GLM-5.2: Z.ai’s new flagship lands on the Coding Plan

GLM-5.2 is Z.ai’s new flagship model, released June 13, 2026 and available immediately to every GLM Coding Plan subscriber. Z.ai describes it as delivering powerful coding capabilities, usable 1M-token context support, and continued strengths in long-horizon tasks — the kind of multi-step, agentic work where a model has to stay coherent over a long session rather than nail a single completion.

What is unusual about this launch is the sequencing. The model is in developers’ hands today, but the standalone API, the Z.ai chatbot, and the open-source weights were all announced for next week, and no benchmarks were published at release. Z.ai shipped the distribution channel — the GLM Coding Plan — ahead of the proof. That is the opposite order of a typical frontier launch, and it tells you who this release is aimed at: existing GLM Coding Plan subscribers who can switch to GLM-5.2 with an environment-variable change, today.

This post covers what Z.ai actually confirmed, why there are no GLM-5.2 numbers to cite yet, the 1M-context and effort-level claims, how the GLM Coding Plan is structured, the exact steps to point a coding agent like Claude Code at GLM-5.2, and the MIT open-source promise. For the model this builds on, see our GLM-5.1 coding-benchmark analysis from earlier this year.

1. 01
   It is a coding-plan rollout, not a benchmark launch.GLM-5.2 went live to all GLM Coding Plan tiers (Lite, Pro, Max, Team) on June 13, 2026, but Z.ai published no benchmark scores at release. The standalone API, the chatbot, and the open weights are all scheduled for next week. Treat today as availability, not proof.
2. 02
   The headline claims are 1M context and long-horizon work.Z.ai positions GLM-5.2 around three qualities: powerful coding, usable 1M-token context, and continued strength on long-horizon tasks. These are vendor descriptions, not measured results — useful as a statement of intent, not as a leaderboard placement.
3. 03
   No GLM-5.2 numbers exist yet — only the lineage does.Every benchmark you can find today belongs to GLM-5 or GLM-5.1, not GLM-5.2. GLM-5's reported 77.8% on SWE-bench Verified set the family's track record, and GLM-5.1 claimed roughly 94.6% of Claude Opus 4.6's coding score. Use those as context for what 5.2 builds on, never as 5.2 results.
4. 04
   Open weights ship under MIT — a more permissive license.Z.ai says GLM-5.2 will be open-sourced next week under the MIT License. The GLM-5 base was Apache-2.0; both are permissive, but MIT is the simplest and most widely accepted, reinforcing Z.ai's open-weight positioning against the closed frontier.
5. 05
   Switching is an env-var change for existing subscribers.Inside the GLM Coding Plan, GLM-5.2 works with Claude Code, Cline, OpenCode, Roo Code, OpenClaw, Kilo Code, Crush, and Goose. In Claude Code you point the model env vars at glm-5.2[1m] and set the auto-compact window to 1,000,000 — the switching cost of a trial is minutes, not a migration.

Z.ai announced GLM-5.2 on June 13, 2026 as its new flagship model and made it available the same day to all GLM Coding Plan users. The confirmed facts are narrow but clear: GLM-5.2 is live across the Lite, Pro, Max, and Team tiers; Z.ai claims powerful coding, usable 1M-token context, and continued long-horizon strength; the model exposes two thinking-effort levels (High and Max); and the company is pairing the release with an open-source commitment. Z.ai’s documentation already lists GLM-5.2 alongside GLM-5.1, GLM-5-Turbo, GLM-4.7, and GLM-4.5-Air as a supported model on every plan.

What is deferred is just as important. Z.ai stated that API and chatbot services will launch next week, and that the model will be open-sourced next week under the MIT License. No technical report, no parameter count specific to 5.2, and no benchmark scores accompanied the announcement. Treat the items below as roadmap, not shipping features — the way Z.ai itself framed them.

As of launch day, June 13: there were no GLM-5.2 benchmark numbers. Not low ones, not high ones — none. (The full scorecard has since landed; see the update at the top of this post and our GLM-5.2 benchmarks breakdown.) Z.ai published the model and the coding-plan availability without a scorecard, and independent labs had no weights and no API to test against. Any specific figure attached to “GLM-5.2” in the first few days was almost certainly a GLM-5 or GLM-5.1 number being mislabeled. The honest position on launch day was that the model’s measured performance was unknown.

What the lineage does provide is a track record to set expectations against. GLM-5, the family’s base model, is credited with 77.8% on SWE-bench Verified — reported in independent coverage as the highest open-source score on that benchmark at the time, which is part of why Z.ai’s self-reported numbers tend to get the benefit of the doubt. GLM-5.1 then claimed roughly 94.6% of Claude Opus 4.6’s coding score on a Claude Code harness, though that figure was self-reported and never independently corroborated. Our GLM-5 architecture analysis and GLM-5.1 benchmark breakdown cover both in detail.

The practical read: GLM-5.2 enters with a credible pedigree and zero verified evidence of its own. That is a normal day-one state for any model — the mistake is acting on the vendor’s adjectives as if they were leaderboard positions. The cross-vendor comparisons that matter (SWE-bench Verified, Terminal-Bench, NL2Repo re-runs) cannot happen until the API and weights land next week. Until then, qualify every claim, including the ones in this post.

Two of the confirmed details are concrete enough to plan around. The first is context: GLM-5.2 supports a 1M-token window, addressed through the glm-5.2[1m] model id, with a maximum output of 131,072 tokens. Z.ai is careful to call this “usable” 1M context — a hedge that points at the real question with every long-context model, which is whether retrieval quality holds across the full window or degrades in the middle. That is exactly the kind of claim independent testing exists to check, so file it under “promising, unverified.”

