MoeRouter Blog
English
Get an API key

GLM-5.3-Flash review roundup: a natively multimodal model at a fraction of a cent, and where it falls short

Zhipu's GLM-5.3-Flash is the first natively multimodal GLM-5 model, with image and video input and a 1M-token context. Vendor scores, a Chinese benchmark and Bilibili creators find solid back-end code; complex flows need watching.

glm-5.3-flash

In short

  • GLM-5.3-Flash was released with open weights around August 27, 2026. It is the model that had gone viral anonymously on OpenRouter as Ox Alpha. It is the first natively multimodal GLM-5 model: text, image and video input, with a 1M-token context.
  • On ReLE, an independent Chinese benchmark, it scores 74.2% and ranks 15th, one point behind Zhipu's own flagship GLM-5.3 at under a tenth of the cost per call.
  • Hands-on tests agree: solid back-end and test code and excellent value. Front-end detail, deep reasoning and keeping state consistent in complex business flows are the weak spots, and it is slow.
  • glm-5.3-flash was the fourth most used model on MoeRouter over the last 7 days and sits in the cheapest tier. It suits high-volume work that can be checked automatically.
On this page
  1. What GLM-5.3-Flash is
  2. What it costs on MoeRouter
  3. Benchmarks
  4. Hands-on
  5. How to use it
  6. How to call it
  7. Frequently asked questions

What GLM-5.3-Flash is#

In late August 2026 an anonymous model called Ox Alpha appeared on OpenRouter: a 1M-token context, text, image and video input, and free. It went straight to the top of the usage chart. A few days later Zhipu confirmed it was the new version of its GLM line, then released it with open weights as GLM-5.3-Flash[1].

GLM-5.3-Flash
Positioning The first natively multimodal GLM-5 model, for code, agents and vision at lower cost
Input and output Text, images and video in, text out, with tool calling
Size 320B parameters in total, 18B active
Context and output 1M-token context, up to about 131K output tokens
List price on Z.ai per 1M tokens $0.03 cached input, $0.15 input, $0.50 output
Model name on MoeRouter glm-5.3-flash
Endpoints on MoeRouter Chat Completions, Responses, Messages

Specifications and list prices are from 子昕's article in the Volcengine developer community[1].

What it costs on MoeRouter#

per 1M tokens
ModelGroupInputOutputCache read
glm-5.3-flashdefaultCN¥0.0072≈ US$0.00108CN¥0.0252≈ US$0.00378CN¥0.00207≈ US$0.000311

Reference prices per 1M tokens in Chinese yuan (CNY), the currency accounts are charged in, from the model square (). Dynamically priced models settle at the route that served each request; the model square is authoritative. Amounts in USD are estimates, converted at 1 CNY = 0.15001 USD ().

This is the cheapest tier on MoeRouter. The price is this low because the default group shows the reference price of the cheapest route at the moment; it moves when the routes do, and the usage log shows what each request was charged. See billing.

Benchmarks#

Reported by Zhipu#

Benchmark GLM-5.3-Flash For comparison
DeepSWE v1.1 63.4 GLM-5.2 46.2
Terminal Bench 2.1 84.3 —
AutomationBench v1.0.6 48.8 GLM-5.2 26.2
Z.ai Code Bench v1.0 (max effort) 29.0 Claude Opus 4.8 29.5
OfficeQA Pro (complex documents) 62.4 Claude Opus 4.8 48.9

Zhipu compares it with models of the Claude Opus 4.8 and GPT-5.6 Terra generation[2]. Its blog also quotes 57 on the Artificial Analysis Intelligence Index; 子昕's caution is that a leaderboard score comes from particular test conditions and cannot be converted into a success rate for a project[1].

Measured by third parties#

  • ReLE, NoneLinear's Chinese-language benchmark of about 15,000 questions across education, medicine, finance, law, reasoning and maths, instruction following, agent tool calling and coding: 74.2% overall, 15th place, where the previous GLM-4.7-Flash scored 49.8% and ranked 119th. Coding rose most, from 22.9% to 64.7%. The one item that slipped is agent tool calling, from 62.7% to 61.6%[2].
  • Against its price class. Among models costing ¥10 to ¥15 per thousand calls it (74.2%, ¥12.7) edges out qwen3.8-flash (74.0%, ¥10). It costs more than deepseek-v4-flash (68.8%, ¥4.9) and scores 5.4 points higher[2].
  • Against the flagship. It is 1.0 point behind Zhipu's own glm-5.3 (75.2%, ¥134.9) at under a tenth of the cost[2].
  • Slow. On ReLE it averages 128 seconds per call where gemini-3.7-flash takes 10[2]. In Atomic Agent's coding-agent test it passed all 96 hidden tests but took 14.1 minutes, 34 steps and $0.059 per task on average; DeepSeek V4.1 Flash 1017 did the same task in 27 seconds for $0.016[3].

