Add ZAI Coding Plan GLM-5.2 with Anthropic provider model

[ririnto]

Summary

• Add new model entry: providers/zai-coding-plan/models/glm-5.2.toml
• Configure model metadata based on https://docs.z.ai/devpack/latest-model.md
  • name = "GLM-5.2"
  • reasoning = true
  • tool_call = true
  • interleaved.field = "reasoning_content"
  • limit.context = 1000000, limit.output = 131072
  • reasoning_options = [{ type = "effort", values = ["high", "max"] }]
• Use provider override block for Coding Plan serving layer:
  • npm = "@ai-sdk/anthropic"
  • api = "https://api.z.ai/api/anthropic"

Motivation

The provider field was intentionally switched to Anthropic SDK because the Z.AI latest-model documentation notes GLM-5.2 supports high/max reasoning options, which aligns with the Anthropic reasoning mode capabilities required here.

Additional Notes

• No provider-agnostic metadata file under models/ was added because this change is specific to the Z.AI Coding Plan wrapper.
• This PR follows the repository contribution pattern used for model additions (for example, issues Add glm-4.6 for zhipuai, zhipuai-coding-plan, zai and zai-coding-plan #250 and Add GLM 5.1 to the list of models for NanoGPT #1291) and builds on the same Z.AI model entry style.

Closes #2276

[leszek3737]

First create a model in the models/zhipuai folder and then refer to it according to the readme:

Model-only facts live in models/, using the same path-style IDs as provider models. For example, models/openai/gpt-5.toml defines metadata for the underlying GPT-5 model, while providers/openai/models/gpt-5.toml defines OpenAI-specific serving details such as pricing.

Use model metadata for provider-agnostic facts:

• name, family, release_date, last_updated, knowledge
• attachment, reasoning, tool_call, structured_output, temperature
• [limit] defaults like context, input, and output token limits
• [modalities] defaults
• open_weights, license, links, weights, and benchmarks

Example:

name = "GPT-5"
family = "gpt"
release_date = "2025-08-07"
last_updated = "2025-08-07"
attachment = true
reasoning = true
temperature = false
tool_call = true
structured_output = true
open_weights = false

[limit]
context = 400_000
input = 272_000
output = 128_000

[modalities]
input = ["text", "image"]
output = ["text"]

[[benchmarks]]
name = "Benchmark Name"
score = 72.5
metric = "accuracy"
source = "https://example.com/results"

[[weights]]
label = "Model weights"
url = "https://huggingface.co/example/model"
format = "safetensors"

Provider TOMLs can inherit these facts with base_model and then keep only provider-specific fields or overrides:

base_model = "openai/gpt-5"

[cost]
input = 1.25
output = 10.00
cache_read = 0.125

[limit]
context = 200_000 # optional provider override
output = 32_000

Provider fields win over model metadata during generation. Use this when the underlying model is the same but a provider serves it with different context limits, modalities, features, or pricing.

[ririnto]

Split the entry into a provider-agnostic models/zhipuai/glm-5.2.toml and reference it from the provider TOML via base_model = "zhipuai/glm-5.2". The provider file now keeps only provider-specific fields (reasoning options, interleaved, cost, and the per-model Anthropic [provider] override). d210e45d

[mugnimaestra]

LGTM, nice work adding both the provider-agnostic metadata and the Coding Plan entry. Hoping this gets merged soon!