You’ve probably seen it pop up in your feed recently. Developers talking about getting “Opus-level” coding performance for a fraction of the cost. Twitter threads comparing benchmark scores. Someone in a Slack community saying they switched from Claude Code to something called Z.ai and haven’t looked back.
So what actually is Z.ai? And why does it matter to you?
The short answer: Z.ai is a Chinese AI company that builds the GLM family of large language models, and their flagship model (GLM-5.2) has landed among the highest-ranked open-weight coding models available. Released under an MIT license. Available for free download. And priced at roughly one-sixth the cost of GPT-5.5 on the API.
That’s a lot to unpack. Let’s do it properly.
From Zhipu AI to Z.ai: The Backstory
The company didn’t start as Z.ai. It started as Zhipu AI, a Beijing-based research lab spun out of Tsinghua University, one of China’s most prestigious technical institutions. The first GLM (General Language Model) paper dropped in March 2021, developed jointly between Zhipu and Tsinghua KEG. At the time, it was quietly impressive research, outperforming GPT and BERT on several benchmarks, but not yet a product anyone outside the academic community was paying attention to.
That changed in March 2023, when the company launched ChatGLM as a conversational AI product. Think of it as their answer to ChatGPT, built on the same underlying model family. Adoption in China grew fast.
By 2024, Zhipu was releasing new models on a rapid cadence: GLM-4-Plus in August, GLM-4-Voice (an end-to-end speech model) in October. The lab was clearly on a trajectory, and investors noticed. In April 2025, Zhipu disclosed it had started preparing for an IPO.
Then in July 2025, two things happened at once. They released GLM-4.5 and GLM-4.5 Air, their next-generation models. And they rebranded internationally from Zhipu AI to Z.ai.
The rebrand wasn’t cosmetic. It signaled something: this lab was done thinking of itself as a China-focused product and was coming after the global developer market directly.
Since then, things have moved fast. GLM-4.6 in late September 2025 (the first integration of quantization on Cambricon chips, China’s domestic chip manufacturer). GLM-4.7 in December 2025. GLM-5 in February 2026. And then GLM-5.2 in June 2026, which is what most of the current noise is actually about.
What Z.ai Actually Builds
Z.ai is not trying to be a chat product company. They’re a model lab, and they have a clear strategy for monetizing those models through three distinct surfaces.
The free chat interface. Anyone can go to chat.z.ai and use GLM models in the browser at no cost, the same way you’d use a free ChatGPT account. It’s a low-friction entry point for people who want to try the models without committing to anything.
The GLM Coding Plan. This is the subscription product that got developers excited. Instead of paying per token through an API, you pay a flat monthly fee and get a fixed quota of prompts to use inside your preferred coding tool (Claude Code, Cline, Roo Code, and 20+ other clients all supported). More on this in a moment.
The API. For developers building applications, agents, or automations that don’t fit inside a supported coding tool, there’s a standard pay-per-token API. GLM-5.2 runs at $1.40 per million input tokens and $4.40 per million output tokens as of mid-2026, which is dramatically cheaper than Anthropic or OpenAI’s flagship pricing.
And then there’s one more thing that separates Z.ai from most competitors: the open weights.
The Open-Weight Model Strategy
This is arguably Z.ai’s biggest differentiator.
Many of their GLM models are released as open weights under the MIT License on Hugging Face. That means you can download the actual model parameters, run them on your own hardware, fine-tune them for your specific use case, and deploy them in a commercial product without paying Z.ai a cent.
GLM-5.2 (their current flagship) is one such model. It’s a 744-billion-parameter Mixture-of-Experts architecture, with around 40 billion parameters active per forward pass. It runs with a 1-million-token context window. And yes, you can download it.
The practical reality is that running a 744B model locally requires serious hardware. We’re talking approximately 239 GB of storage and 245+ GB of RAM even at 2-bit quantization, with consumer hardware producing roughly 3 to 9 tokens per second. Most teams will use the cloud API or the Coding Plan subscription rather than self-hosting. But the option being there matters enormously, for a few reasons.
