For my research, Fable felt like a clear step change … I was excited about the GLM 5.2 hype and tried it; sadly it's nowhere close
Task Evaluating models for research work (alongside Fable and GPT-5.5 Pro).
Z.ai · GLM. Released Jun 16, 2026.
GLM-5.2 is a model from Z.ai in the GLM family, released Jun 16, 2026. evals.report tracks 14 reported GLM-5.2 benchmark scores across ARC-AGI-1, ARC-AGI-2, Humanity's Last Exam, AIME 2026, GPQA Diamond, MathArena HMMT February 2026, SWE-bench Pro, DeepSWE, and 6 more — each shown with its benchmark, metric, source status, and date, and never combined into a single ranking.
| Benchmark | Category | Score | Metric | Status | Date | |
|---|---|---|---|---|---|---|
| ARC-AGI-1 | Reasoning | 77% | accuracy | Official | Jun 16, 2026 | Details |
| ARC-AGI-2 | Reasoning | 22.78% | accuracy | Official | Jun 16, 2026 | Details |
| Humanity's Last Exam | Reasoning | 40.5% | accuracy | Verified | Jun 16, 2026 | Details |
| AIME 2026 | Reasoning | 99.2% | accuracy | Verified | Jun 16, 2026 | Details |
| GPQA Diamond | Reasoning | 91.2% | accuracy | Verified | Jun 16, 2026 | Details |
| MathArena HMMT February 2026 | Reasoning | 92.5% | accuracy | Verified | Jun 16, 2026 | Details |
| SWE-bench Pro | Coding | 62.1% | % resolved | Verified | Jun 16, 2026 | Details |
| DeepSWE | Coding | 46.2% | % resolved | Verified | Jun 16, 2026 | Details |
| Terminal-Bench 2.1 | Agents | 81.0% | task success | Verified | Jun 16, 2026 | Details |
| MCP Atlas | Tool use | 76.8% | pass rate | Verified | Jun 16, 2026 | Details |
| SWE-Marathon | Agents | 13.0% | resolution rate (pass@1) | Verified | Jun 16, 2026 | Details |
| FrontierSWE | Agents | 74% | dominance score | Official | Jun 16, 2026 | Details |
| PostTrainBench | Agents | 34.3% | weighted average score | Verified | Jun 16, 2026 | Details |
| FrontierCode | Coding | 24.5% | weighted score (Main) | Official | Jun 16, 2026 | Details |
Real-world feedback on GLM-5.2 from people using it on actual prompts — praise and criticism alike, each linked to its source. Qualitative, never scored.
Report tone
Report type
Topic
For my research, Fable felt like a clear step change … I was excited about the GLM 5.2 hype and tried it; sadly it's nowhere close
Task Evaluating models for research work (alongside Fable and GPT-5.5 Pro).
GLM 5.2 ranks unusually high on FrontierSWE (long-horizon agentic engineering) … using it with OpenCode is also not far from the quality of Claude Code or Codex.
Task Day-to-day agentic coding with GLM-5.2 in OpenCode.
Genuinely impressed, almost shocked, at how good GLM-5.2 … is at coding. This changes things.
Having an open weight model surpass GPT-5.4 and every Gemini model is dope. That said - it's not cheap. Both Opus 4.8 and GPT-5.5 set to "medium" are cheaper and smarter than GLM-5.2
Asked GLM 5.2 (left) and Opus 4.8 (right) to design a menu and GLM has even better taste while being 4x cheaper. GLM also nailed a lot of small details like adding "chef's pick" and "vegetatian" tags to some dishes.
Task Asked GLM-5.2 and Opus 4.8 to design a restaurant menu (UI/design).