GLM-5.3-Flash: Benchmarks, Price, Weights & API

Z AI · Open weights

The top-10 outlier: open weights, native vision, near-frontier agent scores, and Flash-class serving economics in the same release. Source-checked August 28, 2026.

Compare this model Open live ranking

Live rank: #9
Intelligence v4.1: 57.5
Best blended / 1M: $0.119
Context: 1.05M

The answer in 60 seconds

The first model in this wave to test for cost-sensitive agents

GLM-5.3-Flash is the practical surprise in the top ten. It gives up some coding depth versus the flagship, but its observed price, output speed, native vision, open weights, and strong agentic score create a rare deployment envelope. It is the obvious first benchmark for high-volume agents—provided its weaker SciCode signal does not map to your workload.

Best for

— High-volume tool-using agents and coding loops
— Multimodal workflows that need image input and long context
— Open-weight deployments where throughput and cost matter more than the last capability points

Look elsewhere when

— Your hardest scientific coding tasks expose its lower SciCode result
— You need the flagship’s maximum coding quality more than serving efficiency

Independent data

What the current measurements say

Metrics come from the live Artificial Analysis-backed WhatLLM feed and refresh hourly. Prices and speed are best observed values across matching endpoints; they can come from different providers.

Intelligence Index: 57.5
Coding Index: 71.5
Agentic Index: 58.2
Input / output per 1M: $0.075 / $0.250
Fastest observed: 296.5 tok/s
Lowest observed TTFT: 0.415s

Benchmarks worth reading together

The strongest indexed configuration is shown. Different tests measure different jobs; no single score is a universal model verdict.

GDPval-AA Elo: 1769
GDPval-AA: 64%
GPQA Diamond: 91%
Humanity's Last Exam: 40%
SciCode: 46%
AA-LCR: 78%

Source truth

What GLM-5.3-Flash actually is

Architecture: 320B total / 18B active; hybrid sparse and linear attention; Provider-confirmedSource ↓
Context: 1 million tokens; Provider-confirmedSource ↓
Modalities: Video, image, text, and file input; text output; Provider-confirmedSource ↓
Weights: Official checkpoint and license published on Hugging Face; Provider-confirmedSource ↓
Promotional API price: $0.075 input / $0.25 output per 1M tokens through September 9, 2026; Provider-confirmedSource ↓

Editorial analysis

The decision behind the ranking

We have not claimed hands-on use here. The analysis below combines provider-confirmed facts with the independent data shown above; each section labels that distinction.

Provider-confirmed

Flash is not just a smaller GLM-5.3 endpoint

GLM-5.3-Flash uses a different 320B/18B-active hybrid architecture and adds native multimodality. Z.ai says sparse plus linear attention cuts attention compute and KV-cache requirements substantially versus the flagship. That design choice shows up in the live provider data: Flash is dramatically faster and cheaper while remaining in the same broad frontier band.

This is a separate deployment decision, not an effort setting. Use the flagship for the hardest text-only coding; use Flash when the system needs to look, act, and repeat economically.

Independent signal

The benchmark shape matters more than the average

Flash’s broad and agentic indices are strong, but SciCode is a relative weak spot in the current data. That is exactly the kind of shape an overall ranking hides. A document agent, UI coding loop, or browser worker may love the speed and vision; a scientific-computing assistant needs a much more skeptical trial.

Boundaries

What the headline score misses

01Scientific coding is a weaker signal than the overall rank suggests.
02Reasoning is always enabled on the documented API.
03The current API discount expires September 9; list price is $0.15 input and $0.50 output per million tokens.
04Large open weights still require serious server hardware.
05No dedicated GLM-5.3-Flash safety or system card surfaced in the provider materials reviewed.

Alternatives

The next models to test

Compare the shortlist →

GLM-5.3

A low-priced, agentic coding specialist whose strongest release is hosted rather than open weight.

Qwen3.8 2.4T A95B

Near-Max Qwen capability with downloadable weights—and a 2.4T-scale infrastructure bill hidden behind the word “open.”

GPT-5.6 Terra

The balanced GPT-5.6 tier: much of Sol’s coding capability at half the published token price and materially higher throughput.

Evidence register

Sources and claim status

Checked August 28, 2026. Pricing and availability should be verified again before procurement.

Direct answers

Questions people ask about GLM-5.3-Flash

Is GLM-5.3-Flash open source?

Z.ai publishes the model weights and a license on its official Hugging Face repository. “Open weights” is the more precise description; review the license for your use.

What is the difference between GLM-5.3 and GLM-5.3-Flash?

The flagship is a proprietary, text-only, always-reasoning endpoint optimized for maximum coding and agent quality. Flash is a distinct 320B/18B-active open-weight multimodal model optimized for serving efficiency.

Why is GLM-5.3-Flash so cheap?

Its hybrid sparse/linear-attention architecture activates a small share of total parameters and reduces attention and cache costs. Z.ai is also running a temporary API discount through September 9, 2026; compare against list price for longer-term planning.

Assembling the frontier