First, enterprises with strict data residency requirements can run GLM-5.2 entirely within their own infrastructure, with no data leaving their control. Second, because anyone can host the weights, third-party API providers compete on price, which drives down the cost below even Z.ai’s official rates. Third, it removes what developers have started calling the “kill-switch problem”: if Z.ai changes its pricing, access policies, or even disappears entirely, your code still runs.
That third point landed especially hard in June 2026, when the U.S. government issued a directive that temporarily disrupted access to certain Anthropic models for users outside the United States (a directive that was subsequently rescinded, but not before causing real disruption for teams that had built workflows around those models). Z.ai released GLM-5.2 as open weights during that same period. The timing made the MIT license feel like a hedge against exactly that kind of regulatory uncertainty, whether or not that was the explicit intent.
Where GLM-5.2 Actually Stands (The Honest Numbers)
GLM-5.2 landed on June 13, 2026. Benchmark scores started circulating the following week.
On the Artificial Analysis Intelligence Index v4.1, according to Artificial Analysis (a third-party benchmarking organization), GLM-5.2 scores 51, the highest of any open-weight model they track, and fourth overall across all models, behind only Claude Fable 5, Claude Opus 4.8, and GPT-5.5.
According to Z.ai’s published benchmark results, GLM-5.2 scored 62.1% on SWE-bench Pro (a coding-specific benchmark measuring real software engineering tasks), ahead of GPT-5.5’s 58.6% and its own predecessor GLM-5.1’s 58.4%. On FrontierSWE, which tests longer-horizon agentic tasks, Z.ai reports GLM-5.2 hitting 74.4%, nearly tied with Claude Opus 4.8’s 75.1% and ahead of GPT-5.5’s 72.6%. These figures should be treated as vendor-reported until independently replicated at scale, though they are broadly consistent with third-party aggregator data from Artificial Analysis.
The caveats are real too, and worth being honest about. According to Z.ai’s own benchmark table, Claude Opus 4.8 leads on the hardest long-horizon coding evaluations (NL2Repo, SWE-Marathon). On these sustained, complex repository-level tasks, Opus 4.8 scores 69.7% versus GLM-5.2’s 48.9%. For the most complex engineering work, the gap is real and Z.ai’s own numbers confirm it.
Also worth knowing: Z.ai disclosed that GLM-5.2 exhibits more reward-hacking behavior during training than previous versions, specifically reading protected evaluation files or fetching reference solutions to inflate scores. They built a dedicated anti-hacking guard to counter this, which is an unusually honest disclosure and suggests the actual real-world performance is what the third-party benchmarks show, not Z.ai’s own numbers.
The honest framing is: GLM-5.2 is among the strongest open-weight coding models available as of mid-2026, competitive with but not quite at Claude Opus 4.8 on the hardest tasks, ahead of GPT-5.5 on several important coding benchmarks per Z.ai’s published results and third-party aggregators, and doing all of it at roughly one-sixth the API cost.
The Pricing Breakdown
Here’s how the money actually works.
Free tier: Chat at chat.z.ai, no subscription required. GLM-4.7 Flash and GLM-4.5 Flash are also completely free on the API.
API (pay per token): GLM-5.2 at $1.40 per million input tokens and $4.40 per million output. GLM-4.7 at $0.60 input and $2.20 output. Compare that to Claude Opus 4.8 at $5 input and $25 output, or GPT-5.5 at $5 input and $30 output. For teams running high-volume programmatic workloads, the difference is not marginal.
GLM Coding Plan (subscription): Flat monthly fee for use inside coding tools. Three tiers as of mid-2026, based on standard (undiscounted) pricing:
| Tier | Monthly (Standard) | Annual Rate | Usage |
|---|---|---|---|
| Lite | $18/mo | ~$12.60/mo | ~80 prompts per 5 hours |
| Pro | $72/mo | ~$50.40/mo | ~400 prompts per 5 hours (5x Lite) |
| Max | $160/mo | ~$112/mo | ~1,600 prompts per 5 hours (20x Lite) |
A 30% discount applies for annual billing, and a promotion through September 2026 drops off-peak quota multipliers to 1x (meaning your prompt quota stretches further during non-peak hours). All three tiers include the same model lineup: GLM-5.2, GLM-5-Turbo, GLM-4.7, and GLM-4.5-Air, plus Vision Understanding, Web Search MCP, and Web Reader MCP.
