Z.ai

Z.ai is the international storefront for a family of models: a consumer chat at the main address, and a separate developer platform where you get an API key, top up a balance and call the models. The platform documents a current flagship that it describes as its strongest at software engineering and agent work, a multimodal sibling aimed at visual programming, and the previous generation, with a published migration guide from the older one. Around the core calls it offers web search, function calling, context caching, structured output and an explicit thinking mode, plus a trio of packaged agents for slides and posters, translation, and video effects. Software development kits are provided for Python and Java, and an OpenAI-compatible path exists for callers who want to keep their existing code.

It suits developers who want a capable model at a lower price than the best-known Western options and are willing to adopt a less familiar ecosystem, and it suits anyone whose agent tooling supports a compatible endpoint, since the vendor says its subscription works with a set of popular agent tools. It suits less well a team that needs the widest possible third-party tooling and community answers in English, because the developer documentation is where the detail lives and much of the surrounding ecosystem is elsewhere. It suits less well any organisation whose procurement rules restrict where model traffic and billing are handled.

Not verified: the main chat address is rendered by scripts and returned only its page title when fetched, so nothing about the consumer product's features or quality was read; the description above comes from the developer documentation, which did load. The subscription price is the one printed on the documentation's own promotional line and was seen on that page, but the pricing page itself was not opened, so what the tiers include and how metering works are unknown. The claim of compatibility with a large number of agent tools is the vendor's own. The parent company, its ownership, and how it relates to the models offered on this platform were not checked, and no model was called, so output quality on coding or agent tasks was not observed.