GLM-5.3: strongest open coding model yet, usable in Claude Code now
Same 743B base as 5.2, huge agentic gains for $18/mo — plus emergent exploit-chaining that surfaced thousands of real bugs. Weights land in ~two weeks.
Copy markdown
Every gain came from post-training
GLM-5.3 reuses the exact 743B base from 5.2 — the jumps are all scaled post-training. Terminal-Bench 3.0 climbs 4.6→28.3, DeepSWE v1.1 46.2→66.9, and Z.ai's internal code bench is up ~50%.
Drop it into Claude Code today
It's live now via the Z.ai API and the GLM Coding Plan (Lite $18, Pro $80, Max $168 per month) and slots straight into Claude Code, Codex, OpenCode, and ZCode. A 1M-token route ships as the glm-5.3[1m] model suffix.
Open weights are still ~two weeks out
Zhipu says it'll publish the weights roughly two weeks after launch, once safety hardening finishes — so today it's API/subscription only. No 5.3 token price is posted yet; the table still lists 5.2 at $1.40 in / $4.40 out per million.
The cyber skill it didn't plan for
On CyberGym it hit 84.5% (from 77.2%) and ExploitBench 54.4% (from 24.4%). Z.ai reports it surfaced 2,436 real vulnerabilities across 269 projects — logged publicly at cvd.z.ai — and began chaining multi-stage exploits on its own.
Elsewhere: Claude Code forks subagents by default
Claude Code 2.1.232 turns subagent forking on by default — with full conversation and prompt-cache inheritance — and adds @-mentions so one session can message another live session by name.