GLM 5.3

The llm-stats.com profile corroborates Z.ai's GLM-5.3 as a large-scale reasoning model for software engineering and agent tasks, ranking it 12th overall on the LLM Stats Score composite with tier placements of S in Reasoning (7 of 371), A in Tool Calling (8 of 202) and Coding (9 of 275), and B in Vision. The page reports a GDPval-AA Elo of 1769/3000 and a Terminal-Bench 2.1 score of 0.88, with performance holding near baseline across conversation turns 1 through 10. Quality Tracker shows GLM-5.3 improving at +1.72 sigma with a 20-vote sample as of early October 2026, and a 7-day window of +0.00 sigma against a 30-day baseline of +0.15 sigma. Blended aggregator pricing on llm-stats is shown at $1.33 per million tokens, but this figure reflects routed provider hosts and should not be read as a Z.ai list price for the model itself.

GLM-5.3 from Z.ai is a coding-focused update that reuses the GLM-5.2 base model and derives all gains from scaled post-training rather than retraining, according to a GMI Cloud blog post dated August 25, 2026. The lab scaled training environments, task diversity, and compute using existing infrastructure including IndexCache, SAO, and the asynchronous slime training stack. The model is live through the GLM Coding Plan, ZCode, and API, with open weights held back roughly two weeks for safety review. Lab-reported benchmarks show GLM-5.3 jumping from 4.6 to 28.3 on Terminal-Bench 3.0 (a roughly 6x gain), with Terminal-Bench 2.1 at 88.2, DeepSWE v1.1 at 66.9, Agents' Last Exam at 28.5, HLE with tools at 62.5, Toolathlon Verified at 73.0, and AutomationBench at 48.2. The GMI Cloud post says GLM-5.3 leads open-weight models on Terminal-Bench 3.0, while Claude Fable 5 (33.7) and GPT-5.6 Sol (34.6) score higher among closed models. Gains are attributed to a pipeline that synthesizes runnable long-horizon tasks with verifier-built reward signals.