Z.ai ships GLM-5.3 coding model, weights held for cyber review
Live now on the GLM Coding Plan and ZCode 3.0, it tops GPT-5.6 Sol on agentic and cyber benchmarks but trails on raw SWE - and you can't self-host it yet.
Copy markdown
It's live now - on the plan, not the API
GLM-5.3 shipped Aug 14 via the GLM Coding Plan ($12.60-$117.60/mo, 10k-140k weekly credits) and the ZCode 3.0 agent - not yet as a standalone API. It's a 743B-parameter model tuned for coding and defensive cyber work, usable in your editor today.
The benchmark split: pick your model per task
Vendor numbers put GLM-5.3 first on AutomationBench (48.2%) and CyberGym (84.5%), edging Claude Fable 5 and GPT-5.6 Sol - but it trails on ExploitBench (54.4 vs 78.0) and DeepSWE (66.9 vs 72.7). Strong for agentic and cyber-defense loops, weaker on raw SWE tasks.
Open weights are on hold for a safety review
Unlike past GLM drops, the weights aren't out: Z.ai staged them behind a ~2-week cyber-safety review, targeting roughly Aug 28 (a company timeline, not a promise). Its headline '2,436 vulnerabilities found' figure is mostly older, screened findings across GLM generations - only 53 were publicly disclosed.
Elsewhere: OpenAI's Assistants API sunsets Aug 26
If you built on the Assistants API, it shuts down Aug 26 - migrate to the Responses + Conversations APIs. That's eight days out; audit any agents or threads still pinned to the old endpoint now.
Elsewhere: Sonnet 5's $2/$10 rate reportedly stays
Claude Sonnet 5's introductory $2/$10 per-MTok pricing was set to expire Aug 31 and step up to $3/$15 on Sept 1. Multiple outlets now report Anthropic cancelled that hike and made the low rate permanent - confirm against the official pricing page before you re-budget.