Hands-on#

子昕 (Volcengine developer community) ran three real scenarios in Zhipu's own ZCode, the same ones used earlier on Qwen 3.8 Max and Kimi K3[1]:

Scenario Result
A front-end video workbench Upload and trimming work, and moving a clip snaps into place, which neither of the other two models built; but the layout breaks and the speed and fade controls do nothing
A real-time ticketing system with a waiting list After locking a seat, starting payment throws an error, so issuing, refunds and the waiting list could not be tested; Qwen 3.8 Max and Kimi K3 both got through the main flow
A physics driving game rebuilt from a gameplay video The best of the three: driving forward, reversing and mid-air control all work; the problem is the tuning of the controls

The conclusion, in our translation: the model can produce a first version; consistent state, interaction detail and repair after a failure still need a person watching[1].

程序员阿江-Relakkes (Bilibili) found the strengths concentrated in back-end and test code. His summary: good for 18B active parameters and solid on the back end, somewhat lacking on the front end and in deep reasoning, and very good value[4].

林禾Link (Bilibili) compared it with DeepSeek without rounds of fixes, admits each project had small defects, and finds Zhipu clearly stronger on feature completeness, understanding of the project and self-correction, with a working vision loop in the official harness[5].

塔斯海Tasihi (Bilibili) did not believe Zhipu's claim of Opus 4.8's ability at a fortieth of the price, so he built three front-end projects with identical prompts against DeepSeek-V4-Flash, and found the results surprisingly good and the price lower than even V4-Flash's off-peak rate[6].

How to use it#

Good for

  • Back-end endpoints, test code and scripts: work that can be checked automatically
  • High-volume extraction, classification, summaries and Chinese question answering
  • Tasks that read images or video, which it supports natively

Less suited to

  • Business systems with complex state (flows such as payment and refunds need a human check)
  • Front ends where detail matters
  • Interactive use that needs a quick answer: it is on the slow side

The other choice at this price is DeepSeek V4.1 Flash: on the same coding task in Atomic Agent's test it was far faster. GLM-5.3-Flash's advantages are video input and its overall score in Chinese.

How to call it#

With the OpenAI SDK:

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://moerouter.com/v1",
    api_key=os.environ["MOEROUTER_API_KEY"],
)
reply = client.chat.completions.create(
    model="glm-5.3-flash",
    messages=[{"role": "user", "content": "Add unit tests for this function's edge cases, using pytest"}],
)
print(reply.choices[0].message.content)

It also answers on the Messages and Responses endpoints, so it can be set as the model in Claude Code or Codex. Use streaming for requests that take long; see endpoints and protocols. The docs are in Chinese for now.

Frequently asked questions#

Is GLM-5.3-Flash the anonymous Ox Alpha?#

Yes. Before launch it was tested anonymously on OpenRouter as Ox Alpha; Zhipu then confirmed it was the new version of its GLM line and released it as GLM-5.3-Flash.

How much does the GLM-5.3-Flash API cost?#

Zhipu's list price on Z.ai per 1M tokens is $0.15 input and $0.50 output, with cached input at $0.03. MoeRouter's reference prices are in the table above; accounts are charged in Chinese yuan.

Does GLM-5.3-Flash accept images and video?#

The model takes text, image and video input and is the first natively multimodal GLM-5 model. Through MoeRouter, image input works directly; video input depends on the route in use, so test with a small file first.

GLM-5.3-Flash or DeepSeek V4.1 Flash?#

Both are very cheap. GLM-5.3-Flash scores higher on the Chinese benchmark and can read video; DeepSeek V4.1 Flash is much faster on coding-agent tasks. Try DeepSeek first for code, and GLM for documents, Chinese question answering and multimodal work.

Why is GLM-5.3-Flash slow?#

Reasoning is on by default and it spends many tokens thinking. It averages 128 seconds per call on the Chinese benchmark, and on one coding-agent task it took 14 minutes, with most of its output being reasoning tokens. It does not suit latency-sensitive use.

Can it replace the GLM-5.3 flagship?#

For most tasks it is worth trying first. The two are one point apart on the Chinese benchmark at under a tenth of the cost; where a task involves a complex business flow, compare on your own cases.

Sources

The test results and figures quoted in this post come from the public sources below and belong to their authors; each quotation is marked where it appears.

  1. 牛来原来是智谱,GLM-5.3-Flash 实测有惊喜也有硬伤 (in Chinese) Volcengine developer community · 子昕
  2. 【非线智能测评】智谱 GLM-5.3-Flash 实测:Coding与多模态大幅跃升,重塑轻量模型性价比标杆 (in Chinese) NoneLinear (非线智能) · published 2026-09-03
  3. DeepSeek V4.1 Flash 1017: $0.016 per Coding Task Atomic Agent
  4. 匿名爆火的「牛来」模型,原来真的是智谱家的!实测GLM-5.3-Flash! (video, in Chinese) Bilibili · 程序员阿江-Relakkes · published 2026-08-27
  5. 夯爆了!全面超越Deepseek!智谱GLM5.3Flash性能实测!效果过于惊艳! (video, in Chinese) Bilibili · 林禾Link · published 2026-08-27
  6. GLM-5.3-Flash性价比实测如何?对比V4Flash (video, in Chinese) Bilibili · 塔斯海Tasihi · published 2026-08-27

Disclosure: MoeRouter sells API access to the models discussed here. The test results in this post come from the third-party sources cited; prices and endpoint details come from MoeRouter's own data.