August 8, 2026

7 min read

Artificial Intelligence

GLM-5.2: Zhipu’s Open-Weight Model Nears Opus 4.8

> Zhipu’s GLM-5.2 is an MIT-licensed 744B MoE model that nears Claude Opus 4.8 on coding benchmarks at a fraction of the cost, with real self-hosting tradeoffs.

ShareX LinkedIn

🎧 Listen — ~7 min

Ready · GLM-5.2: Zhipu’s Open-Weight Mod

0:00 / 7:00

Verified by Essa Mamdani

Zhipu AI, operating under the brand Z.ai, released GLM-5.2 on June 13, 2026 as an MIT-licensed, 744-billion-parameter Mixture-of-Experts model built specifically for long-horizon coding agents. Independent trackers have since converged on a consistent read: it is the strongest open-weight coding model available, it lands within a few points of Claude Opus 4.8 on several benchmarks, and it does so at roughly a fifth to a sixth of the API cost. It is not a frontier-beating model outright, but for developers deciding what to route agentic coding traffic to, it has become one of the defaults worth testing.

Architecture and access

GLM-5.2 is a sparse MoE model with roughly 744 billion total parameters and about 40 billion active per token, depending on which provider's model card you read (Together AI lists 40B active against a 256K context window for its serving configuration, while the headline spec most vendors cite is a 1,048,576-token context window with up to 131,072 output tokens). It supports two selectable reasoning effort levels — High and Max — with Z.ai recommending Max for complex, multi-step coding work that needs planning and revision across long sequences, and High for faster everyday use. The weights are MIT licensed, meaning unrestricted commercial use, self-hosting, and modification, and they're already available through Hugging Face, Ollama, and llama.cpp/Unsloth GGUF builds for local inference.

API pricing varies noticeably by provider — anywhere from roughly $0.68 to $1.50 per million input tokens and $2.14 to $4.40 per million output tokens, depending on whether you're hitting Z.ai directly or a reseller like Together AI, OpenRouter, or DeepInfra. Even at the higher end of that range, GLM-5.2 costs a fraction of Claude Opus 4.8's per-token pricing, which is the core of its value proposition: near-frontier agentic coding performance at open-weight economics.

Benchmark reality: close, but not equal, to Opus 4.8

Z.ai's own published numbers put GLM-5.2 within striking distance of Claude Opus 4.8 on several coding benchmarks, and clearly ahead of GPT-5.5:

Terminal-Bench 2.1: GLM-5.2 scores 81.0 against Opus 4.8's 85.0 — a 4-point gap, and a large generational jump from GLM-5.1's 63.5.
SWE-bench Pro: GLM-5.2 scores 62.1, ahead of GPT-5.5's 58.6 but behind Opus 4.8's 69.2 — the widest gap in Z.ai's own table.
FrontierSWE: GLM-5.2 scores 74.4 against Opus 4.8's 75.1, a gap under one point on long-horizon engineering tasks.
MCP-Atlas (tool use): GLM-5.2 scores roughly 77.0 against Opus 4.8's 77.8.
AIME26 (math): GLM-5.2 leads the field at 99.2%, ahead of Inkling (97.1%) and Kimi K2.6 (96.4%).

The picture flips on the hardest, longest-running agentic tasks. On SWE-Marathon, an extended multi-hour engineering benchmark, GLM-5.2 scores 13.0 against Opus 4.8's 26.0 — Opus is exactly twice as capable by that measure. NL2Repo shows a similar gap (48.9 vs 69.7), and Tool-Decathlon shows GLM-5.2 at 48.2 against Opus 4.8's 59.9. The pattern across independent write-ups (Layer3Labs, Groundy, Emergent, BitsMinds) is consistent: GLM-5.2 is competitive to near-parity on standard coding and reasoning benchmarks, but the gap widens on ultra-long-horizon, many-step agentic workloads where sustained coherence over hours matters more than any single capability.

It's worth being explicit about evidence class here, the same way independent trackers do: the specific point scores above (SWE-bench Pro, Terminal-Bench, FrontierSWE, MCP-Atlas) are Zhipu's own reported numbers, not independently reproduced by a third party running the exact same harness. They are directionally corroborated — GLM-5.2 does rank #1 on Design Arena's blind human-preference coding leaderboard, and firms like Artificial Analysis, Scale AI, and BenchLM have each published independent evaluations treating it as genuinely competitive — but the precise decimal-point numbers in Zhipu's table should be read as vendor-reported until someone else runs the same suite and publishes matching figures. One community benchmark write-up also flagged harness variance: Opus 4.8 scored 85.0 on Terminal-Bench 2.1 in Zhipu's table but 74.6 in another Z.ai report using a different harness configuration, which is a reminder that "which harness" changes the number as much as "which model."

