What GLM-5.2 Instant is: GLM-5.2 without the reasoning phase
The GLM-5.2 Instant API gives you the same GLM-5.2 model with one change. It writes the answer directly, with no reasoning phase in front of it. The price is the same, so the choice rests on two things: your client, and how soon you need the visible answer.
So which ID should you put in your client, and what do you give up? Start with what GLM-5.2 does by default. It thinks before it answers. That thinking is a block of text called the reasoning trace. Z.ai's API turns thinking on by default for GLM-5.2. The chat template is the text format that wraps every request to the model. GLM-5.2's template also turns reasoning on, unless a request sets `enable_thinking` to false.
The trace is useful on hard problems. It also costs tokens and time before the first word of the answer appears. GLM-5.2 Instant removes that step. Our model page for GLM-5.2 Instant lists reasoning as disabled.
The base model comes from Z.ai. It has 744 billion parameters, the learned weights of the model. Only 40 billion of them work on each token, so each token needs far less compute than the full size suggests. It is released under the MIT licence. Z.ai describes GLM-5.2 as built for long coding tasks. On its own API, Z.ai offers several thinking effort levels for GLM-5.2. On our API, treat reasoning as on or off. Specs and benchmarks for the base model are on our GLM-5.2 page.
When the Instant ID helps: clients that cap max_tokens
Many coding tools send a fixed `max_tokens` value with every request. That setting caps how many tokens one answer may use. On GLM-5.2, the reasoning trace counts towards that cap.
Some of these tools also never show the trace to you. Our serverless inference docs name GitHub Copilot, Cursor and similar setups. In these tools you paste in your own API key for another provider. In such a client, the model can spend the whole budget on reasoning. The visible answer then comes back empty or cut off.
Here is a worked example with assumed numbers. A tool sends a question with `max_tokens` set to 1,024. Assume GLM-5.2 writes 1,000 tokens of reasoning first. The 1,000 is a hypothetical figure for illustration, not a measurement. The other two inputs are real: GLM-5.2 writes about 55 tokens per second in our status checks, and output costs $4.50 per million tokens.
reasoning time = 1,000 tokens / 55 tokens per second = about 18 seconds
answer left = 1,024 - 1,000 = 24 tokens
reasoning cost = 1,000 x $4.50 / 1,000,000 = $0.0045 per requestSo the reader waits about 18 seconds and gets 24 tokens of answer. At 10,000 such requests, the hidden reasoning adds $45 of output tokens. With z-ai/glm-5.2-instant, all 1,024 tokens go to the answer.
The example does not show answer quality. Without reasoning, the model may do worse on hard, multi-step problems. Real reasoning length also varies a lot from prompt to prompt.
If your client can send extra request fields, you have another option. You can keep z-ai/glm-5.2 and turn reasoning off per request. Our API accepts either of two fields for this, and one is enough. Here is the shorter one:
{
"model": "z-ai/glm-5.2",
"reasoning_effort": "none"
}The other is `chat_template_kwargs` with `enable_thinking` set to false. The Instant ID is for clients that cannot send either field.
GLM-5.2 Instant price, context window and request limits
GLM-5.2 Instant costs the same as GLM-5.2. You pay per token, with no base fee.
| Price per 1M tokens | GLM-5.2 Instant | GLM-5.2 |
|---|---|---|
| Input | $1.50 | $1.50 |
| Cached input | $0.38 | $0.38 |
| Output | $4.50 | $4.50 |
Prices are from our models page, read 23 October 2026. The cached input price comes from prompt caching. That means a request that starts with the same text as an earlier one pays $0.38 instead of $1.50 for the repeated part. Caching is automatic and best effort. Keep the stable part of your prompt at the front, and read `usage.prompt_tokens_details.cached_tokens` to see what was cached.
The context window is 1M tokens. That is the amount of text one request can hold, prompt and answer together. Z.ai lists the same 1M for GLM-5.2. Within it, one answer can be up to 131,072 tokens long, which is Z.ai's 128K counted in blocks of 1,024. On our API, one request can also run for up to 300 seconds. Set `stream: true` for long outputs.
GLM-5.2 Instant reads text only. If you send an image, the request succeeds but the image is ignored. Tool calling works in the standard OpenAI format.
GLM-5.2 Instant speed in our status checks
We publish speed figures from our own status checks on each model page. Two numbers matter here. Time to first token is the wait before the first token arrives. Output speed is how many tokens per second follow after that.
| Median, 7 days to 23 October 2026 | GLM-5.2 Instant | GLM-5.2 |
|---|---|---|
| Time to first token | 409 ms | 398 ms |
| Output speed | 53 tokens/s | 55 tokens/s |
Medians from our status checks on short requests, measured 23 October 2026 at 13:08 UTC.
On short requests, the two models run at about the same speed. The 11 ms gap is too small to matter. Our status checks count reasoning tokens in both figures. On GLM-5.2, the first token can be a reasoning token, so its time to first token is not the time until the visible answer starts. On Instant, every token is answer, so the figures describe the answer itself. On prompts that make GLM-5.2 think, its visible answer starts later. That gap is what Instant removes.
For live availability and incidents, check our status page.
Calling z-ai/glm-5.2-instant through our OpenAI-compatible API
Lyceum is a sovereign European AI inference provider for open-source models. In plain terms, we run open models on our GPUs in Europe and you call them over an API. Serverless inference means you send requests to models we keep running and pay per token, with no servers of your own. Our serverless models run in Paris and Finland. We don't keep your prompts or outputs, and we never use them for training.
You call GLM-5.2 Instant through our OpenAI-compatible API. Set the base URL, your API key and the model ID. Create the key in the Lyceum dashboard.
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["LYCEUM_API_KEY"],
base_url="https://api.lyceum.technology/openai/v1",
)
response = client.chat.completions.create(
model="z-ai/glm-5.2-instant",
messages=[{"role": "user", "content": "Explain this stack trace in two sentences."}],
max_tokens=1024,
)
print(response.choices[0].message.content)In a coding tool, enter the same base URL and paste z-ai/glm-5.2-instant as the custom model name.
Instant model IDs on our API, and which ones to use
Our API serves five model IDs that end in "-instant". They do not all behave the same way today.
| Model ID | Based on | Use in tools that hide reasoning |
|---|---|---|
| z-ai/glm-5.2-instant | GLM-5.2 | Yes, answers directly |
| qwen/qwen3.8-27b-instant | Qwen3.8 27B | Yes, answers directly |
| qwen/qwen3.8-flash-next-instant | Qwen3.8 Flash Next | Yes, answers directly |
| z-ai/glm-5.3-instant | GLM-5.3 | Avoid for now |
| z-ai/glm-5.3-flash-instant | GLM-5.3 Flash | Avoid for now |
Behaviour as listed in our serverless inference docs, read 23 October 2026.
The two GLM-5.3 Instant IDs currently write their reasoning into the answer text. A tool that hides reasoning would show it to your users. For GLM-5.3 itself, see our GLM-5.3 page. For running other GLM versions in Europe, see Running GLM 5.1, 5.2 and 5.2 Instant in Europe.
When to choose GLM-5.2 instead
Choose GLM-5.2 when the task needs planning over many steps and your client shows or handles the reasoning trace. Give it a `max_tokens` value that leaves room for both the thinking and the answer.
Choose GLM-5.2 Instant when your tool caps the answer and hides reasoning, or when you want answers without the wait. Price and context are the same, so neither choice costs more per token.
Your next step: create an API key and send the same five prompts to both IDs. Compare when each answer starts and how many output tokens you were billed.