct feedback from compilers, tests and static scans. Long context, multi-tool calling and coding agents have pushed products from single-line completion to independent repository-level execution. That convergence shrinks the gap between top labs and sets up the homogenisation and price war to come.
The cost trap
Lower barriers invite a ‘hundred-regiment war’ of copycats and a bottomless compute-subsidy sink. Overseas, OpenAI and Anthropic’s ‘unlimited free, extended subscription’ tug-of-war is mutual attrition, a question of who breaks first under token cost. In 2026, AI coding is fully in the agent era: where a user once spent dozens of tokens with Copilot, an agent now reads related files and runs sandbox tests until the bug is fixed, so per-user daily token use explodes. A 20 dollar monthly plan becomes a gold-devouring black hole, and even giants now subsidise. In China, universal coding tools, completion, IDE plugins, even Cursor-like standalone IDEs, have lost the right to charge independently, with giants offering near-free service to individuals and small firms. The differentiator is no longer the model but who converts free usage into paying enterprise.
Read the original report (LeiPhone)
Translated and adapted from LeiPhone (leiphone.com).