Where independent testers push back

Reddit and community benchmarking threads add useful texture that vendor tables don't capture. One widely shared test ran GLM-5.2 against Opus 4.8 on 50 real Go and Rust pull requests pulled from open source repositories and found GLM-5.2 came in last on output quality — and wasn't even the cheapest option once total token usage was accounted for, since it tends to use more tokens per task to reach a comparable result. Another thread reported GLM-5.2 matching Claude on 45 terminal-bench coding-agent tasks at less than half the cost, with the caveat that task selection matters enormously in these informal comparisons. The honest takeaway is that GLM-5.2's benchmark strength shows up most reliably on structured evals and less reliably on messy, real-world repositories with idiosyncratic tooling — which is exactly the gap you'd expect between "scores well on FrontierSWE" and "handles your actual codebase's weird build system."

Self-hosting: what it actually costs

Because the weights are MIT licensed and openly published, self-hosting is a real option, not just a marketing bullet point. Community write-ups on running GLM-5.2 with vLLM put full-precision self-hosting at roughly 8×H200 GPUs, which puts it well outside hobbyist territory and into serious infrastructure spend — one comparison pegged the break-even against a $30/month cloud API tier at a scale most individual developers will never hit. For anyone without dedicated GPU infrastructure, the practical path is the hosted API through Z.ai directly or a reseller, not local deployment.

How to actually decide between GLM-5.2 and Opus 4.8

The independent comparisons that have accumulated since June converge on a reasonably simple decision rule. Reach for GLM-5.2 when you're running high-volume agentic coding workloads where cost per task matters more than squeezing out the last few points of quality, when you need self-hosting or open weights for compliance or data-residency reasons, or when your tasks fit inside standard coding-benchmark shapes (bug fixes, feature implementation, test writing) rather than multi-hour autonomous engineering runs. Reach for Opus 4.8 when the task is the hardest 10% of your backlog — long-horizon refactors, ambiguous multi-file changes, or anything where SWE-Marathon-style sustained reasoning matters — or when you're working in a codebase with real-world mess that benchmark suites don't capture well.

For teams already running a model router across coding agents, the practical move many teams are making is exactly that: route routine agentic coding tasks to GLM-5.2 for the cost savings, and escalate to Opus 4.8 (or another closed frontier model) for the tasks that actually need it. That router pattern is cheaper to build than it sounds — a simple complexity heuristic on task description length and file count gets most of the value.

FAQ

Is GLM-5.2 actually open source, or just "open weights"?

The weights are published under the MIT license, which is a genuinely permissive open-source license allowing commercial use, modification, and redistribution. That's a stronger claim than many "open" model releases, which use custom or more restrictive licenses.

What's the real context window — 1M tokens or 256K?

Different providers serve different configurations. Z.ai's own documentation and most resellers list 1,048,576 tokens as the maximum context window with up to 131,072 output tokens, but some serving configurations (like certain Together AI deployments) cap at 256K. Check your specific provider's model card before relying on the full 1M window.

Does GLM-5.2 beat GPT-5.5?

On the coding benchmarks Zhipu published, yes — GLM-5.2 leads GPT-5.5 on SWE-bench Pro (62.1 vs 58.6), FrontierSWE, and AIME26 math. It's a closer comparison against Claude Opus 4.8, where GLM-5.2 trails on most benchmarks but narrows the gap significantly compared to GLM-5.1.

Can I run GLM-5.2 on a single GPU?

Not at full precision — community estimates put full self-hosting at roughly 8×H200 GPUs. GGUF quantized builds from Unsloth reduce that requirement substantially for local experimentation, though with the usual quality tradeoffs of quantization on a MoE model this large.

Is GLM-5.2 good for long-running autonomous agent tasks?

This is its weakest area relative to Opus 4.8. On SWE-Marathon, an extended multi-hour engineering benchmark, GLM-5.2 scores roughly half of Opus 4.8's result. For short-to-medium agentic coding tasks it's competitive; for the longest autonomous runs, the gap widens.

Source notes

Author context

I write about AI systems and full-stack engineering from the point where product behavior meets production reality. My bias is toward explicit contracts, narrow tool surfaces, and enough telemetry to explain failures without exporting private data everywhere.

For help building a model-routing layer that sends routine agentic coding tasks to cheaper open-weight models like GLM-5.2 and escalates the hard ones, I'm available through /hire.