The Goldie Open Genius™ + GLM-5.2.
A frontier coder that's open, free to own — and today it out-scored the giants on the jobs that actually matter.
z.ai shipped GLM-5.2 on 13 June with a 1-million-token brain and almost nothing about how good it was.
No SWE-bench. No LiveCodeBench. Just "here's the model, go try it."
Today the numbers landed. A tiny, open, downloadable model just matched the big paid ones — for a sixth of the cost.
This is the guide to what dropped, what it means for you, and how I run it for pennies inside my Agent OS.
Read it — and run it — yourself.
Every number in this guide is sourced. Here's where today's drop was reported, plus z.ai's own pages so you can check it first-hand:
"Z.ai's open-weights GLM-5.2 beats GPT-5.5 on multiple long-horizon coding benchmarks — for one-sixth the cost."
— VentureBeat headline, 17 June 2026
here's z.ai announcing it ↓
The benchmarks finally landed. And they're loud.
Here's the headline.
On the long, grinding jobs — the multi-hour, multi-step work agents actually do — GLM-5.2 beats GPT-5.5.
And it lands within a single point of Claude Opus 4.8, the most expensive coder on the board.
Look at the long-horizon score first. This is the one that matters for real agent work:
FrontierSWE — long-horizon task completion
Now look at the really long jobs — the multi-hour builds.
On PostTrainBench the gap to GPT-5.5 isn't close at all:
PostTrainBench — multi-hour engineering
Same story on broad reasoning with tools turned on.
GLM-5.2 slots in ahead of GPT-5.5 again, just behind Opus:
Humanity's Last Exam — with tools
One more, gen over gen.
On SWE-bench Pro it steps up to 62.1 from GLM-5.1's 58.4. A clean jump, not a rounding error.
SWE-bench Pro — generation over generation
I gave it one sentence each. It built these.
Benchmarks are one kind of proof. Here's the other.
Every panel below is a single self-contained file GLM-5.2 wrote — a playable game or a live toy, one shot, no asset packs, no second prompt.
The visual toys run live as you scroll. The 3D games stay paused until you hit ▶ Play — so the page stays smooth — or open any of them fullscreen.
toys run live · games are click-to-play so the page stays smooth · "play fullscreen" gives full mouse-look.
all single HTML files, one shot each, written by GLM-5.2.
I was you. Then I found the open path.
Before
I was locked into the most expensive models on earth.
Every long agent job ran up a tab I could watch ticking.
I'd kick off a big build and pray it didn't loop and burn through credits.
The grinding multi-hour stuff — the work I most wanted agents to do — was the work that cost the most to run.
And I owned none of it. Pull the plug on the bill and the whole thing went dark.
Then the open models caught up — and GLM-5.2 was the one that changed it.
After
Now the long, boring builds run on a frontier model for a sixth of the price.
It holds my whole project in its 1-million-token head, so it stops forgetting halfway through.
It lives in my Agent OS next to Claude and Kimi — I point the cheap grinder at the long jobs and save the pricey one for the hardest 5%.
And the weights are open. I can download the whole brain and run it myself. Nobody can switch it off.
You can have this too. Same model. Same path. It's free to own.
Real people. Real wins. Inside the Boardroom right now.
Here's what's already happening for the members running this stack — agency owners, ecom founders, course creators, solo operators. Different businesses. Same result.
What's already happening for members building with cheap, open agents
Commit to switching one job over today. Not tomorrow.
You've seen the proof. Real people. Real results.
The next few minutes show exactly what dropped and how I run it for pennies.
So here's the deal.
Promise yourself one thing right now. You'll finish this guide and move one task — just one — onto a cheaper, open model before you sleep tonight. Because the moment you make that switch, the cost of running AI all day stops being the thing that holds you back.
The people sitting still are watching their spend climb. The people switching today are the ones who'll look back in six months and say "that was the moment it got cheap."
Be one of those people.
Commit to the switch. Commit to taking action today. This changes what AI costs you forever.
The Goldie Open Genius™.
Five things make GLM-5.2 the model I reach for first. Put together, that's the Open Genius — a frontier brain that's open, cheap, and yours.
