GLM-5.3

Best for

Agentic coding teams that want near-frontier terminal/SWE performance with open weights and low cost

Price & access

$1.40 in / $4.40 out per 1M with $0.26 cached input (Z.ai docs; cached-input storage free for a limited time). The GLM Coding Plan uses points, with 50% off-peak discount outside 14:00-18:00 UTC+8 on weekdays.

Evidence summary

First-party Z.ai launch post and Hugging Face card give detailed benchmarks with context/output settings, and Z.ai's own pricing page gives exact rates. Weakened by the missing SWE-bench-family and Design Arena coverage and by unclear licence text.

Edition status

No prior snapshot. Tracked in the September 2026 baseline.

Strengths

• Terminal Bench 2.1 88.2 is effectively tied with Kimi K3 (88.3) and DeepSeek V4 Pro (87.9)
• Huge generational jump: Terminal Bench 3.0 4.6 -> 28.3 and DeepSWE 46.2 -> 66.9 vs GLM-5.2
• Open weights (753B) with wide inference-engine support and $1.40/$4.40 hosted pricing
• TTFT 2.14 s with 84.2 tok/s - responsive for a frontier-scale model

Tradeoffs

• No SWE-bench Pro, SWE-bench Verified or Design Arena row exists for 5.3
• Thinking cannot be disabled, forcing a migration for non-thinking apps
• Licence terms are not stated on the Hugging Face card; AA calls it a restricted commercial licence
• Text-only on the evidence available - no documented image input for frontend screenshots