GLM 5.2 leaves Mistral on Oct 31: what to move and test

What is retiring, and what the vendor says to use

Mistral's changelog lists Z.ai GLM 5.2 (zai-glm-5-2) as deprecated, retiring on 31 October 2026, and says to use Z.ai GLM 5.3 (zai-glm-5-3) instead, at the same price.¹ GLM 5.3 became generally available on Mistral on 28 September.¹ The page does not say how a request to the retired id will fail, so plan for an error and do not assume an automatic redirect.

GLM 5.3 Flash (zai-glm-5-3-flash) arrived on 22 October. Mistral describes it as a multimodal mixture-of-experts model with 320B parameters, 18B active and a 1M-token context, at $0.15 input, $0.03 cached input and $0.50 output per million tokens.¹ It is not the named successor. It is a different model, and the changelog gives no rate for zai-glm-5-3, so compare the two on your own bill before choosing.

This is a different class of problem from the usual calendar entry. The model is a third party's, served under a host's id, and the host sets the date. INFRO's own tracker lists it as a Mistral shutdown, with the date four days out when it was last updated.²

Find every call, including the ones that are not in code

Hosted third-party ids are spelled differently per platform. Mistral uses zai-glm-5-2. Alibaba Model Studio's recommended list shows GLM 5.3 as both glm-5.3 and ZHIPU/GLM-5.3.³ Perplexity's Agent API uses provider/model ids such as openai/gpt-5.4.⁴ Search for all spellings:

rg -n -i 'zai-glm-5-2|glm[-_. ]?5[-_.]?2|zhipu/glm' \
  --glob '!node_modules' --glob '!*.lock' .
# also: .env files, Helm values, feature-flag exports, prompt/config tables

Code is the easy half. The harder places are a config table, a feature-flag payload, an eval harness, a notebook and a vendor's no-code workflow that stores a model string. Then check traffic, not just source: group the last 30 days of requests by model id in your gateway or provider logs. A string that appears there but not in your repo is a caller you have not found. Perplexity's October retirement notice told users to update "direct model selections and fallback chains" before the cutoff, which is the same two places to look.⁴

Choose the successor and test it first

Start with zai-glm-5-3, since the vendor names it. Add zai-glm-5-3-flash only if you want the lower-rate tier and are willing to qualify it separately. Before moving traffic, replay a sample of real prompts against each candidate and check:

• Output shape: JSON validity, tool-call arguments and refusal behaviour on your own schemas.
• Length and latency: time to first token and total tokens per task, since a model that answers longer costs more at the same rate.
• Cost per task: use the list rates your provider bills, then your contract rate if you have one.
• Region and key: Model Studio regions differ in endpoints, supported models and pricing, and API keys are not interchangeable across regions.⁵ A model id that works in one region may not exist in another.

Put a failover rule in front

A migration done in one deploy has no way back. A rule in front of the call lets you change the model without shipping code, and keeps one working path if the new one misbehaves.

With INFRO, name a route in your application and put the model choice in the route's targets. Priority targets are tried in order, and weighted targets let you send a small share to the candidate first.⁶ Failover applies before the first token; a provider 5xx, timeout, 429 or even a 404 moves the request to the next target.⁷ If you call a model directly, a request-level fallbacks array takes up to five catalog model ids, tried in order once the primary has failed.⁷

from openai import OpenAI

client = OpenAI(base_url="https://api.infro.io/v1", api_key="INFRO_API_KEY")
resp = client.chat.completions.create(
    model="<primary: tested GLM 5.3 id>",
    messages=[{"role": "user", "content": "..."}],
    # ids below are placeholders; use catalog ids from /providers
    extra_body={"fallbacks": ["<second: Flash candidate>"]},
)

fallbacks is ignored when model names a route, so use one mechanism per call site.⁷ A fallback is cross-model: your outputs will change when it fires. That is why the replay test comes first. For teams that already record retirements in INFRO, the model-lifecycle capability warns organizations still sending a retiring model and, from the shutdown date, refuses it before any provider call, naming the replacement.

After the date

Check three things on 1 November: no request still names zai-glm-5-2, the fallback is not carrying more than its expected share, and cost per task moved the way your replay predicted. The same checks apply to the next hosted-model retirement. See the deprecation calendar for dated entries and the migration playbook for the general sequence.