The Open Door
The weights are MIT. You download the whole brain, run it yourself, and own it forever. No lock-in, no off-switch someone else holds.
The Giant's Score
It matches the big paid models on the jobs that count — beats GPT-5.5 on long-horizon coding, ties Opus inside a point. You stop paying a premium for the same result.
The Penny Price
A sixth of GPT-5.5. A fifth of Opus on output. The long grinding jobs stop costing a fortune, so you can actually run them.
The Long Memory
One million tokens of context. It holds your whole codebase or project at once, so it stops forgetting what it was doing halfway through.
The Night Shift
Cheap plus tireless means you point it at the multi-hour builds and let it grind while you sleep. You wake up to finished work.
The "coding" model is really a long-job engine. It runs agents that do research, write content, sort leads, and handle ops.
Members run agencies, ecom, coaching and content on this exact stack. The engine room doesn't care what you sell.
— real member, non-technical, running agents anywayThe same job. A sixth of the price.
Here's the shift in one picture — running a long agent build the old way versus the Open Genius way.
- Locked to one pricey frontier model
- Watch the meter tick on every long run
- Avoid the big grinding jobs to save spend
- Re-feed context every session — it forgets
- You rent the brain — cancel and it's gone
- One vendor, one price, take it or leave it
- Frontier-level results at a sixth of the cost
- Run the long jobs freely — the meter barely moves
- Point the cheap grinder at the multi-hour builds
- 1M context holds the whole project in its head
- Open weights — download it, own it, run it yourself
- Route by job: cheap for most, pricey for the hard 5%
Not here. Look back at the bars — it beats GPT-5.5 on the long jobs and ties Opus inside a point.
You're paying less for the same result, not paying less for a weaker one.
— real member, cut spend without losing qualitythe hands-on takes started landing fast ↓
Tool use and the terminal: basically a tie.
It's not just the long jobs. On tool orchestration it's a point off Opus.
MCP-Atlas — tool-use orchestration
Terminal work is the one spot it trails the top two — but it still clears Gemini 3.1 Pro with room to spare.
Terminal-Bench 2.1 — terminal-heavy workflows
people lined it up against everything ↓
Where it still loses — read this.
I'm not going to pretend GLM-5.2 is the new king. It isn't.
On the deepest, repo-scale software jobs, Claude Opus 4.8 still pulls clear. These gaps are real and big:
The gaps Opus still owns
So the honest read: GLM-5.2 has closed the gap to a single point on a lot of agent work — but on the very hardest jobs, Opus still earns its price.
That's exactly why you run both, and route by job.
Keep Claude. This isn't a swap — it's a router.
Run GLM-5.2 for the long, cheap, grinding 95%. Save Claude for the hardest 5% where it still wins. Your spend drops and your quality holds.
— real member, runs more than one model nowNow look at the price tag.
Benchmarks are half the story. Cost is the other half.
"Within a point of Opus" reads very differently once you see what each one charges to run:
Output price per 1M tokens — the real cost driver for long agent runs
That's the whole pitch in one chart.
Frontier results on most agent work, for roughly a sixth of GPT-5.5 and a fifth of Opus on output.
For an agent that burns millions of tokens a day, that's not a discount. It's a different way to run a business.
and because the weights are open, you can run it yourself ↓
The catch is small. The API is cheap, not always free — but the weights are MIT, so the brain itself is yours to download and run.
You go from renting intelligence by the token to owning it outright.
— real member, stopped paying for what they could run cheapThree ways to plug it in. Pick the job.
The games up top are just the warm-up.
The real point is this — you can build or automate almost anything with GLM-5.2.
We wired it into the Agent OS three ways. Same cheap, open brain. Three doors. You pick the one that fits what you're doing.
Type it. Watch it build.
The GLM 5.2 panel in your dashboard. You type what you want in plain English, it streams the code, and the finished build lands in your workspace — one click to preview or play.
Best for fast, one-shot things. Every game on this page was made this way.
→ it writes the file → you click Preview → you're playing it.
