Benchmarking GLM-5.2 Fast: 100 Requests Across Chat, Coding, and Math

GLM-5.2 Fast is now available through AIHubMix. We benchmarked it against GLM-5.2 using real prompts and a production API endpoint.

TL;DR

Across three 100-request benchmark categories, GLM-5.2 Fast delivered:

• 83-94% higher per-user throughput
• 53-77% higher system throughput
• 41-46% lower inter-token latency
• 38.5-42.5% lower median end-to-end request latency

Its clearest advantage appears after generation begins: output streams faster and complete responses finish sooner. This makes it especially useful for real-time chat, coding agents, streamed long-form generation, and multi-step agentic workflows.

Benchmark setup

We used three real-prompt datasets representing different workloads:

Category	Dataset
Chat	ShareGPT
Coding	instruct_coder
Math reasoning	AIMO

The benchmark configuration was:

• 100 requests per category for each model
• Identical prompt samples selected with random seed 42
• Concurrency: 5
• Streaming responses
• Server-reported token counts
• Production AIHubMix chat completions endpoint
• Benchmark runner: aiperf

The models tested were alicloud-glm-5.2 and alicloud-glm-5.2-fast-preview.

Using the same seed ensures both models receive the same sampled prompts. It does not eliminate normal infrastructure variation, but it prevents one model from receiving an easier prompt set.

Result 1: Per-user throughput increased by 83-94%

Per-user throughput is the number of output tokens delivered to one user per second after generation begins. It is the metric most closely related to how fast a streaming response feels.

Workload	GLM-5.2	GLM-5.2 Fast	Improvement
Chat	50.2 tok/s	97.2 tok/s	+94%
Coding	53.3 tok/s	98.1 tok/s	+84%
Math reasoning	61.4 tok/s	112.7 tok/s	+83%

Across the three categories, GLM-5.2 Fast delivered about 1.8-1.9x the per-user throughput of GLM-5.2.

Result 2: System throughput increased by 53-77%

System throughput measures the total number of output tokens generated per second across all concurrent requests. It is useful for understanding aggregate serving capacity under load.

Workload	GLM-5.2	GLM-5.2 Fast	Improvement
Chat	158.6 tok/s	242.5 tok/s	+53%
Coding	245.6 tok/s	416.3 tok/s	+70%
Math reasoning	262.4 tok/s	463.6 tok/s	+77%

These values were measured with concurrency set to 5.

Result 3: Inter-token latency fell by 41-46%

Inter-token latency measures the average delay between consecutive streamed output tokens. Lower values produce a denser, more continuous response stream.

Workload	GLM-5.2	GLM-5.2 Fast	Reduction
Chat	20.58 ms	12.12 ms	-41%
Coding	18.93 ms	10.31 ms	-46%
Math reasoning	16.40 ms	8.93 ms	-46%

This was the most consistent low-level speed signal in the benchmark.

Median end-to-end latency fell by 38.5-42.5%

End-to-end latency measures the complete request duration, from sending the request until the streamed response finishes.

Workload	GLM-5.2	GLM-5.2 Fast	Reduction
Chat	7.68 s	4.73 s	-38.5%
Coding	29.20 s	17.84 s	-38.9%
Math reasoning	33.41 s	19.22 s	-42.5%

These figures are for individual model requests, not complete agent tasks. In a multi-step workflow, however, time saved on sequential model calls can compound across the full task.

What about time to first token?

We measured TTFT, but the P50 results did not move consistently in one direction:

• Chat: 1.6% slower
• Coding: 2.9% faster
• Math reasoning: 6.4% faster

Tail results also varied by workload. We are therefore not making a blanket claim that GLM-5.2 Fast always produces the first token sooner.

The more reliable conclusion is that once generation starts, tokens arrive faster and the complete response finishes sooner.

Best use cases

GLM-5.2 Fast is particularly well suited to:

• Real-time AI chat, where faster streaming directly improves perceived responsiveness
• Coding agents, where long code and reasoning outputs can dominate task duration
• Multi-step agentic workflows, where latency savings accumulate across sequential calls
• High-concurrency applications, where system throughput affects aggregate capacity
• Streamed long-form generation, where lower inter-token latency creates a more continuous output stream

Production fallback configuration

The GLM-5.2 Fast route is currently provided by Alibaba Cloud only. For production traffic, we recommend:

Primary:  glm-5.2-fast-preview
Fallback: glm-5.2

Key-level fallback can route eligible requests to standard GLM-5.2 if the Fast route is temporarily unavailable. It reduces exposure to a single upstream provider, but it is not a guarantee of uninterrupted service. Production systems should still implement timeouts, retries, monitoring, and error handling.

Pricing and model details

At the time of publication, the model page lists:

Item	Value
Input price	$2.254 per 1M tokens
Output price	$7.889 per 1M tokens
Context length	1M tokens
Fast-route provider	Alibaba Cloud

Pricing and availability may change. Check the model page before deploying.

Benchmark caveats

For the Fast coding run, 100 requests were issued, while 99 responses contained usable server-reported token counts for token-derived metrics.

Results also depend on prompt distribution, output length, concurrency, network conditions, provider load, and test timing. Treat these numbers as evidence from this controlled test, not a guarantee for every request.

Try GLM-5.2 Fast

View current pricing, provider availability, and API details on the GLM-5.2 Fast model page.

For production traffic, enable fallback to GLM-5.2 before routing critical workloads to the Fast route.