GLM 5.2 — the 1M-context flagship. The same day Claude got banned, China gave one away.

Here's the announcement.

Zhipu announced it on X themselves. Read it — then I'll break down what each part actually means.

Nearly a million views in a day. Three claims stand out — a 1M context, it's coding-first, and the weights go fully open under MIT next week. Let's take each one.

The specs, in plain English.

The headline is that context window. A million tokens means you can drop an entire codebase into one session and it holds the thread. That's the lane where Claude has always had the edge — large-codebase work and long, multi-hour agent runs. GLM 5.2 is aiming straight at it.

One honest flag: Z.ai calls the 1M context "usable" — a careful word. Retrieval quality across the whole window hasn't been independently tested yet. Treat it as promising, not proven.

A frontier-class coding model with 1M context on a flat plan makes per-token API spend look steep. That's the whole point.

The company behind it is blacklisted in America.irony

Here's the part that makes today surreal.

Zhipu AI — the lab behind GLM — was added to the US Entity List back in January 2025.

It was the first Chinese large-model company ever blacklisted by the US government.

And it didn't get the soft version. It got a "Footnote 4" designation — the harshest tier, which blocks not just US tech but any foreign product built with US parts.

So sit with the timing.

On the same day the US government switched off Claude's best models for national security, a Chinese company that America has officially blacklisted handed those same Americans a free, frontier-class coding model.

The thing you were told was too dangerous to export — and the thing built by a company you're not allowed to trade with — landed on the same Tuesday.

That's the real story under the model. The US tried to wall this company off. The company answered by giving its best work away for free, under the most permissive licence there is. Whatever you think of that — it's a strategy, and it's working.

Why it can offer 1M context this cheap.

You don't need to be an engineer for this — but two pieces of tech explain why it runs so efficiently.

1. It only pays attention to what matters

Normal models make every word look at every other word. That's what makes long context expensive — costs explode as the text grows.

GLM uses something called DeepSeek Sparse Attention. Instead of everything looking at everything, the model dynamically focuses on the tokens that actually matter and skips the rest.

That's the trick that makes a one-million-token window affordable instead of bankrupting. Same idea as skim-reading a book for the important bits instead of re-reading every page.

2. It trains itself faster than anyone

Zhipu built a training system they call Slime. It splits the "generate practice problems" step from the "learn from them" step so both run at full speed without waiting on each other.

That's the boring-but-huge reason GLM went from 5 to 5.1 to 5.2 in a matter of months. They can retrain and ship faster than almost anyone.

Translation: this line was built specifically for long, autonomous coding work — not chat. That's the lane it's gunning for.

They shipped it with zero benchmarks.

Here's the part to be honest about. At launch, Zhipu published no benchmark numbers. No SWE-bench, no LiveCodeBench, no HumanEval. Nothing.

The X replies noticed instantly. Some read the silence as a red flag — "why hide the numbers if they're good?" Others who actually tried it came away impressed.

For the lineage, though, we have real, independently-checked numbers — and they're strong.

So the honest read: GLM is genuinely top-tier on coding and agent work, a touch behind on the very hardest pure reasoning, and unbeatable on price. GLM 5.2 is that same family with five times the context and a deeper thinking mode. Wait a week for independent 5.2 numbers before you trust the hype — but the bloodline is already proven.

The efficiency gap is not even close.

Strip away the hype and this is the real reason people are switching.

So you're getting ~95% of the best model on the planet, for a fraction of the effort.

And here's the kicker — switching is basically free. You don't change tools. You change one endpoint and keep working in Claude Code exactly like before.

For a long time, Claude's one untouchable edge was native 1M context. That was the reason to pay up. GLM 5.2 just closed that gap too.

That doesn't make Claude bad. Claude still wins the hardest reasoning, and it has multimodal that GLM doesn't. But it does mean the "just pay for the best, it's worth it" argument got a lot weaker this week.

GLM 5.2 landed in a crowded field.

This is the bit most coverage misses. GLM 5.2 isn't alone. There's a whole wave of Chinese open models trading blows at the top of the coding charts right now.

Notice what every single comparison site concludes.

Read that twice. The experts aren't telling you to pick one model. They're telling you to run several and send each job to whichever one is best for it.

DeepSeek for hard algorithms. Kimi for parallel agents. GLM for long agentic builds. Claude for the trickiest reasoning. Whatever's cheapest for the simple stuff.

That's not a model decision. That's a system decision. Which is exactly where this is going.

How to switch your agent to GLM 5.2 today.

This is the bit that matters. You don't need a new tool. GLM 5.2 plugs into the agents you already run — Claude Code, OpenClaw, Cline, OpenCode, Roo Code, Goose, Crush, Kilo Code. Here's the exact setup.

Step 1 — get on the Coding Plan

Grab a GLM Coding Plan key from z.ai (any tier works). The model runs on the coding endpoint — not the pay-per-token one — so the subscription covers it:

https://api.z.ai/api/coding/paas/v4

Step 2 — point Claude Code at it

Open ~/.claude/settings.json and set GLM 5.2 as your Opus and Sonnet models. The [1m] suffix turns on the full million-token context:

{
  "env": {
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL":  "glm-4.5-air",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.2[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL":   "glm-5.2[1m]"
  }
}

Step 3 — turn the thinking up

Inside a session, type /effort and pick max. For coding, max effort is where GLM 5.2 does its best, most stable work:

# low / medium / high  → maps to GLM "high"
# xhigh / max          → maps to GLM "max"  (use this for code)
/effort max
/status   # confirm it now shows glm-5.2[1m]

On OpenClaw it's the same idea — add the glm-5.2 model to your zai provider and set it as primary with a fallback. That's it. New brain, same hands.

And if Claude Code says the [1m] model "doesn't exist" — just update Claude Code to the latest version and try again. That's the one gotcha.

One model gets banned. Another gets given away. Same week.

That's the new normal. The model layer changes weekly now. The people who win aren't chasing each drop — they're running a system that takes the best of all of them.

Every source, first-hand.

The whole thing — at a glance.