Hand it a job. Walk away.
GLM-5.2 as a Hermes agent. Give it a task and it works on its own — plans the steps, runs them in the background, and pings you when it's done.
Best for automation. The long, multi-step jobs you don't want to babysit.
→ it runs in the background → you come back to 10 drafts.
Run the whole brain yourself.
The weights are open, so you can run GLM-5.2 on your own machine through Ollama, then drive it from Hermes or Claude Code.
Best for private work and zero per-token cost. Your data and the model never leave your computer.
→ build + automate offline, nothing sent to the cloud.
That's the whole idea behind the Open Genius.
One cheap, open, frontier brain — and three ways to point it at real work.
Want GLM, Claude and Kimi in one dashboard?
If you want to actually use what you just saw — the cheap long-horizon grinder, routed against Claude for the hard jobs — that's the Agent Operating System inside the AI Profit Boardroom.
It's a full operating system I built that connects Claude, Kimi and GLM-5.2 into one dashboard.
Your agents share one memory. They know your goals. They know your business. So when you point the cheap model at a long build, it already has your full context — and you keep the expensive one for the 5% that needs it.
The Agent Operating System
You get the full zip, every prompt, the memory setup, and coaching calls where I walk you through the whole thing.
- One dashboard running Claude, Kimi and GLM-5.2 side by side
- A 30-day roadmap for wiring cheap open models into real work
- Four coaching calls a week with people running agents in production
- Daily tutorials as each model and update ships
- A prompt library + a member map to find operators near you
- 3,600+ founders building this right now · someone online 24/7
Three beliefs in your way.
158 pages of members who already broke through these exact beliefs. Real businesses. Real wins. All documented.
Read the 158-page testimonials doc →My honest advice.
Don't rip out what works. Add, don't replace.
Move one long, expensive, grinding job over to GLM-5.2 this week and watch two things: the quality, and the cost.
Keep Claude on the hardest repo-scale work where it still wins.
The open weights are still rolling out under MIT, so if you want to self-host, check that before you build a product on it.
And remember every number here is z.ai's own, from today's drop — strong, but wait for the outside labs to re-run it before you treat it as gospel.
The people who figure out cheap, open agents now, while the tools move fast, are going to be way ahead when everything settles. Every workflow you build, every job you move over — it all compounds.
Six months from now, running a frontier model for pennies is just normal.
The window where most people are still overpaying — and you aren't — is open right now. It closes fast.
— real member, wishes they'd started soonerthe reaction kept rolling all day ↓
Live X embeds need a connection to load — offline they show as plain links. That's normal for X.
Direct: @Zai_org ·
@ai_for_success ·
@Designarena ·
@ollama ·
@atomic_chat_hq
What you walk away with.
Frontier-level coding for a sixth of GPT-5.5.
It beats GPT-5.5 on long jobs, ties Opus inside a point.
1M context holds the whole project at once.
Open MIT weights — download the brain and own it.
Cheap model for the 95%, Claude for the hard 5%.
Point it at the multi-hour builds and wake up to finished work.
The frontier didn't get smarter today. It got six times cheaper. That's the bigger deal.
Make it actually save you money every day.
If you want this to be more than a model you tried once, go grab the Agent Operating System inside the AI Profit Boardroom.
It turns Claude, Kimi and GLM-5.2 into one system with shared memory, shared context, and one dashboard you control.
Your agents understand your business. They remember everything. And every new model — like GLM-5.2 today — makes the whole system cheaper and stronger automatically.
The Agent Operating System
I built it in one session. You get the zip. Every prompt. The memory setup. Coaching calls where we set it up together, step by step.
- 3,600+ members · daily tutorials · a 30-day roadmap
- One dashboard for every model, routing cheap-vs-pricey for you
- A member map to find operators near you · someone online 24/7
- The full 158-page testimonials doc — read the wins before you join
Move one job over first. Decide second.
I'll see you in the next one.
Sources · today's drop: VentureBeat · llm-stats · morphllm · OpenRouter · launch context: Codersera, i-scoop. All figures vendor-reported; independent re-runs pending.