Z.ai releases GLM-5.2: a truly usable 1 million tokens and two levels of deep thinking

Z.ai has released GLM-5.2, an updated coding model with a genuinely working one-million-token context window, two modes of deep thinking and instant integration into popular development tools.

One million tokens: "usable" is the point

Long context windows have long since become a standard line in marketing copy. Claiming a million tokens is easy; making sure the model actually works with all of them is much harder. Most competitors degrade near the limit: they "lose" information from the middle of a long document, start ignoring early instructions or give noticeably less accurate answers.

Z.ai deliberately put the word "usable" in the release description. That signals the team is betting on real processing of the entire context, not on a number in a spec sheet.

For developers this opens up concrete scenarios: load an entire large codebase into a single request, several long documents at once, or the full history of an issue-tracker discussion, and work with them without losing context.

Two levels of "thinking"

Instead of a single generation mode, GLM-5.2 offers two levels of effort:

• High — a balanced mode for everyday tasks: fast, accurate, without excess compute
• Max — extended thinking: the model builds an internal chain of reasoning before answering, which adds depth on complex tasks

The approach is already familiar from OpenAI products (the o1/o3 series) and Anthropic (extended thinking in Claude 3.7). GLM-5.2's advantage is that both modes are available in one model through one endpoint, with no switching between versions.

High suits refactoring and autocomplete; Max suits architectural review, writing tests and debugging tangled errors.

Integration in minutes

GLM-5.2 ships through an Anthropic-compatible API format. For a developer already using one of the supported tools, connecting takes minutes, with no adapters or rewriting of logic:

• Claude Code — Anthropic's terminal-first AI assistant
• Cline — a popular open-source agent inside VS Code
• OpenClaw — Z.ai's own multi-agent platform

The release covers all GLM Coding Plan tiers without exception. No waitlists or priority-access programs.

No benchmarks, for now

Z.ai has not published standard evaluations with the release. In a market where announcements usually open with comparison tables on HumanEval, MMLU and Codeforces, this is an unusual move.

Instead, the company promises open MIT weights within the next week. Once the weights are out, independent researchers will be able to test the model's capabilities themselves, without relying on figures from a press release.

This is either a sign of confidence in the results or a deliberate choice not to disclose details until the open version ships.

What this means

GLM-5.2 arrives in the developer's working environment without friction: one endpoint, two thinking modes, a large context that really works. Z.ai continues to occupy a position between closed commercial models and the open-source community, and MIT weights in a week will make it available for local deployment without any restrictions.