[FEATURE]: Add GLM-5.2 model support for Z.AI provider

[phalla-doll]

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

Z.AI just released GLM-5.2 — their latest and most capable reasoning model. It is available to all Z.AI Coding Plan subscribers (Max, Pro, Lite) and supports a 1M context window, reasoning with configurable effort levels, and tool calling.

Specs (from Z.AI docs):

Feature	Value
Model ID	glm-5.2
Reasoning	Yes — effort levels high (default) and max
Context window	1,000,000 tokens (1M)
Max output tokens	131,072
API	OpenAI-compatible
Coding Plan endpoint	https://api.z.ai/api/coding/paas/v4
Tool calling	Yes

Current state: GLM-5.2 is not listed on models.dev (the Z.AI provider page lists up to GLM-5.1, GLM-5-Turbo, GLM-5V-Turbo), and there are no existing PRs or issues mentioning GLM-5.2.

What needs to happen

1. Add GLM-5.2 to models.dev zai provider TOML — this is the prerequisite for opencode to pick it up via its model catalog (fromModelsDevProvider() in provider.ts). I will file a separate issue/PR there.

2. Reasoning effort mapping for Z.AI API: GLM-5.2 supports high and max effort levels (Z.AI docs map Claude Code's /effort levels to these). PR feat(bedrock): add GLM-5 reasoning support on AWS Bedrock #28264 adds GLM-5 reasoning via variant: "high" on AWS Bedrock, but we need to confirm the Z.AI OpenAI-compatible API provider also maps effort/variant correctly.

3. 1M context window handling: Z.AI docs use a [1m] suffix convention (e.g., glm-5.2[1m]) to enable 1M context. Opencode's model definition uses contextWindow directly — we should either define one model with contextWindow: 1000000, or investigate whether two variants are needed.

4. Coding Plan base URL: The Coding Plan uses https://api.z.ai/api/coding/paas/v4 which differs from the standard endpoint on models.dev (https://api.z.ai/api/paas/v4). Users on the Coding Plan need to configure the correct base URL.

5. Tool-call parsing: GLM-5 has known tool-call JSON parsing issues ([bug] GLM 5 keeps screwing up the json parsing of read tool calling #13982, fix(opencode): recover malformed GLM/NVIDIA tool calls #25971). GLM-5.2 should be verified against these to avoid regressions.

References

• Z.AI model switching docs: https://docs.z.ai/devpack/latest-model
• PR feat(bedrock): add GLM-5 reasoning support on AWS Bedrock #28264 (GLM-5 reasoning on Bedrock — reference implementation): feat(bedrock): add GLM-5 reasoning support on AWS Bedrock #28264
• models.dev Z.AI provider page: https://models.dev/providers/zai

[gmaynez]

While this gets added to models.dev, you can add the following to your opencode.json configuration (if you have Z.AI Coding Plan) and use it right away:

  "provider": {
    "zai-latest": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Latest Z.AI",
      "options": {
        "baseURL": "https://api.z.ai/api/coding/paas/v4",
        "apiKey": "YOUR_API_KEY_HERE"
      },
      "models": {
        "glm-5.2": {
          "name": "GLM 5.2",
          "limit": {
            "context": 1000000,
            "output": 131072
          },
          "options": {
            "reasoningEffort": "max"
          }
        }
      }
    }
  }

[ririnto]

glm-5.2 not provided for api. for now, can use with coding-plan.

API will arrive on next week.

And, at model page, high / max guide is for only claude code. I'm not sure /chat/completions endpoint support this.

• https://x.com/Zai_org/status/2065704919299235870
• https://docs.z.ai/devpack/latest-model

Need official announce.