What are the strengths of this model?

Priced ~1/40 of Claude Opus 4.8 with comparable Artificial Analysis intelligence First native-multimodal GLM-5 (image + video input) Sparse + linear attention hybrid cuts 1M-context serving cost MIT license: fine-tune, resell, or self-host with no constraints

What are the weaknesses of this model?

4-bit self-host still needs ~190GB of combined RAM/VRAM — not a single-GPU model The 100k-domestic-chip serving claim is unverified and reads as marketing AA Index 57 is frontier-adjacent, not frontier-leading on the hardest reasoning and agentic tests Launch promo ($0.075 input) expires Sep 9; the real baseline is $0.15/$0.50

What are the best use cases?

High-volume coding, classification, extraction and summarization where cost dominates Document + image multimodal workflows without a second vision model Replacing or downgrading Claude Opus subscriptions at ~1/40 the cost Self-hosted fine-tuning research where the MIT license matters

Zhipu AI (Z.ai)Open Source

GLM-5.3 Flash

Name: GLM-5.3 Flash
Author: Zhipu AI (Z.ai)

Compare this model

GLM-5.3 Flash is Zhipu AI (Z.ai)'s 320B-parameter MoE (18B active) open-weight model, released August 26, 2026. It debuted anonymously as 'Ox-Alpha' and held OpenRouter's weekly top slot before Zhipu confirmed authorship on August 26. A sparse + linear attention hybrid lets it serve a 1.31M-token context cheaply, and it is the first native-multimodal GLM-5 release. Priced at $0.15/$0.50 per million tokens — about 1/40 of Claude Opus 4.8 — it ties Opus 4.8 on the AA Intelligence Index (57) and tops it on the vendor's own Code Bench (63.4). MIT-licensed.

Share:X Hatena LINE

Parameters

320B total / 18B active (MoE, sparse + linear attention hybrid)

Context Window

1.31M (131K max output)

License

MIT

Release Date

2026-08-26

API Pricing

Input Price (per 1M tokens)

$0.15

Output Price (per 1M tokens)

$0.5

Billing Mode: launch promo input $0.075 through 2026-09-09, then standard $0.15/$0.50

Strengths

•Priced ~1/40 of Claude Opus 4.8 with comparable Artificial Analysis intelligence
•First native-multimodal GLM-5 (image + video input)
•Sparse + linear attention hybrid cuts 1M-context serving cost
•MIT license: fine-tune, resell, or self-host with no constraints

Weaknesses

•4-bit self-host still needs ~190GB of combined RAM/VRAM — not a single-GPU model
•The 100k-domestic-chip serving claim is unverified and reads as marketing
•AA Index 57 is frontier-adjacent, not frontier-leading on the hardest reasoning and agentic tests
•Launch promo ($0.075 input) expires Sep 9; the real baseline is $0.15/$0.50

Use Cases

•High-volume coding, classification, extraction and summarization where cost dominates
•Document + image multimodal workflows without a second vision model
•Replacing or downgrading Claude Opus subscriptions at ~1/40 the cost
•Self-hosted fine-tuning research where the MIT license matters

Deep Analysis

Context Window

1.31M tokens

131K max output; native image + video input; first native-multimodal GLM-5 release

API Price

$0.15 / $0.50 per 1M

Input/output per million tokens; launch promo $0.075 input through Sep 9

AA Intelligence Index

Ties Claude Opus 4.8 on Artificial Analysis composite index

Z.ai Code Bench

63.4

Ahead of Claude Opus 4.8 (58.0) and DeepSeek-V4-Vision-Exp (59.3)

License

MIT (open weights)

Weights published on Hugging Face Aug 26, 2026

Parameters

320B total / 18B active

MoE; sparse + linear attention hybrid; first open frontier model with that mix

Strengths

・Price-performance is the headline: ~1/40 the cost of Claude Opus 4.8 at comparable coding quality on Z.ai Code Bench.
・First open-weight GLM-5 model with native multimodal input — can judge when to look and steer with visual feedback.
・Sparse + linear attention hybrid cuts 1M-context serving cost sharply; runs on far less hardware than the parameter count suggests.
・Shipped under MIT, so fine-tuning, resale, and self-hosting need no permission or attribution.