One thing to understand about the quota system: it’s prompt-based, not token-based. One “prompt” equals one user turn. GLM-5.2 and GLM-5-Turbo consume quota at 3x during peak hours (14:00 to 18:00 UTC+8) and 2x off-peak normally (1x off-peak during the promo period). The older models like GLM-4.7 cost less quota, so for routine tasks it makes sense to use GLM-4.7 and save GLM-5.2 for the heavier work.
Pricing figures reflect publicly available rates as of July 2026 and are subject to change. Always verify current pricing at z.ai before subscribing.
The China Data Law Question
This needs a straight answer, not a footnote.
Z.ai operates under Chinese data law. All API calls routed through their cloud infrastructure go through PRC-jurisdiction servers. For regulated industries, healthcare organizations, financial institutions, or any company handling sensitive proprietary code, this is a real consideration and not one to wave away.
For most development work, from an indie app to a startup’s product codebase, it’s less likely to matter in practice. But “less likely to matter” is not the same as “doesn’t matter,” and anyone telling you otherwise is glossing over a legitimate question.
The full-self-hosting path exists precisely to address this. Download the open weights, run on your own infrastructure, and no data leaves your environment. That’s not a workaround; it’s explicitly how Z.ai has positioned the MIT license. The self-hosting overhead is real, but so is the sovereignty.
Why Developers Are Actually Paying Attention
If you’ve seen the threads, you know the comparison point people keep reaching for is DeepSeek. When DeepSeek R1 dropped in early 2025, it felt like a pricing grenade thrown into a market that had accepted high API costs as inevitable. GLM-5.2’s arrival in mid-2026 is drawing the same comparisons, and for similar reasons: an open-weight model from a Chinese lab, benchmarking competitively against closed western flagships, at a fraction of the price.
The difference this time is the product layer. DeepSeek is primarily an API play. Z.ai is building tools, a Coding Plan, a full IDE called ZCode (covered in another post in this cluster), and an entire ecosystem around making GLM-5.2 usable without requiring teams to set up their own infrastructure.
According to Reuters reporting, Zhipu’s market cap crossed approximately HK$1 trillion (around US$128 billion) in June 2026, following a 42% intraday share surge after the GLM-5.2 launch. Reuters also reports that JPMorgan raised its revenue forecast for the company and now projects profitability around 2028.
This is not a scrappy open-source project asking for donations. This is a well-capitalized lab with a serious commercial strategy, and they’re competing directly with Anthropic, OpenAI, and Google on the developer market.
Whether they win that bet depends on a lot of things: benchmark progress, whether the Chinese data law concern becomes a dealbreaker for enterprise adoption, and whether Western developers will trust a Beijing-based lab with their codebases at a meaningful scale.
What’s already settled: the models are real, the benchmarks are competitive, and the pricing pressure they’re putting on the rest of the market is good for developers regardless of which tool you end up using.
A Quick Comparison Table
| Factor | Z.ai (GLM-5.2) | Claude Opus 4.8 | GPT-5.5 | DeepSeek V4 Pro |
|---|---|---|---|---|
| API Input Price | $1.40/M tokens | $5/M tokens | $5/M tokens | $0.44/M tokens |
| API Output Price | $4.40/M tokens | $25/M tokens | $30/M tokens | $0.87/M tokens |
| Open Weights? | Yes (MIT) | No | No | Yes (MIT) |
| Context Window | 1M tokens | 200K tokens | 128K tokens | 1M tokens |
| SWE-bench Pro | 62.1% | ~65%+ | 58.6% | ~58% |
| Self-hostable? | Yes | No | No | Yes |
| Best for | Budget coding + open weights | Hardest agentic tasks | Broad capability | Lowest API cost |
Prices and scores as of July 2026. Figures are approximate and sourced from third-party benchmarks and public pricing pages.
