GLM-5.2

Zhipu AIOpen weights

The strongest open-weights model. Five hosts serve it at different quantizations — headline scores only tell you about the official serving.

Index @ Medium

69.9

Context

200K

$/M in · out

$0.60 · $2.20

Median speed

95 tok/s

Released

May 20, 2026

02·Which effort for which task

The sweet spot, per task

Cheapest effort tier that keeps ≥96% of this model's peak quality on the benchmarks behind each task type. Below the floor number, quality is not okay — it just fails quietly.

Everyday writing & Q&A

Medium

Chat, drafting, summaries, general questions.

keeps 97.5% of peak−87% cost

Code assist

High

Autocomplete, single functions, small fixes.

keeps 98.3% of peak−62% cost

Coding agents

High

Real repos, multi-file changes, end-to-end tickets.

keeps 96.5% of peak−62% cost

⚠ at Low: only 78% of peak — avoid

Math & quantitative

X-High

Derivations, statistics, financial models.

keeps 100.0% of peakpeak tier needed

⚠ at Low: only 66% of peak — avoid

Hard research problems

High

Novel, frontier, genuinely difficult questions.

keeps 96.0% of peak−62% cost

⚠ at Low: only 75% of peak — avoid

Agents & tool use

High

Long workflows with tools, policies, and state.

keeps 97.8% of peak−62% cost

03·The effort ladder

Quality and cost, tier by tier

On everyday tasks, Medium keeps 97.5% of X-High quality at $0.0072 vs $0.055 per task.

Intelligence Index by effort

Cost per typical task by effort

04·Benchmark detail

Scores at Medium

Low Medium High X-High

at Mediumat X-High (peak)

Benchmark	Medium	Peak	What it predicts
MMLU-Pro	83.4	86.0	Everyday Q&A · Drafting emails & docs · Summarizing articles
GPQA Diamond	74.8	85.0	Technical analysis · Root-cause debugging · Scientific writing
Humanity's Last Exam	22.9	26.0	Novel research questions · Frontier problem solving
AIME 2025	71.8	92.0	Financial modeling · Statistics · Algorithm design
SWE-bench Verified	69.3	77.0	Coding agents · Bug fixing in real repos · Multi-file refactors
LiveCodeBench	72.9	81.0	Writing functions from scratch · Algorithms · Code autocomplete
τ²-bench	73.0	83.0	Tool-using agents · Customer-facing automation · Multi-step workflows
IFEval	88.2	90.0	Structured output (JSON) · Templated generation · Batch pipelines

05·Host variants

Same weights, different model

Hosts serve GLM-5.2 at different quantizations and stacks. Index below is measured per host at Medium effort — the headline number only applies to the reference serving.

Host	Quant	Index	Δ vs reference	$/M in · out	tok/s	TTFT	Uptime	Gateway id
Z.ai (official)reference	bf16	69.9	—	$0.60 · $2.20	95	0.45s	99.5%	glm-5.2
Fireworks	fp8	69.5	-0.3	$0.55 · $2.19	210	0.30s	99.3%	glm-5.2@fireworks
Together AI	fp8	69.2	-0.7	$0.59 · $1.99	170	0.35s	99.0%	glm-5.2@together
DeepInfra	int8	67.9	-2.0	$0.40 · $1.60	120	0.60s	98.2%	glm-5.2@deepinfra
SiliconFlow	fp8	69.0	-0.8	$0.45 · $1.80	140	0.55s	97.8%	glm-5.2@siliconflow

06·Drift

Is it still the same model this week?

Z.ai (official)FireworksTogether AIDeepInfraSiliconFlow

Weekly Intelligence Index · dashed rules mark serving-change events

May 31, 2026DeepInfra-2.4 ptsrecovered Jun 21, 2026

Quantization silently switched int8 → int4 during a capacity crunch; reverted after three weeks of community reports. The canonical "降智" case.

07·Call it

One key, this model, your effort

The gateway speaks the OpenAI Chat Completions dialect. Pick a host with glm-5.2@host syntax, set reasoning_effort, and the sweet spot above becomes one line of config. Full docs →

chat with GLM-5.2

curl https://models.fyi/api/v1/chat/completions \
  -H "Authorization: Bearer $MFYI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.2",
    "reasoning_effort": "medium",
    "messages": [{"role": "user", "content": "Hello"}]
  }'
# local dev: replace https://models.fyi with http://localhost:3000