Weaknesses

・Self-hosting is not trivial: a practical 4-bit build still needs ~190GB of combined RAM/VRAM, not a single consumer GPU.
・The 100k-domestic-chip cluster claim behind the stealth Ox-Alpha run is marketing-grade — Zhipu named no hardware and it is unverified.
・Capabilities still trail the top closed flagships on the hardest reasoning and agentic tests; treat 57 AA Index as frontier-adjacent, not frontier-leading.
・Steep launch promo (input at $0.075) expires Sep 9, after which the real $0.15/$0.50 price is the baseline.

Competitor Comparison

Model	Arena	SWE	GPQA	Price
GLM-5.3 Flash	57 (AA Index)	63.4 (Z.ai Code Bench)	N/A	$0.15/$0.50
Claude Opus 4.8	57 (AA Index)	58.0 (Z.ai Code Bench)	N/A	$10/$50
DeepSeek V4 Pro	53 (AA Index)	N/A	N/A	$1.32/$3.96
Qwen3.8-Max-0902	#1 WebDev	70.0 (QwenSWE V2)	N/A	$2/$6

Overview

GLM-5.3 Flash, released August 26, 2026 by Zhipu AI (Z.ai), is a 320B-parameter MoE (18B active) open-weight model that debuted under the anonymous Ox-Alpha handle and promptly topped OpenRouter's weekly charts. Priced at $0.15/$0.50 per million tokens — about 1/40 of Claude Opus 4.8 for near-equal coding quality — it is the clearest signal yet that China's open-weight labs are competing on price-performance, not parameter bragging rights. The model is the first native-multimodal GLM-5 and uses a sparse + linear attention hybrid to make 1M-context serving cheap.

Benchmarks & Performance

Zhipu reports an Artificial Analysis Intelligence Index of 57, tying Claude Opus 4.8, and a Z.ai Code Bench score of 63.4, ahead of Opus 4.8 (58.0) and DeepSeek-V4-Vision-Exp (59.3). On the separate KingBench evaluation it scored 63/80. All figures are Zhipu-reported; the anonymous Ox-Alpha stealth run pulled 62 trillion tokens in its first days on OpenRouter and OpenCode before the reveal.

Detailed Comparison

Against Claude Opus 4.8 it matches composite intelligence at roughly 1/40 the price, which is the entire point — the disruption is to the API subscription, not the GPU. Against DeepSeek V4 Pro it undercuts on price after DeepSeek's own 14x increase. Against Qwen3.8-Max-0902 it trades the coding-leaderboard crown for a far cheaper, MIT-licensed, self-hostable option. The honest framing: GLM-5.3 Flash is the Pareto play, not the peak.

Community Feedback

The Ox-Alpha mystery build drew a grassroots following on OpenRouter and OpenCode for a week before Zhipu confirmed it, ending DeepSeek's 56-day OpenCode lead. Zhipu's Hong Kong-listed shares rose 12% on the reveal. Independent researchers fingerprinted the model to 99% Zhipu before the official claim. The 100k-domestic-chip serving claim is the part to treat as unverified.

Use Cases

Best fit: high-volume coding, classification, extraction, and summarization where a 6-40x cost cut matters more than marginal quality. The native multimodal input makes it usable for document-and-image workflows without a separate vision model. Teams evaluating self-hosting should budget ~190GB (4-bit) and validate the 1M-context latency on their own hardware before committing — the runs-on-hardware-you-own pitch is directionally true but not single-GPU true.

Latest News

Zhipu open-sourced GLM-5.3 Flash on August 26, 2026 under MIT, the same day it revealed Ox-Alpha as the stealth build. The launch promo cuts input to $0.075 through September 9. Zhipu reported H1 2026 revenue of RMB 954M, up ~400% YoY, and says it now serves large-scale inference on a 100k domestic-chip cluster. The model sits at or near the top of OpenRouter's weekly usage charts as of early September.

Sources

Analysis generated: 2026-09-06