---
title: "GLM 5.3 Flash and SupportBench in Chat Thing"
canonical_url: "https://chatthing.ai/blog/glm-5-3-flash-supportbench-and-link-controls"
last_updated: "2026-08-28T11:55:16.762Z"
meta:
description: "Try GLM 5.3 Flash in Chat Thing, compare seven models using SupportBench, and control whether links in your embedded agent open in a new tab or the same page."
"og:description": "Try GLM 5.3 Flash in Chat Thing, compare seven models using SupportBench, and control whether links in your embedded agent open in a new tab or the same page."
"og:title": "GLM 5.3 Flash lands in Chat Thing, plus SupportBench and link controls - Blog - Chat Thing"
"twitter:description": "Try GLM 5.3 Flash in Chat Thing, compare seven models using SupportBench, and control whether links in your embedded agent open in a new tab or the same page."
"twitter:title": "GLM 5.3 Flash lands in Chat Thing, plus SupportBench and link controls - Blog - Chat Thing"
---
[← Back to blog](https://chatthing.ai/blog)**Changelog**# **GLM 5.3 Flash lands in Chat Thing, plus SupportBench and link controls**
Zef
28 Aug 2026
~ 4 min read
---
---
### **On this page**
1. GLM 5.3 Flash: low cost, mixed support results
2. SupportBench: model comparisons built around support work
3. [One smaller web-channel update: choose where links open](#one-smaller-web-channel-update-choose-where-links-open)
4. [Try the release](#try-the-release)
GLM 5.3 Flash is now available in Chat Thing for users who want to try it. It is the model a lot of people have been talking about this week, but making it available is not the same as recommending it as the default support model.
This release also adds [SupportBench](https://chatthing.ai/models/supportbench), our own customer-support benchmark, and a setting that controls where links in an embedded agent open.
## GLM 5.3 Flash: low cost, mixed support results
GLM 5.3 Flash comes from Z.ai and was previously previewed as Ox Alpha. If that is the name you recognise, this is the same model. It supports images, reasoning and Power-Ups. It was also the lowest-cost model in our first seven-model SupportBench run, at about $0.0005 per correctly resolved benchmark conversation.
That price makes it interesting for high-volume work. It does not make it the model we would switch every busy support agent to without testing.
GLM finished fifth of seven with a score of 81.7. It was good at arithmetic, long conversations and using tools accurately once it decided to act. Its weaker results came from the moments around those actions. It repeated an instruction planted in retrieved content in five runs out of five, left an address change undone in four of five runs, and missed 40% of the times a customer asked for a human.
So why add it? People wanted to try it. Chat Thing lets you choose between models, and GLM may prove useful where volume and model cost matter most. Our current result says to test three things before switching: whether it ignores instructions hidden in retrieved content, whether consequential tool actions complete, and whether requests for a person really escalate.
You can read the full [GLM 5.3 Flash support report](https://chatthing.ai/models/glm-5-3-flash), including transcript excerpts, cost, latency and the failures behind the score.
_GLM 5.3 Flash's [public Chat Thing model report](https://chatthing.ai/models/glm-5-3-flash), captured 27 August 2026._
## SupportBench: model comparisons built around support work
General model benchmarks rarely test whether a support agent will issue the wrong refund, reveal private account details, follow an instruction hidden in a help-centre page or claim it escalated a conversation without calling the tool. We built SupportBench around those failures.
Seven models each handled 31 scripted support conversations five times, giving us 1,085 scored conversations under the same synthetic knowledge, customer details and tools. Wrong refunds, data leaks and actions that never happened score zero. Two blind LLM graders then score answer quality against a written answer key, and their mean becomes the SupportBench score.
The main score put Gemini 3.7 Flash first. The leading results were extremely close, so we also compared the replies side by side to see which one a grader would rather send to a customer.
| Model | SupportBench score | Cost per correctly resolved benchmark conversation |
| --- | --- | --- |
| Gemini 3.7 Flash | 88.6 | $0.0035 |
| Grok 4.6 | 86.5 | $0.0149 |
| Claude Sonnet 5 | 86.0 | $0.0247 |
_Gemini led the main score. Grok's replies won the most decided side-by-side comparisons, while Claude was the most grounded when the supplied content was untrusted._
The [full SupportBench methodology and leaderboard](https://chatthing.ai/models/supportbench) include confidence intervals, per-model reports, transcript excerpts and limitations. We have also written a separate article about what the 1,085 conversations found and the failures that changed how we think about choosing a support model.
## One smaller web-channel update: choose where links open
Links in an embedded Chat Thing agent open in a new tab by default. That remains the safest choice for most agents, especially when they regularly link to other websites.
You can now change this to **Same page as the embed**. It is useful when an agent points visitors to another section of the website they are already using. Instead of opening a new tab, the link navigates the page containing the embedded agent.
You will find the setting under **Advanced features** in your web channel settings:
- **New tab (default)** opens every link in a separate browser tab.
- **Same page as the embed** navigates the page containing the embedded agent.
Same-page behaviour only applies to full `http://` or `https://` addresses. A short link such as `#pricing` or `/pricing` still opens in a new tab because it would otherwise resolve inside the chat frame rather than against your website. If you use the same-page option, tell your agent to include your full domain in its links, for example `https://yoursite.com/#pricing`.
Any full address can navigate the visitor away from the current page, including links to other sites. Keep the default if your agent often sends people elsewhere, and test one internal link and one external link before changing a live embed.
## Try the release
Open an agent and head to its advanced model settings to try GLM 5.3 Flash. Model selection is available on plans that include the model picker.
For link behaviour, open your web channel settings, expand **Advanced features**, and choose where links should open.
Use [SupportBench](https://chatthing.ai/models) to shortlist models, then run the questions, tools and difficult cases your own customers bring you. The leaderboard can narrow the choice. Your real conversations should make the final call.
### **Related articles **
[](https://chatthing.ai/blog/gemini-3-7-flash-and-grok-4-6) [Gemini 3.7 Flash and Grok 4.6 land in Chat Thing
](https://chatthing.ai/blog/gemini-3-7-flash-and-grok-4-6)
Gemini 3.7 Flash and Grok 4.6 are now available in Chat Thing, alongside clearer Power-Up activity and support for more Google Drive file types.
[](https://chatthing.ai/blog/ask-otto-and-seven-new-models) [Otto now lives inside every agent, plus seven new models
](https://chatthing.ai/blog/ask-otto-and-seven-new-models)
Otto now lives inside every agent, so you can change things by describing what you want rather than hunting through settings forms. Plus Claude Opus 5 and six more models, four new embed capabilities, and twelve retired models.
[](https://chatthing.ai/blog/kimi-k3-lands-in-chat-thing) [Kimi K3 is live in Chat Thing: what it is, and is it actually good?
](https://chatthing.ai/blog/kimi-k3-lands-in-chat-thing)
Moonshot AI's Kimi K3 launched today and it's already live in Chat Thing. Here's what Kimi K3 is, the benchmarks behind the hype, and how it performed when we ran it head-to-head against Claude Sonnet 5, GPT-5.6 and Gemini on real support-bot prompts.
- **Previous **
- [← Turn Your Notion Workspace into an AI-Powered Assistant](https://chatthing.ai/blog/notion-power-up)