GLM-5.2 benchmarks: an open model at the frontier’s edge

GLM-5.2 benchmarks are finally here. On June 16, 2026 — three days after Z.ai pushed the model to its GLM Coding Plan with nothing but adjectives — the company published a full scorecard, opened the weights under the MIT License, and turned on the standalone API and chatbot. The headline a benchmark-less launch could only gesture at is now measured: GLM-5.2 is the strongest open-weight coding model available, and on several agentic benchmarks it trades blows with Claude Opus 4.8.

The single most quotable result comes from outside Z.ai. On Arena.ai’s Code Arena Frontend leaderboard — a human-preference board, not a self-report — GLM-5.2 (Max) ranks #2, behind only Anthropic’s Fable 5 and ahead of Claude Opus 4.7 and Opus 4.8 in thinking mode. An MIT-licensed model you can download today is beating two of the closed frontier’s flagships on frontend coding, as judged by developers.

This post reads the numbers honestly: the Code Arena placement, the agentic and long-horizon coding benchmarks, the full cross-vendor table, the open-weight architecture, the economics, and a clear verdict on where GLM-5.2 wins and where Opus 4.8 still leads. For how the launch itself was sequenced, see our GLM-5.2 launch-day coverage; for the model it improves on, our GLM-5.1 benchmark analysis.

When GLM-5.2 first appeared on June 13, it was a distribution-first launch: live inside the GLM Coding Plan, but with no API, no chatbot, no public weights, and — conspicuously — no benchmarks. We covered it as exactly that, a coding-plan rollout shipped “before the scoreboard.” Three days later the scoreboard arrived, and it is substantial. Everything Z.ai deferred is now shipped.

The model card is public, the architecture paper is out, the standalone API is priced and live, the Z.ai chatbot hosts GLM-5.2, the MIT-licensed weights are on Hugging Face and ModelScope, and a full cross-vendor benchmark table is published. Independent corroboration showed up too: Arena.ai’s Code Arena ranked GLM-5.2 against the field, and three external labs ran the long-horizon coding benchmarks. The qualifier we attached to every day-one claim — “unverified” — can now be lifted on most of them.

Self-reported benchmarks invite skepticism, so the most valuable GLM-5.2 result is the one Z.ai did not run. Arena.ai’s Code Arena Frontend leaderboard ranks models by blind, head-to-head human preference on real frontend coding tasks. There, GLM-5.2 (Max) lands at #2 with an Elo of 1,595 — behind only Fable 5 and ahead of every other model on the board, including Claude Opus 4.7 (Thinking) at 1,566 and Opus 4.8 (Thinking) at 1,561. Arena puts the gain at +29 points over Opus 4.7 Thinking.

For context, GLM-5.1 sits at #9 on the same board (1,531), so the jump from one generation to the next is large. Arena also reports GLM-5.2 as #2 on the React sub-board and #4 on HTML, and as the top-ranked model in several frontend categories — brand and marketing, reference-based design, data and analytics, consumer product, gaming, and simulations. It is, by Arena’s framing, the best open model on the board by a wide margin over Kimi K2.6 and MiniMax M3.

GLM-5.2’s pitch is long-horizon agentic work — staying coherent across hours of multi-step coding rather than nailing a single completion — so the benchmarks that matter most are the agentic ones. Three were run by external labs (Proximal, PostTrainBench, and Abundant AI) at 1M context and Max effort, which adds credibility. The pattern is consistent: GLM-5.2 is the top open model on every one, beats GPT-5.5 on most, and sits just below Opus 4.8 — with the gap widening only on the very hardest tasks.

Read together, these say something specific: GLM-5.2 has closed most of the distance to the closed frontier on mainstream agentic coding, but Opus 4.8 still pulls clear on the most demanding, longest tasks. That is a meaningful improvement over GLM-5.1, which our earlier analysis measured at roughly 94.6% of Opus 4.6’s coding score — GLM-5.2 is now comparing itself to Opus 4.8, not 4.6, and holding its own.

Z.ai published a wide cross-vendor table spanning reasoning, coding, and agentic tool use. The selection below pulls the load-bearing rows and places GLM-5.2 against its predecessor and the three closed frontier models most teams weigh. Across coding, GLM-5.2 is also the strongest open model versus Qwen 3.7 Max, MiniMax M3, and DeepSeek-V4-Pro (not shown). Higher is better in every row.

Unlike launch day, the weights are real now. GLM-5.2 is published on Hugging Face and ModelScope under the MIT License — the most permissive mainstream open license, and a step looser than the Apache-2.0 that GLM-5 shipped under. Z.ai frames it as “pure open”: no regional limits, no usage borders. The model card and config confirm the shape below, and our GLM-5 architecture analysis covers the lineage this builds on.

Benchmarks decide whether a model is good enough; economics decide whether you switch. GLM-5.2’s most consequential detail is that it costs exactly what GLM-5.1 did — $1.40 per million input tokens, $4.40 output, $0.26 cached — while delivering a generation’s worth of improvement. A near-frontier coding model at that price, with an open-weight fallback, is what reshapes the build-versus-buy calculation. There are three ways to run it, each with a different cost shape.

GLM-5.2 is the strongest open-weight coding model available in June 2026, and frontier-adjacent on agentic work — but it is not a frontier-beater across the board, and the honest matrix says so. Here is how it stacks against the realistic alternatives a team would weigh this month, now scored on confirmed numbers rather than launch-day adjectives.

Run your own evaluation — the cost is now near zero. With the API live at GLM-5.1 pricing and weights on Hugging Face, there is no longer a reason to take any benchmark on faith. Point a coding agent at glm-5.2 (or GLM-5.2[1m] in Claude Code), set Max effort, and run one real multi-file refactor or failing-test fix on a branch. Score end-to-end completion without intervention — that is the long-horizon claim, tested directly.

Match the model to the task tier. The benchmarks suggest a clean split: GLM-5.2 for the bulk of day-to-day agentic coding, where it is frontier-adjacent and far cheaper, and reserve Opus 4.8 for the hardest, longest tasks where it still pulls clear (large-scale refactors, multi-hour autonomous runs). A two-model routing setup captures most of the savings without giving up the ceiling.

Treat the open weights as a real lever now, not a someday. Unlike launch day, the MIT weights are downloadable, so a self-hosting or data-residency plan can actually be scoped. For regulated workloads it is a genuine path; for most teams it is a negotiating lever and a compliance fallback. Our AI transformation services cover model selection and deployment, and our web development team ships with these agentic stacks daily.

Re-check the leaderboards in a fortnight. Arena’s Elo will keep moving as vote volume grows, and independent SWE-bench and Terminal-Bench re-runs typically firm up within one to two weeks of an open-weight release. The picture here is strong and unusually well-corroborated for a day-of report, but the open weights mean the community will pressure-test it fast — let the standardize-on-it decision follow that.