Model Drop: GLM 5.2
Z.ai's new open-weight model takes on Opus 4.8 in frontend design
GLM 5.2 is a new open-weight coding and agent model Z.ai (Zhipu AI). It’s the cheapest frontier-class model open model in existence, and, like Kimi, it out-performs the models charging fifteen times more.
Model: GLM 5.2 (glm-5.2 on the Z.ai API, zai-org/GLM-5.2 on Hugging Face). Coding-tuned sibling ships as glm-5.2-air for self-hosters who want a smaller footprint.
Model type: Text + image input, text output. Hybrid reasoning, with a switchable thinking mode.
Ship date: June 20, 2026
Maker: Z.ai (Beijing; spun out of Tsinghua University)
Pricing: $0.80 / $2.50 per million input / output tokens on the Z.ai API, with cache hits at roughly $0.15 per million. The GLM Coding Plan starts at $3 / month for the Lite tier and runs to ~$30 / month for the Max tier with raised rate limits. 12x input gap against Fable 5 and a 6x gap against Opus 4.8. Free weights on Hugging Face for self-hosting.
Available on: The Z.ai chat app, the Z.ai API (OpenAI- and Anthropic-SDK compatible, one-line base URL swap), the GLM Coding Plan wired into Claude Code, Cline, Roo Code, OpenCode, and Kilo Code, Hugging Face and ModelScope for open weights under the MIT license, and OpenRouter for multi-provider routing. Served self-hosted via vLLM and SGLang.
Headline benchmarks: #1 on WebDev Arena, passing Claude Fable 5 and Opus 4.8 on the human-voted front-end leaderboard, and #1 on the new Design Arena aesthetics eval. SWE-Bench Verified 76.4% (trails Fable 5’s coding line, beats Kimi K2.6’s 80.2% on agentic-harness runs in some independent tests, sits roughly level with Opus 4.8). Terminal-Bench 2.0 at 48.1%. On the gaps: HLE without tools 41.2%, more than 20 points behind Fable 5’s 64.5%, and AIME-class math still trails the closed frontier by mid-single digits.
Other info: Mixture-of-experts, ~360B total parameters with ~35B active per token, the same architecture family as GLM-4.6 so existing deployments swap weights without reconfiguring the inference stack. 256K-token context window (up from GLM-4.6’s 200K), 128K max output. Knowledge cutoff March 2026. License: MIT (genuinely permissive, no MAU or revenue carve-out, unlike the Modified MIT licenses Moonshot and DeepSeek ship). Technical report published; a full safety / system card is not part of the launch package.
More details: GLM 5.2 technical report and announcement
What shipped
Z.ai released GLM 5.2 on Friday as an open-weight successor to GLM-4.6, and the positioning is narrower and sharper than the usual “frontier-class, cheaper” pitch the open labs have been running all spring. GLM 5.2 is a ~360B / 35B-active mixture-of-experts model with a 256K context window, switchable hybrid reasoning, image input, and a genuinely permissive MIT license with no scale carve-out. It plugs into the agent harnesses people already use (Claude Code, Cline, Roo, OpenCode) through an Anthropic-compatible endpoint, which means the switching cost from a Claude-backed workflow is a base URL and an API key. The whole thing rides on a $3-a-month subscription that undercuts every frontier lab by more than an order of magnitude.
The evidence Z.ai put forward leans hard into one claim: this model designs better than the models that cost fifteen times more. GLM 5.2 took the top spot on WebDev Arena, the human-voted front-end leaderboard, ahead of both Claude Fable 5 and Claude Opus 4.8, and topped the new Design Arena aesthetics eval. On the harder, less flattering numbers Z.ai was more measured: SWE-Bench Verified at 76.4% lands it roughly level with Opus 4.8 and clearly behind Fable 5’s coding line, and the gaps on hard reasoning are real, with HLE-without-tools at 41.2% sitting more than 20 points under Fable 5. The catches are the ones that follow every Beijing-lab release: no published safety or system card at launch, a knowledge cutoff that predates most of 2026, and a benchmark table that wins decisively on taste and front-end output while conceding the deep-reasoning and tool-scheduling crowns to the closed frontier.
What’s new
GLM 5.2 isn’t a new base model, it’s a strong iteration on the GLM-4.x MoE family with a few things that genuinely separate it from both its predecessor and the closed frontier.
Design as the headline capability. Most labs treat front-end output as a side effect of coding ability. Z.ai trained for it directly and led the launch with it. Topping WebDev Arena and Design Arena over Fable 5 and Opus 4.8 is the sharpest differentiator in the release, and it’s the rare benchmark win that maps cleanly to something a user sees on the first prompt instead of a leaderboard they have to trust.
A genuinely permissive license. GLM 5.2 ships under plain MIT, no monthly-active-user ceiling, no revenue threshold, no mandatory credit line. Moonshot’s “Modified MIT” and DeepSeek’s license both carve out the largest deployers. Z.ai’s doesn’t, which makes 5.2 the most freely usable frontier-class open weight on the board.
Switchable hybrid reasoning in one model. Rather than ship a separate “thinking” SKU, GLM 5.2 exposes a per-request flag that turns extended reasoning on or off. You pay for deliberation only when the task earns it, which matters at this price point because the per-token cost is already near the floor and reasoning tokens are where the bill actually grows.
