GLM-5.3 Hands-on Guide: Always-on Thinking, Three Reasoning Tiers, and the API Support Matrix (AIHubMix)

Core content summary

GLM-5.3 is Z.ai's flagship model released on 2026-08-14. It uses exactly the same base as GLM-5.2 and is upgraded purely through post-training. Every "verified" conclusion in this article comes from real calls made through the AIHubMix API (using the Chat Completions, Responses, and Messages protocols) on 2026-08-14.

1. Model specification at a glance

2. Key API differences from GLM-5.2 (most important)

• Official migration recommendation: change applications that previously sent {"type":"disabled"} to {"type":"enabled"} and set reasoning_effort:"low".

• Measured: sending disabled through AIHubMix still returns 200 and thinking still occurs (automatically converted according to the semantics of the official channel). If a client depends on "turning off thinking to save tokens," switch to reasoning_effort:"low".

• Measured: an invalid enum value returns 200 and falls back to the default max; on the same arithmetic problem, thinking tokens were 27 with low versus 39 with max.

3. Calling the three protocols and reading thinking content

Chat Completions (thinking content is in the reasoning_content field, and in delta.reasoning_content for streaming):

from openai import OpenAI
client = OpenAI(base_url="https://aihubmix.com/v1", api_key="<KEY>")
completion = client.chat.completions.create(
    model="coding-glm-5.3",
    reasoning_effort="max",   # low / high / max, max by default
    extra_body={"thinking": {"type": "enabled"}},
    messages=[{"role": "user", "content": "Compute the square root of (17*23-19*11), rounded down. Digits only."}],
)
print(completion.choices[0].message.reasoning_content)
print(completion.choices[0].message.content)  # observed: "13"

• Measured: usage.completion_tokens_details.reasoning_tokens reports thinking usage—for the same problem, low used 27 and max used 39.

Responses API (thinking content is a reasoning output item; the text is in summary_text within the summary array):

response = client.responses.create(model="coding-glm-5.3", input="What is the capital of France? City name only.")
# output item types: ["reasoning", "message"]
# reasoning item: {"type": "reasoning", "summary": [{"type": "summary_text", "text": "..."}]}
# usage.output_tokens_details.reasoning_tokens: 80

• Measured: even without passing any reasoning parameters, a reasoning item is returned by default (no explicit opt-in is required).

Messages (Anthropic protocol) (thinking content is a native thinking content block):

client = Anthropic(api_key="<KEY>", base_url="https://aihubmix.com")
response = client.messages.create(model="coding-glm-5.3", max_tokens=4096,
    messages=[{"role": "user", "content": "What is the capital of France? City name only."}])
# content block types: ["thinking", "text"]

4. Tool calls and parallel tools

• All three APIs were verified to work; the Responses API showed parallel tool calls within a single turn (the official declaration is supports_parallel_tool_calls: true).

• Upstream limits: at most 128 functions in tools; the native tool_choice support is limited to auto.

• In Chat Completions, tool_choice:"none" was measured to work (no further tool calls); in the Messages protocol, tool_choice:{type:"none"} was still observed to produce tool_use. To disable tools, remove the tools parameter directly, or use tool_choice:"none" with Chat Completions.

5. Structured output

• response_format supports text and json_object; the upstream service does not provide a json_schema mode. When a strict schema is required, put the JSON Schema in the prompt and validate it on the client.

• Measured: response_format={"type":"json_object"} returned valid JSON containing the requested key.

6. Other details

• Context caching / automatic caching is available (on the AIHubMix side).

• This article reflects the AIHubMix aggregation channel (coding-glm-5.3 is its preview route); the model ID on the official channel is glm-5.3, so keep the two distinct.

Key quotations from the original

"thinking.type no longer supports disabled — thinking cannot be turned off."

"If your client relied on 'turn off thinking to save tokens', switch to reasoning_effort: 'low'."