The second is effort. GLM-5.2 exposes two thinking-effort levels, High and Max, and Z.ai recommends Max for coding tasks to enable deeper reasoning. Inside Claude Code the mapping is explicit: the low, medium, and high effort settings all map to GLM-5.2’s High, while xhigh, max, and ultracode map to its Max. In other words, if you want the model’s deepest reasoning you have to ask for it — the default-tier efforts land on High, not Max.

“Available to all GLM Coding Plan users” is the whole distribution story, so it is worth understanding the plan. The GLM Coding Plan is a subscription that meters usage in prompts rather than tokens, where Z.ai estimates each prompt invokes the model 15-20 times — a reflection of how agentic sessions fan a single request into many model calls. Z.ai lists the plan from around $18/month in its documentation, across four tiers (Lite, Pro, Max, Team), with periodic discounts; check z.ai/subscribe for current per-tier pricing.

Usage is governed by two simultaneous caps — a rolling 5-hour limit and a weekly limit — and the more capable models (GLM-5.2, GLM-5.1, GLM-5-Turbo) draw down quota faster at peak hours. The figures below are Z.ai’s approximate per-tier prompt allowances. They matter because the difference between Lite and Max is not capability — every tier gets GLM-5.2 — but how much of it you can use before hitting a wall.

The reason this launch is frictionless for existing subscribers is that the GLM Coding Plan exposes an Anthropic-compatible endpoint, so agents built for Claude work with a base-URL and key swap. By default, Z.ai maps Claude Code’s model tiers to GLM-4.7 (for Opus and Sonnet) and GLM-4.5-Air (for Haiku) — so out of the box you are not on GLM-5.2. To use the new flagship you edit the model env vars in your ~/.claude/settings.json. Our GLM-4.7 setup guide walks through the provider configuration this builds on.

The change is three model variables plus the compact-window setting that the 1M context needs. After the provider base URL and API key are in place, set the defaults to glm-5.2[1m] as below, then relaunch and confirm with /status in Claude Code. To pick the deepest reasoning, switch effort with /effort — recall that Max effort is what maps to GLM-5.2’s Max.

Beyond Claude Code and Cline, Z.ai lists GLM-5.2 as working with OpenCode, Roo Code, OpenClaw, Kilo Code, Crush, and Goose — most of which support custom model configuration the same way. The takeaway is that for anyone already paying for a GLM Coding Plan, evaluating GLM-5.2 is a settings edit and a relaunch, not a procurement decision. That low switching cost is the entire point of releasing to the plan first.

The open-source commitment is the strategic spine of the release. Z.ai framed the launch with the line “the future of AI is open, and it belongs to the people,” and backed it with a concrete promise: GLM-5.2 weights open-source next week under the MIT License. That is a small but real shift from the GLM-5 base model, which shipped under Apache-2.0. Both are permissive, OSI-approved licenses; MIT is the shorter, simpler one, with no explicit patent grant but the least friction for adoption.

For teams, an MIT-licensed frontier-adjacent coding model — if the weights land as promised and the capability holds — is a genuine self-hosting and data-sovereignty option. It is the same lever that makes models like Kimi K2.7-Code attractive to regulated industries: code never leaves the perimeter, and there is no per-token meter. The caveat is identical too — open weights for a large model are a serious serving commitment, so for most teams self-hosting is a compliance fallback and negotiating lever, not the default path.

GLM-5.2 enters one of the most crowded quarters the coding-model market has seen. Anthropic’s Claude Opus 4.8 anchors the closed frontier; Moonshot just shipped Kimi K2.7-Code as an open-weight coding specialist; Alibaba’s Qwen 3.7 Max took closed-model benchmark wins in late May. The matrix below places the realistic alternatives a team would weigh this month — with the standing caveat that GLM-5.2’s independent placement is a week away, so it is positioned on confirmed attributes (availability, context, license), not on scores it does not yet have.

If you already pay for a GLM Coding Plan, trial it today. The cost is a settings edit: point your model env vars at glm-5.2[1m], set Max effort, and run one real mid-complexity task — a multi-file refactor or a failing-test fix — on a branch. Score it on end-to-end completion without intervention. That is the long-horizon claim Z.ai is making, so test it directly rather than trusting the adjective.

If you do not, wait a week — there is little to evaluate yet. With no API, no chatbot, no weights, and no benchmarks until next week, there is no low-friction way to assess GLM-5.2 from outside the plan right now. Subscribing solely to test a benchmark-less day-one model is premature; the standalone API next week is the natural entry point for non-subscribers.

Treat the MIT weights as a future option, not a current one. If self-hosting or data residency is on your roadmap, the MIT release is worth tracking — but it is announced, not shipped. Verify the weights are actually published before scoping any deployment, and remember a large model is a real serving commitment. For organizations weighing that build-out, our AI transformation services cover model selection and deployment strategy, and our web development team ships with these agentic stacks daily.

Hold platform decisions until the independent numbers land. GLM-5 earned its credibility by backing a headline claim externally, but that is GLM-5, not GLM-5.2. SWE-bench Verified, Terminal-Bench, and NL2Repo re-runs typically follow an open-weight release within one to two weeks. Start a contained trial now if you are on the plan; make the standardize-on-it decision after the third-party data exists.

A distribution-first launch, with the proof to follow.