Public benchmarkExternal source reviewReviewed by DanPublished Source checked Updated
GLM-5.3: public coding benchmark receipts
The report
Verdict
Published benchmark evidence
Z.ai reports 66.9% on DeepSWE v1.1, 42.5% on SWE-Marathon v1.1 and 28.3% on the older Terminal-Bench 3.0: different tasks and benchmark generations.
On the tier list: The November 17 tier is Dan’s own ranking; public benchmark numbers are a separate evidence channel, not an automatic tier adjustment.
What the benchmarks report
- DeepSWE v1.1: 66.9%. GLM-5.3; mini-swe-agent harness. Reported by Z.ai vendor release.
- Terminal-Bench 3.0: 28.3%. GLM-5.3; Claude Code 2.1.207; max effort; avg@3. Reported by Z.ai vendor release.
- SWE-Marathon v1.1: 42.5%. GLM-5.3; long-horizon software engineering. Reported by Z.ai vendor release.
Original sources
Source 1 — DeepSWE v1.1, Terminal-Bench 3.0, SWE-Marathon v1.1.
What this evidence tells a reader
DeepSWE v1.1 at 66.9% uses the mini-swe-agent harness, while the 28.3% Terminal-Bench 3.0 figure names Claude Code 2.1.207 at max effort and avg@3. That last notation changes the interpretation: the reported value reflects the publisher's aggregation method, not necessarily a single run a developer will experience.
Terminal-Bench 3.0 must not be placed directly beside Terminal-Bench 4.0 values for Opus, Sonnet or Kimi as though they were common-exam rankings. SWE-Marathon v1.1's 42.5% offers evidence about another category—longer engineering work—but it cannot calibrate the earlier terminal percentage.
Readers selecting GLM for a coding agent should begin with a fixed project repair and record code correctness, task completion, tool compatibility and review time. A vendor's three-table profile is useful for deciding what to test next, not enough to declare a cost-efficient winner.
How to use these results
- Check both benchmark generation and harness before making cross-model comparisons.
- Treat avg@3 separately from single-attempt success.
- Run a representative multi-step fix and count accepted patches and correction passes.
Limits of this report
- The cited results were published in Z.ai's model release, without an independent matched local reproduction here.
- Different suite generations and aggregation rules prevent a reliable score ranking against Terminal-Bench 4.0 pages.
Follow the lab
New test reports and tier-list changes get announced in my Discord's #news channel. The full videos are on YouTube.
Receipts
DanManREAL · AI lab
Test receipt
GLM-5.3
- DeepSWE v1.1
- 66.9%Z.ai vendor release (model publisher). GLM-5.3; mini-swe-agent harness. Not independently rerun by DanManREAL.
- Terminal-Bench 3.0
- 28.3%Z.ai vendor release (model publisher). GLM-5.3; Claude Code 2.1.207; max effort; avg@3. Not independently rerun by DanManREAL.
- SWE-Marathon v1.1
- 42.5%Z.ai vendor release (model publisher). GLM-5.3; long-horizon software engineering. Not independently rerun by DanManREAL.
- Source review
- November 17, 2026Date these source claims were reviewed; not a claim about when the published benchmark was executed.
Verdict
Published benchmark evidence
Tier list The November 17 tier is Dan’s own ranking; public benchmark numbers are a separate evidence channel, not an automatic tier adjustment.
danmanreal.com/tests/glm-5-3-public-benchmarks
Models covered by external benchmarks
On the tier list's bench
- GLM-5.3 · 2 receipts on file
Related
Independent testing. Sponsorships are always disclosed, and verdicts are never for sale.
Disagree? Bring your receiptsAll reports
- One inbox for my AI coding agents: triage took 34% less time, resuming work did not improveHead-to-headNov 20
- Mythos 5.2: research and security benchmark source reviewPublic benchmarkNov 17
- Fable 5.2: agentic coding benchmark source reviewPublic benchmarkNov 17
- Claude Opus 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- Astra 6.2: new model and public task-evaluation evidencePublic benchmarkNov 17
- Sol 6.2: new model and public coding benchmark evidencePublic benchmarkNov 17
- Gemini 4 Carbon: benchmark source and usage profilePublic benchmarkNov 17
- Qwen 4: Qwen4-Max launch and benchmark sourcesPublic benchmarkNov 17
- Kimi K3.1 Pro: public coding benchmark receiptsPublic benchmarkNov 17
- GLM-5.3: public coding benchmark receiptsPublic benchmarkNov 17
- Gemini 4 Argon: public coding benchmark receiptsPublic benchmarkNov 17
- Claude Sonnet 5.5: public coding benchmark receiptsPublic benchmarkNov 17
- DeepSeek V4.1 Flash: public coding benchmark receiptsPublic benchmarkNov 17
- Muse Wonder Full 1.0: public coding benchmark receiptsPublic benchmarkNov 17
- Grok 4.7: separate release and public coding resultsPublic benchmarkNov 17
- ai-design-check on two of my FABLE screens: 10 warnings, 4 of them false alarmsField testNov 2
- Vercel Agent Desktop used a sample Trip app in 6 actions without moving the mouseField testNov 1
- Cerberus Guard on sample data: one crawler name, two verdicts, and a block earned by behaviorField testOct 31
- GPT-6.1 Terra vs Opus 5.5: three real projects, and the designs I picked blindHead-to-headOct 29
- Steps → Reduced: opening a terminal agent went from 98 seconds to 11AuditOct 28
- One ChatGPT chat directed Opus and Codex at once through DMR_REMOTEField testOct 28
- Claude Duet ran my channel for a week and caught 269 fake AI crawlersField testOct 27
- Kimi K3.1 Pro built my animation family, after a quota stop and 21 minutes of failed requestsField testOct 24
- ChatGPT + Codex built 33 new DMR_REMOTE toolsField testOct 23
- Gemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistencyField testOct 21
- Qwen 3.8 Flash Next on my RTX 5080: 81 tokens a second with 202K tokens readAuditOct 21
- Does an extra AI review step pay off? A small DeepSeek trial said noHead-to-headOct 19
- My real DeepSeek bill: 12.31 billion tokens for $122.28Cost receiptOct 17
- Gemini 4 Argon kept hanging in Gemini CLI. The cause was a 60-second timeoutIncidentOct 14
- I asked 10 AI coding agents for their biggest weaknessSelf-reportOct 14
- Jarvis 27B on my own PC: 30 tokens a second, and it still recalls at 251KAuditOct 10
- My agents' first game benchmark was wrong. Here's how it was caughtIncidentOct 10
- Pixal3D + Blender turned my 2D art into animated 3D charactersField testSep 22
- DeepSeek Flash answered 200 OK and sent nothing for two hoursIncidentSep 14
- Grok, Codex and DeepSeek played WorldOS for Rank 1. Grok won with 98Head-to-headAug 20
- DeepSeek V4 Flash audited 18 of my real coding sessions. Two said 'done' too earlyAuditAug 4