glm-5.2
Meant to excel at
-
Z.ai's flagship for long-horizon tasks, with a stable 1M-token context and selectable thinking effort levels.
Vendor's own site huggingface.co -
Strong agentic and terminal work: Terminal Bench 2.1 81.0, with big gains in FrontierSWE over GLM-5.1; AIME 2026 99.2.
Vendor's own site huggingface.co -
Open weights under MIT (753B MoE, about 40B active); works as a drop-in model in Claude Code, Cline and OpenCode.
Vendor claim, relayed rits.shanghai.nyu.edu -
Structured extraction from messy HTML and tool calling matched Sonnet at roughly one-seventh to one-eighth the cost.
Independent evaluator zyte.com
Watch out for
-
Behind Opus 4.8 on agentic coding (SWE-bench Pro 62.1 vs 69.2; Tool-Decathlon 48.2 vs 59.9); numbers self-reported.
Vendor's own site huggingface.co -
Verbose and slower: about 43K tokens per task versus 26K for GLM-5.1; 24 to 55s on scraping tasks.
Independent evaluator zyte.com -
Stale library knowledge: generated third-party API config missing required settings.
Independent evaluator zyte.com -
React output less consistent under multi-step agentic pressure and it stumbled on TypeScript compile errors before recovering.
Practitioner report lennysnewsletter.com
Benchmarks
Coding percentile 0.04
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| code arena webdev | 1604.6 | 0.00 | Independent evaluator llmrun.dev | undated |
| lmarena code arena webdev | 1604.6 | 0.09 | Independent evaluator lmarena.ai | undated |
Agentic coding percentile 0.42
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| swe-bench pro | 62.1 | 0.50 | Vendor's own site huggingface.co | 2026-06-17 |
| swe-bench verified | 79 | 0.40 | Independent evaluator epoch.ai | undated |
| terminal-bench 2.1 | 81 | 0.36 | Vendor's own site huggingface.co | 2026-06-17 |
Reasoning percentile 0.17
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| gpqa diamond | 92 | 0.17 | Independent evaluator epoch.ai | undated |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +/-
A month-long user says it root-caused a main-thread lag and a battery polling loop, verified fixes with tests and kept PRs focused, but is a token glutton, cannot read screenshots and has immature tool calls; close to Opus 4.8 for daily coding but short on visual work.
Task: bug-fix | Compared with claude-opus-4.8 | 2026-07-08 older than 90 days
Practitioner report rafay99.com - +/-
A 45-minute Claude Code run on Sentry and Vercel logs found two P0 issues monitoring had missed for $3.36, but React consistency was a worry; she recommends using it in rotation.
Compared with claude-opus-4.8 | 2026-06-29 older than 90 days
Practitioner report lennysnewsletter.com - -
Kilo's code-review test with planted bugs: GLM-5.2 caught 13-15 of 16 obvious bugs but only 4-7 of 10 subtle cross-route bugs, trailing GPT-5.5 (9/10) and Opus 4.8 (8/10); results varied run to run and with prompt wording.
Task: code-review | Compared with gpt-5.5 | 2026-06-26 older than 90 days
Practitioner report blog.kilo.ai - +/-
ChatPRD's live review: fast, solid at repo summaries and HTML/UI iteration, and ran a 30-45 minute autonomous Sentry/Vercel triage into a 14-fix plan, but stalled on complex React/TypeScript and needed human review.
Task: operations-support | 2026-06-24 older than 90 days
Practitioner report chatprd.ai - +/-
Zyte's scraping tests found GLM-5.2 matched Sonnet 4.6 on extraction and tool calling at roughly 2-8x lower cost, but it was verbose, slower, and used outdated Zyte API setup; not yet trusted for hardest long-horizon planning where Opus leads.
Task: feature | Compared with claude-sonnet-4.6 | 2026-06-19 older than 90 days
Practitioner report zyte.com - +/-
Matched Sonnet on structured JSON extraction from messy HTML and tool calls at a fraction of the cost, but was verbose, slower and out of date on library details.
Compared with claude-sonnet-5 | 2026-06-19 older than 90 days
Practitioner report zyte.com - +
Added a search feature to a Next.js blog for about $0.27 in OpenCode; asked sensible clarifying questions and kept the page server-rendered, but used more tokens than GLM-5.1.
Compared with glm-5.1 | 2026-06-18 older than 90 days
Practitioner report dev.to - +
Built a Matrix bot with encryption and a Rust tool-using agent for about $20 via Fireworks; found it fast and pleasant, far cheaper than Opus or GPT.
Compared with claude-opus-4.8 undated
Practitioner report news.ycombinator.com - +/-
Benchmark author says multi-agent coding is just shy of Opus 4.6 on average, but cost-adjusted GLM is frontier; not the pick if you want only the top model.
Compared with claude-opus-4.6 undated
Practitioner report news.ycombinator.com - +
Calls GLM-5.2 the third model, after Opus 4.6+ and GPT-5.5, able to sustain long-horizon agentic work.
Compared with gemini-3.1-pro undated
Practitioner report news.ycombinator.com - -
Needs more guidance than Opus to reach an acceptable plan, which takes Opus less time.
Compared with claude-opus-4.8 undated
Practitioner report news.ycombinator.com - +
For Elixir work it beat Google and Anthropic models that rank higher on general benchmarks.
Compared with gemini-3.1-pro undated
Practitioner report news.ycombinator.com