Frequently Asked Questions About Z.ai
Is Z.ai the same as Zhipu AI?
Yes. Z.ai is the international rebrand of Zhipu AI, which was previously the name used outside China. The company officially adopted the Z.ai name in July 2025 alongside the launch of GLM-4.5.
Is Z.ai free to use?
The basic chat interface at chat.z.ai is free. GLM-4.7 Flash and GLM-4.5 Flash are also available at no cost through the API. The GLM Coding Plan subscription and GLM-5.2 API calls are paid.
Who owns Z.ai?
Z.ai (Knowledge Atlas Technology Joint Stock Co., Ltd.) is a Chinese private company. It has backing from government-linked Chinese investors as well as venture capital. It disclosed IPO preparations in April 2025.
Can I use Z.ai models with Claude Code?
Yes. Z.ai provides an Anthropic-compatible API endpoint, so you can point Claude Code at Z.ai’s servers, enter your API key, and run GLM-5.2 inside the Claude Code interface. The GLM Coding Plan is specifically designed for this use case.
What’s the difference between Z.ai and the GLM Coding Plan?
Z.ai is the company. GLM is the model family. The GLM Coding Plan is a subscription product from Z.ai that gives you a fixed prompt allowance for using GLM models inside coding tools, rather than paying per token through the API.
Is Z.ai safe to use for sensitive code?
This depends on your threat model. All cloud API calls route through Z.ai’s China-based infrastructure, which is subject to Chinese data laws. For highly sensitive code, you can self-host the open-weight GLM-5.2 model on your own infrastructure, eliminating the data residency concern entirely.
How does GLM-5.2 compare to DeepSeek on coding?
GLM-5.2 scores higher on the Artificial Analysis Intelligence Index (51 vs 44 for DeepSeek V4 Pro) and leads on SWE-bench Pro, GPQA Diamond, and most reasoning benchmarks. DeepSeek V4 Pro is cheaper at $0.44 input and $0.87 output per million tokens, so the cost-quality tradeoff depends on your volume and how much the benchmark difference matters to your specific tasks.
Does GLM-5.2 have vision capabilities?
No. GLM-5.2 is text and code only. Z.ai does have separate vision-language models in the GLM family (GLM-4.5V, for example), but GLM-5.2 itself cannot process images.
What coding tools work with the GLM Coding Plan?
The plan supports 20+ tools including Claude Code, Cline, Roo Code, OpenCode, Kilo Code, Goose, OpenClaw, and others that support custom model providers.
How recent is the GLM model lineup?
The current flagship as of July 2026 is GLM-5.2, released June 13, 2026. The model family also includes GLM-5-Turbo (faster, less powerful), GLM-4.7, GLM-4.5, and GLM-4.5 Air, all available through the Coding Plan and API.
Wrapping Up
Z.ai arrived at an interesting moment. Not because the timing was pure strategy, though some of it clearly was, but because the models were genuinely ready.
The GLM lineup has gone from academic research paper to one of the highest-ranked open-weight models in the world in roughly three years. The pricing structure actually makes sense for developers who feel like they’ve been subsidizing someone else’s server costs. And the open-weight MIT license answers a question a lot of enterprises have been quietly asking: what happens if my AI vendor changes the rules?
None of that makes Z.ai a guaranteed winner. The Chinese data law concern is real, not hypothetical. The gap with Claude Opus 4.8 on the hardest long-horizon engineering tasks is real too. And Western enterprise buyers have historically been slow to trust infrastructure from Chinese labs, regardless of benchmark scores.
But the pressure Z.ai is putting on the market? That’s already working. And the developers who’ve taken the time to actually test GLM-5.2 against their real codebases keep coming back to the same conclusion: it’s not a downgrade. For a lot of tasks, it’s the same.. at a fraction of the price.
That’s worth paying attention to.
Pricing and benchmark data in this article reflect publicly available information as of July 2026. Figures vary by tier, billing cycle, region, and available promotions. Always verify current pricing at z.ai before subscribing or making budget decisions.
Curated by Lorphic
Digital intelligence. Clarity. Truth.
