GLM-5.2 Benchmarks Beat GPT-5.5 at One-Sixth the Cost. Here's Why That Matters.

TL;DR

- GLM-5.2 scored 77.8 on SWE-bench Verified and 56.2 on Terminal Bench 2.0, surpassing every open-weight model to date and beating GPT-5.5 on multiple long-horizon coding benchmarks. - The model costs roughly $1.40 per million input tokens and $4.40 per million output tokens via API, compared to GPT-5.5's ~$7.50/M output on OpenRouter. - Its 1,000,000-token context window means you can feed an entire mid-size codebase into a single task without chunking. - MIT licensed and open-weight on Hugging Face, meaning you can self-host, fine-tune, or route through any API proxy without vendor lock-in. - Vercel CEO Guillermo Rauch posted "Genuinely impressed, almost shocked, at how good GLM-5.2 is at coding. This changes things" — 1.8 million views and climbing.

---

Z.ai dropped GLM-5.2 on June 16, 2026. Within 48 hours it was the top post on Hacker News, racking up 364 points and a thread that kept burning for days.

The pitch was simple: the best open-weight coding model ever released, MIT licensed. And cheap enough that solo operators and small agencies could finally stop negotiating with their closed-model providers.

The numbers backing that claim are real.

GLM-5.2 hit 77.8 on SWE-bench Verified and 56.2 on Terminal Bench 2.0. The two benchmarks that actually predict whether a model can survive a full coding session without falling apart. On the Artificial Analysis Intelligence Index v4.1, it scored 51, eleven points clear of its predecessor and higher than any open-weight model ever tested. Third-party reviews and developer hands-on reports have been consistent: this is the first open model that doesn't require you to hold your breath when it touches your codebase.

But benchmarks are theater.

What actually matters is whether it works in a harness.

The One Thing That Makes GLM-5.2 Different

Every few months a model posts impressive numbers and then disappears because it falls apart the moment a human actually uses it. GLM-5.2 is other in one specific way: the IndexShare optimization in its architecture reduces per-token compute by 2.9x at the full 1,000,000-token context length.

That's not a marketing claim. That architecture detail means you can actually use that million-token window in practice. Most models can technically support long contexts but become slow and expensive at the extremes. GLM-5.2 doesn't. Feed it a 200,000-line codebase in one shot and it processes it without the cost scaling that kills your per-task budget.

For small agencies running coding agents on real client work, that changes the workflow math.

You stop treating context as a scarce resource you must ration. You stop breaking projects into artificial chunks that force the agent to re-orient every time. You hand it a repo and let it work.

The 744-billion-parameter MoE architecture activates roughly 40 billion parameters per token. That puts it in the range where a well-configured 8xH100 setup can actually run it for self-hosting. Z.ai trained the whole thing on Huawei Ascend chips.

A deliberate choice that underscores how far China's hardware ecosystem has moved away from NVIDIA dependency.

The Cost Math That Should Make Anthropic Nervous

Here is what actually got people's attention on social media.

GLM-5.2 via API runs about $1.40 per million input tokens and $4.40 per million output tokens. GPT-5.5 on OpenRouter sits closer to $7.50 per million output tokens. The per-token gap looks modest until you factor in that output tokens. The generated code, the reasoning traces, the test outputs — are where your budget actually goes in a coding workflow.

Third-party analysis put the real-world cost differential at roughly one-sixth. That is not a rounding error.

That is the difference between a feature that pencils out and one that you quietly drop as the API bill got away from you.

The MIT license compounds the impact.

You are not renting access to someone else's infrastructure. You can download the weights from Hugging Face today, set up your own inference endpoint. And never worry about a provider changing their pricing, restricting your use case, or going down during a client delivery. For agencies with compliance requirements or data residency preferences, that option did not exist at this performance level six months ago.

Low built-in moderation is part of the package. If you have spent any time with heavily safety-filtered models, you know the moment where you are deep in a legitimate debugging session and the model refuses to engage with the actual problem since some edge case tripped a content filter. GLM-5.2 runs looser out of the box. Whether that is a feature or a concern depends on your use case. But for a coding agent doing real work in production, it is mostly a feature.

What You Should Actually Do With This

Stop treating this as a benchmark curiosity.

The real question is whether your current stack has a credible fallback if your primary coding model provider raises prices, changes terms, or gets throttled during a critical sprint.

If you run Claude Code for client work and you are on an annual contract, you probably should not rip that out this week. But if you are starting a new internal project, spin up GLM-5.2 via OpenRouter as a secondary provider and run it in parallel. Compare actual output quality on your specific codebase. The benchmark scores are predictive but not definitive — your workflow is the final judge.

The small-team angle is where this gets genuinely interesting.

Solo developers and two-person agencies do not have the use to negotiate API pricing or the infrastructure team to absorb an outage. An open-weight model that performs at this level gives you something you did not have before: bargaining power. You can credibly tell your primary provider that you have an alternative, and that alternative is not a downgrade.

If you have been watching the coding-agent space waiting for the moment when open models caught up with closed ones, this looks like that moment. Not a demo. Not a research paper. A model you can download, run, and actually ship work with today.

The Takeaway

GLM-5.2 is not a theoretical argument about open-model parity. It is a working tool that scored higher than GPT-5.5 on the benchmarks that matter, at roughly one-sixth the cost, under an MIT license, with a context window that makes long-horizon coding sessions practical for the first time at this price point.

The Vercel CEO was not being hyperbolic when he said this changes things.

He was reading the room. And the room was a lot of developers who had been waiting for exactly this.

Set up an OpenRouter account. Route one project through GLM-5.2 this week. See for yourself whether the hype matches your actual workflow. That is the only test that matters.