GLM-5.2 Benchmarks
How Z.ai’s GLM-5.2 stacks up against the frontier (Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro) and its predecessor GLM-5.1 — across long-horizon coding, standard coding, reasoning, and agentic benchmarks, plus the MTP speculative-decoding gain. The strongest open-source model at launch.
Score (higher = better). GLM-5.2 trails Opus 4.8 narrowly and is the top open-source model.
View data & sources →Score (higher = better). GLM-5.2 closes much of the gap to the closed-source frontier.
View data & sources →Score (higher = better). GLM-5.2 leads on AIME 2026; competitive across the board.
View data & sources →Score (higher = better). Tool use and multi-step agent tasks.
View data & sources →Score (higher = better). The biggest jumps over the predecessor.
View data & sources →Ablation on coding scenarios. Each technique lifts acceptance length; +20% over baseline.
View data & sources →Data table
| glm51 | glm52 | gpt55 | gemini | opus48 | series | benchmark | source_ref | value_basis | step | acc_len |
|---|---|---|---|---|---|---|---|---|---|---|
| 30.5 | 74.4 | 72.6 | 39.6 | 75.1 | coding_long | FrontierSWE | zai-glm52 | Full Benchmark Table — FrontierSWE (dominance as of 2026-06-16) | ||
| 20.1 | 34.3 | 28.4 | 21.6 | 37.2 | coding_long | PostTrainBench | zai-glm52 | Full Benchmark Table — PostTrainBench | ||
| 1 | 13 | 12 | 4 | 26 | coding_long | SWE-Marathon | zai-glm52 | Full Benchmark Table — SWE-Marathon | ||
| 63.5 | 81 | 84 | 74 | 85 | coding_std | Terminal-Bench 2.1 | zai-glm52 | Full Benchmark Table — Terminal-Bench 2.1 (Terminus-2) | ||
| 58.4 | 62.1 | 58.6 | 54.2 | 69.2 | coding_std | SWE-bench Pro | zai-glm52 | Full Benchmark Table — SWE-bench Pro |
23 more rows + CSV download
The full 28-row dataset, one-click CSV export, and the AI-ready context file are free with an account. Prefer to verify it yourself? The full methodology and sources are published below.
Methodology & sources
Last updated: Jul 17, 2026Methodology
Source-backed values for all six charts come from the Z.ai GLM-5.2 launch post (2026-06-16): long-horizon coding (FrontierSWE, PostTrainBench, SWE-Marathon), standard coding (Terminal-Bench 2.1, SWE-bench Pro, NL2Repo, DeepSWE, ProgramBench), reasoning (HLE, CritPt, AIME, HMMT, GPQA-Diamond, IMOAnswerBench), agentic (MCP-Atlas, Tool-Decathlon), the GLM-5.2 vs GLM-5.1 generational leap, and the MTP acceptance-length ablation. Every numeric point carries a sources[].ref and a value_basis naming the table row. Scores are reproduced verbatim from the Full Benchmark Table; cells the vendor published as "-" (not reported) are omitted, never estimated. CAVEAT: these are the model developer’s self-reported figures under their stated harnesses and prompts (see the post’s footnotes); some competitor HLE/CritPt cells are full-set scores. They are comparison-as-published, not an independent eval. Architecture context (not charted): 1M-token context (up from 200K), IndexShare cuts per-token indexer FLOPs ~2.9× at 1M, MIT-licensed open weights. Re-verified 2026-06-22.
Comparisons are informative, not definitive. See each source for definitions and limits.
How relevant was this information?
Continue researching
Co-related indexes
Related articles
No articles in this sector yet. Browse all research →