Moving from GLM 5.2 Fast
GLM-5.2-Fast is withdrawn from new requests. Its past usage and billing history
remain available. See retired model migration.
Choose a replacement explicitly. Primex preserves public model IDs and billing history; existing requests and saved selections are not automatically renamed. Check live availability and prices before switching.
Choose by workload
| Workload | Starting point | What to check |
|---|---|---|
| Ordinary text chat and streaming | GLM-5.3-Flash | Expected answers, output budget and final streaming usage. |
| Client-executed function tools | GLM-5.3-Flash or GLM-5.3 | Tool selection, argument validation and your full tool-result conversation, including the tool-choice modes your application uses. |
Strict response_format.json_schema | GLM-5.3 | Validate the returned object against your schema and allow enough completion tokens for reasoning and visible output. JSON object mode and strict JSON Schema are different contracts. |
| Long prompts or cached prefixes | Either candidate after a workload check | Context limits, complete output and actual cache usage. A past cache hit does not guarantee a hit for your prefix. |
Compare actual output and charges for your application before moving its traffic. Prices and cache rates differ; use the live catalog and your Usage.
Data-handling requirements
Switching models does not preserve data-location or processing commitments. If your workload requires a specific region or residency commitment, confirm the replacement's terms with Primex before sending data. Review the Privacy Policy and team data-handling guidance.
Change the selected model
For an OpenAI-compatible client, keep your Primex base URL and key, and choose the replacement using its public Primex ID:
import os
from openai import OpenAI
client = OpenAI(
base_url="https://gateway.primexos.com/v1",
api_key=os.environ["PRIMEX_API_KEY"],
)
response = client.chat.completions.create(
model="GLM-5.3-Flash",
messages=[{"role": "user", "content": "Reply with one short greeting."}],
max_tokens=1024,
)
print(response.choices[0].message.content)
This is a normal billable generation when you run it. Review reasoning budgets and billing when choosing the completion limit. Use standard assistant/tool message fields rather than copying provider-specific response extensions back into request messages.
In Primex AI, select the replacement for the conversation you intend to move. Review separately configured presets, assistants and API clients; changing one selection does not update the others.
Review before moving more traffic
Confirm usable final output and the expected stop reason, not just HTTP 200. For tools, execute only validated arguments and check the subsequent tool-result turn. For structured output, validate your schema even when JSON parses. For streaming, check completion and authoritative usage rather than treating a partial answer as a finished response.
Verify the selected model and charge in Usage, then update only the reviewed client or conversation. Keep the earlier configuration for comparison while its selected model remains available. A retired model is not a dependable rollback.