GLM-5.2, reviewed: an open-weight model with a real coding claim

Open-weight releases usually mean a trade: you get the weights and a low price, and you give up a few points of capability to whatever closed frontier model currently tops the charts. GLM-5.2 narrows that trade further than most. Zhipu AI, operating its hosted API as Z.AI, released it as a 753-billion-parameter Mixture-of-Experts model under the MIT license, with weights published on Hugging Face and a hosted API priced well below the frontier tier — while posting provider-reported benchmark numbers that land ahead of GPT-5.5 and Claude Opus 4.7 on more than one coding test.

The catch isn't the license or the price. It's the size. At 753B total parameters, this isn't a model you run on a gaming GPU, MIT license or not — plan on a serious multi-GPU server if you want to self-host it, and the hosted API is the realistic path for most teams.

A sourcing note before the numbers do any work: the pricing and context figures come from Z.AI's own pricing and model documentation. The parameter count and license come from the official Hugging Face model card under the zai-org organization. The benchmark figures below are Z.AI's own provider-reported scores from the same model card and release notes — not an independent benchr test, and not third-party verification. Treat them as a vendor claim with a named benchmark, which is a much stronger evidence tier than an unnamed marketing number, but still one company's own run.

What Z.AI actually claims, and against whom

Z.AI's release notes and Hugging Face card list a specific, named set of results: SWE-bench Pro at 62.1 (up from GLM-5.1's 58.4), Terminal-Bench 2.1 (Terminus-2 harness) at 81.0, GPQA-Diamond at 91.2, AIME 2026 at 99.2, and Humanity's Last Exam at 40.5 text-only or 54.7 with tool use. Z.AI frames the release as its "highest-ranked open-source model" across the suites it tested, and says it outperforms GPT-5.5 and Claude Opus 4.7 on multiple benchmarks — while landing about a percentage point behind Claude Opus 4.8 on its own FrontierSWE comparison.

The Terminal-Bench line is the honest one to sit with: 81.0 against Opus 4.8's 85.0 is a real gap, not a rounding error, even by Z.AI's own numbers. That's the shape of this release — a genuinely strong open-weight model that closes most, not all, of the distance to the priciest closed frontier option, on Z.AI's own scoreboard.

The price makes the benchmark gap easy to live with

Here's where GLM-5.2 actually wins the argument. At $1.40 input and $4.40 output per million tokens, it costs a fraction of GPT-5.5's $5/$30 or Claude Opus 4.8's $5/$25 — roughly a 20x-plus difference on output tokens against Opus 4.8. Cached input at $0.26 per million pushes repeated-context work even lower. For a team running heavy agent loops or long coding sessions, that per-token gap compounds fast, and it's large enough to absorb the Terminal-Bench deficit against Opus 4.8 many times over in pure budget terms.

It's not the cheapest open-weight option on the market — DeepSeek-V4-Pro runs $0.435/$0.87 and Alibaba's open-weight Qwen3.6-27B is free to self-host — but neither of those posts the same coding-benchmark numbers GLM-5.2 claims. The open-weight tier roundup covers where DeepSeek, Qwen, Llama, and Mistral each fit; GLM-5.2 slots in as the pricier-but-stronger coding option among that group, not the rock-bottom one.

MIT license, real weights, serious hardware

The MIT license on the Hugging Face card is about as permissive as open-weight licensing gets: commercial use, modification, and redistribution with almost no restriction. That matters if your legal team has ever flagged a more restrictive "open" license from another provider. But permissive licensing doesn't shrink the model. At 753B total parameters in a Mixture-of-Experts architecture, this is not a laptop or single-GPU model — multi-GPU hardware in the 8xH100 range is the kind of setup this size class typically needs, and Z.AI's own docs don't spell out a lighter self-host path.

For nearly everyone, that makes the hosted API at api.z.ai the practical way to use GLM-5.2, license notwithstanding. The MIT license is real and valuable if you eventually want to self-host or fine-tune, but budget for the hosted rate today rather than planning around free self-hosted inference.

The verdict

GLM-5.2 earns a place in the coding-model conversation on the strength of its own numbers: real benchmark gains over its GLM-5.1 predecessor, a genuine (if not category-leading) showing against GPT-5.5 and Opus 4.7, and pricing that undercuts every closed frontier competitor by a wide margin. The gap to Claude Opus 4.8 on Terminal-Bench 2.1 is real, and every score here is Z.AI's own benchmark run rather than an independent one — read the comparison set, not just the headline claim.

Go with GLM-5.2 if coding and agent work is the job and budget matters, and you're comfortable with a hosted API rather than self-hosting a 753B model. Stick with Claude Opus 4.8 if you need the top score on Terminal-Bench regardless of cost. For a wider field comparison, the best free coding model piece and the coding assistants shootout both sit next to this review.