Z.ai · open weights

GLM-5.2 API

A 1M-token context window, an MIT licence and strong published reasoning scores.

This model has been retired from the router. No upstream serves it any more, so requests for zai-org/GLM-5.2-FP8 no longer resolve on any plan. The prices and benchmarks below are kept for reference — see the models we serve now.

Input / 1M
$0.38
Output / 1M
$1.19
Blended (4:1)
$0.542

Last billed rates before retirement — no longer purchasable. Prices verified November 9, 2026. Full market table on /pricing.

How it was called

Model id: zai-org/GLM-5.2-FP8. This id no longer resolves — the snippet is kept for reference.

from openai import OpenAI

client = OpenAI(
    base_url="https://router.mingles.ai/v1",
    api_key="sk-your-free-key",
)

resp = client.chat.completions.create(
    model="zai-org/GLM-5.2-FP8",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

Published benchmarks

Benchmark Score As published
SWE-bench Pro 62.1 —
Terminal-Bench 2.1 (Terminus-2) 81.0 not comparable to the 2.0 figures elsewhere
AIME 2026 99.2 —
GPQA-Diamond 91.2 —
HMMT Feb. 2026 92.5 —
IMOAnswerBench 91.0 —
MCP-Atlas (public set) 76.8 —
Tool-Decathlon 48.2 —

Source: Z.ai — GLM-5.2-FP8 model card. Full comparison with sources on /compare/claude-api.

Specs

Architecture Mixture-of-experts with sparse attention, 753B total parameters (per model card; active count not published)
Context 1M tokens (per model card)
Precision FP8 (F8_E4M3); the card also ships F32 and BF16 weights
Strengths Long-horizon reasoning and coding; selectable thinking-effort levels
License MIT
Availability Paid plans only — free keys get a 402 on this model id

On a plan

Plans hold a priced allowance, not a token count, so what it buys depends on the model. While GLM-5.2 was served, the allowance worked out to roughly:

Mini · $9/mo
up to ~41M tokens
Standard · $29/mo
up to ~147M tokens
Team · $79/seat
up to ~437M tokens

On its own card GLM-5.2 posts the strongest reasoning numbers of anything we serve — 62.1 on SWE-bench Pro against 58.6 for Kimi and 56.2 for MiniMax, and 99.2 on AIME 2026 — with the usual caveat that each figure comes from its own vendor under its own harness, so they are three claims rather than one comparison. One genuine comparison does exist: DeepSeek ran GLM-5.2 under its own harness and beat it on all seven agentic rows both models completed, at about a seventh of the price. So the case for GLM is now the 1M-token context and the reasoning scores, not throughput — and both cards agree on its 81.0 Terminal-Bench 2.1, which is the one number here nobody has to take on trust. It is on paid plans only, and it spends a plan allowance at about ×5.4 — roughly Kimi's rate.

FAQ

▸How do I call GLM-5.2?

POST /v1/chat/completions with "model": "zai-org/GLM-5.2-FP8" (case-sensitive) against https://router.mingles.ai/v1. Any OpenAI SDK works. The model id resolves only for keys on a paid plan.

▸Why is GLM not on the free plan?

It is served through a different upstream than the rest of the catalogue, and that upstream bills us per token. Free keys get the models we serve ourselves; paid plans get the full catalogue plus failover when the free network is saturated.

▸GLM, DeepSeek, Kimi or MiniMax?

DeepSeek V4 Flash for volume — it is the cheapest and the model plan sizes are quoted in. MiniMax in the middle, Kimi when a task needs more depth. GLM when you want the 1M-token context or the reasoning scores on its card, and you are on a paid plan. Burn rates in the console: DeepSeek ×1, MiniMax ×2.9, GLM ×5.4, Kimi ×5.7.

▸What does the 1M-token context mean in practice?

You can put a large codebase or a long conversation history in one request without chunking. Note that you pay for every token you send on every turn — a long context is a real cost, not a free upgrade.