GLM-5.2's three-tier mapping, and why it ended
How the GLM-5.2 template split opus, sonnet, and haiku across three models, why GLM-5.3 collapsed it, and the two-model mapping that replaced it.
Superseded by GLM-5.3. This post documents the June 2026 GLM-5.2 template. The current template maps
glm-5.3[1m]to primary, opus, sonnet, and fable, andglm-5.3-flash[1m]to haiku. See the GLM-5.3 post for the current setup.
GLM-5.2 arrived on the Coding Plan in June 2026 with a 1,000,000-token context window. For a few weeks, claude-multi had a real three-tier spread: GLM-5.2 for opus, GLM-5.1 for sonnet, and GLM-5-Turbo for haiku. It was the first time each Claude Code tier got a different GLM model.
The mapping at 5.2 launch
| Claude Code tier | GLM model | Context window | Role |
|---|---|---|---|
| Opus | GLM-5.2 | 1M | Long, hard, multi-file work |
| Sonnet | GLM-5.1 | 200K | Day-to-day coding |
| Haiku | GLM-5-Turbo | 200K | Quick edits, shell, subagents |
Thinking stayed on at high with 8,000 thinking tokens. The 1M window came from the [1m] suffix on GLM-5.2. GLM-5.1 and GLM-5-Turbo were 200K models, so they matched Claude Code’s default assumption and needed no suffix.
GLM-5-Turbo was the old template choice for the haiku slot. It is not a Coding Plan model, and the current template does not use it.
Why 5.3 collapsed it
Z.ai released GLM-5.3 on 2026-08-14 and stopped serving 5.2 and 5.1 on the Coding Plan endpoint. Requests for either ID route to 5.3. claude-multi v0.11.0 put 5.3 in both the opus and sonnet slots, so the middle tier disappeared. On 2026-08-26, v0.12.0 moved sonnet to GLM-5.3-Flash. Version 0.13.0 moved Flash to haiku and returned GLM-5.3 to sonnet. The current template has two models, not the three this post describes.
The sources are Z.ai’s GLM-5.3 announcement and the GLM-5.3-Flash announcement.
Current mapping
| Claude Code tier | GLM model | Context window |
|---|---|---|
| Primary, opus, fable | glm-5.3[1m] |
1M |
| Sonnet | glm-5.3[1m] |
1M |
| Haiku | glm-5.3-flash[1m] |
1M |
Primary, opus, sonnet, and fable share GLM-5.3. Flash carries three times the plan quota of GLM-5.3 and accepts image input, so it fits the frequent, short calls in the haiku slot. The full spec is on the GLM provider page.
What to do with an old instance
Do not follow the setup instructions from the original version of this post. They describe a template that no longer ships. If you created a GLM instance on 0.10.0, 0.11.0, or 0.12.0, run:
claude-multi doctor fixThe sync rewrites the model IDs, the compaction window, and the output cap, and it preserves your API key and any custom env vars. The old glm-5.2 and glm-5.1 IDs still route to 5.3, so a stale instance will not fail immediately. A 0.10.0 instance keeps the old 64,000 output cap. A 0.11.0 instance has the new cap but still points sonnet at 5.3. A 0.12.0 instance has Flash in sonnet; the 0.13.0 sync moves sonnet back to GLM-5.3 and puts Flash in haiku.
The original post also described the 3x peak / 2x off-peak quota and a promo through September 2026. Z.ai discontinued that system for new users on 2026-07-30 and moved the plan to credits. The GLM-5.3 post has the current formula and plan limits.
What still applies
The mechanics did not change. The [1m] suffix opts a model into the million-token window, and Claude Code compacts around the window it reads from the model name. That was true for GLM-5.2, and it is true for GLM-5.3 and Flash. When the tiers had different windows, the template had to set context per model. Both current models are 1M, so the current template uses one global window at 1000000.
See the GLM provider page for the current spec table, or compare against Kimi K3.1 Pro and MiniMax M3.1 Flash. The CHANGELOG has the full diff.