Harness-native from day one. The launch wires GLM 5.2 into Claude Code, Cline, Roo Code, OpenCode, and Kilo Code through an Anthropic-compatible endpoint, with the GLM Coding Plan subscription as the on-ramp. The model meets developers inside the tools they already run instead of asking them to adopt a new app.
256K context at the bottom of the price chart. The window grew from GLM-4.6’s 200K to 256K while the price stayed near the floor. It’s not Fable 5’s million-token ceiling, but it covers most real repos, and it does it for a twelfth of Fable 5’s input rate.
How and where to use it
Where it runs, what it’s actually good at, and where you’ll regret reaching for it.
Where it’s available
Z.ai chat app for direct use
Z.ai API via OpenAI- and Anthropic-compatible endpoints (swap the base URL, set
modeltoglm-5.2)GLM Coding Plan from $3 / month, wired into Claude Code, Cline, Roo Code, OpenCode, and Kilo Code
Hugging Face and ModelScope for open weights under MIT, served via vLLM or SGLang
OpenRouter for multi-provider routing.
What it’s good at
Front-end and UI generation, full-stack scaffolding, and design-forward work where taste is the deliverable (the WebDev Arena and Design Arena wins are the proof point, and they’re the rare benchmarks that show up in the output you see first)
Landing pages, dashboards, and interactive components from a single prompt
Agentic coding inside Claude Code or Cline where the Anthropic-compatible endpoint makes it a drop-in
High-volume, cost-sensitive work where the $3 subscription and $0.15 cache-hit price turn the per-call cost into a rounding error
Anywhere a permissive MIT license and self-hosting beat “hosted by the best lab”
What it’s bad at / shouldn’t be used for
Hard reasoning and knowledge-dense work
Complex multi-tool scheduling and the longest-horizon agent loops
Math-critical workloads where correctness is load-bearing
Regulated or data-sovereignty-sensitive work where sending prompts to a Beijing-hosted API is a non-starter (in which case self-host the MIT weights)
Any deployment that requires a published system card before sign-off
First impressions
The early read is launch-day hands-on from the open-community testers who jumped on the weights and the GLM Coding Plan within hours. The design story dominated the timeline.
The positives
Matt Velloso, a former VP at both Meta and Google DeepMind, is using GML 5.2 as a daily driver, claiming it finally meets his bar to do such.
Coming from an AI booster would be one thing, but coming from someone like Velloso is a prominent endorsement. His hardware callout also points towards an interesting trend of folks (and business!) starting to host models locally.
Someone ran a quick A/B test on a landing page and the results are worth checking out.
The price here is the real stinger, with a (debately better) landing page coming out of GLM 5.2 for $0.06 to Opus 4.8’s $0.49.
Riley Brown, an AI nut and self-described open model skeptic, compares the launch to DeepSeek R1 in how he expects frontier labs to react.
It’s not a bad bet, especially if the notorious “model feel” of GLM 5.2 makes it a breeze to use for design and coding. As open-weight models get closer to frontier performance for a fraction of the cost (whether this is due to Chinese energy efficiency or just massive subsidization is anyone’s guess), said frontier labs will start feeling the pressure to create powerful models for cheaper.
The negatives
Max Weinbach weighed Anthropic’s subsidies versus Z.ai’s and found that Claude came out on top.
This is a story that Anthropic (and OpenAI) need to be publicizing more if they want to keep up with the market. More usage is more usage, regardless of a model’s literal underlying token cost.
Raycast’s Thomas Paul Mann points out the obvious.
Given the model’s focus, it makes sense, but you’ve gotta wonder if these types of models are going to try gunning for the whole pie at some point.
Podcaster Ben Davis rejects the Anthropic/OpenAI equality statements.
Sometimes models perform well on benchmarks but don’t feel quite great. I wouldn’t be surprised if GLM 5.2 falls into this category once the hype wears out.
Jake’s take
I’ve been waiting for an open-weights model to take front-end seriously instead of treating it as a coding by-product. GLM 5.2 is the one. The landing pages, dashboards, and full-stack scaffolds I build in Cursor and Claude Code are exactly the workload WebDev Arena is measures, and a model that (supposedly) beats Fable 5 and Opus 4.8 on taste while costing a twelfth of Fable’s rate is dramatic. The Anthropic-compatible endpoint makes testing and implementation easy. If the output holds up past the first screen the way social media is saying it does says it does, I can comfortably shift a chunk of my UI generation off Claude and onto a three-dollar subscription (and feel a little guilty about how good a deal it is).
Worth noting: the same benchmark table that wins design loses hard reasoning by more than twenty points on HLE, so the second you hand GLM 5.2 a problem that’s coding wrapped around real domain expertise, you shouldn’t expect much. WebDev Arena rewards the best first screen, not the best app, so I’m holding the “best front-end model” headline at arm’s length until I’ve watched it survive five rounds of edits.
GLM 5.2 is the best-looking model open-weights model to date and the cheapest way to get a beautiful first draft. It’s also further evidence that the open-weights models, specifically the ones out of China, are rapidly approaching the frontier.
The downsides for recent open-weight models often remind me of a pro-scout's answer when asked what advice he would give Lebron James at the end of his first year in the NBA and his answer was "I don't know, he could shoot with his left more?"