GLM-5.3 lands: frontier agentic coding, but the weights are held back
Z.ai's terminal and SWE gains route into Claude Code today; its cyber skills delay the weights. Plus DeepSeek V4-Pro ships open and ChatGPT gains memory.
Copy markdown
The agentic-coding jump is real
On Z.ai's own numbers GLM-5.3 hits 28.3 on Terminal-Bench 3.0 (up from 4.6) and ~66.9 on DeepSWE (up from 46.2) — a real leap on terminal and repo-scale agent tasks, not just chat.
Use it today — even inside Claude Code
It's live now through the GLM Coding Plan (from about $18/mo) and Z.ai's ZCode harness, and it routes into Claude Code, Cline, and Roo Code — so you can A/B it against your current model without switching tools.
But the open weights are on hold
Unlike GLM-5.2's same-week MIT drop, 5.3's API and weights ship 'in stages' after safety review — roughly two weeks out — because of its emergent cyber-exploit skills (CyberGym 84.5, ExploitBench 54.4). Self-hosters wait.
DeepSeek V4-Pro goes GA — and fully open
The GA build is MIT-licensed on Hugging Face (~893GB, fp8/fp4) and served automatically via the deepseek-v4-pro alias, with DeepSWE jumping 12.8 to 62.7. Re-run your agent evals — and note the peak/off-peak pricing that starts Aug 16.
ChatGPT desktop starts remembering your screen
OpenAI's opt-in Computer History (macOS) turns your app and browser activity into memories ChatGPT and Codex can draw on. Pro, Business, and Enterprise only, and Business/Enterprise admins must switch it on first.