GLM 5.2 + ZCode is HERE (Better than Claude Code?)

ZCode launched as the native agentic IDE for GLM 5.2 — 1.5x quota, 1M context, BYOK.

What ZCode actually ships

ZCode is Z.ai's first-party dev environment for GLM 5.2, available on macOS, Windows, and Linux. It supports bring-your-own-key, so any Chinese model subscription plugs in — not just a GLM coding plan. The editor is built specifically for long-horizon tasks: expect 30 minutes or an hour from your task completions rather than the short bursts Claude Code optimizes for. That matters because GLM 5.2 ships with a 1 million context window, which Claude Code cannot fully exploit in an interactive loop.

The pricing math vs. Claude Code

If you already pay for a GLM coding plan, quota multipliers inside ZCode are: 1.5x usage quota, peak hours drop from 3x to 2x, and off-peak goes from 1x to 0.67x. Net effect: higher effective throughput at lower peak cost. Boxmining's mid-recording conclusion: maybe for the cost factor alone, Zcode might make more sense than using it on Claude code.

Benchmarks and ecosystem signal

GLM 5.2 is the first open model to lead a category on Apex SWE — specifically 55.3% pass@1 on integration, with a 37.3% pass@1 overall (#6 on the leaderboard). LangChain published official integration guides and called out developers using GLM 5.2 as a daily driver, which Boxmining frames as an open-model credibility threshold being crossed.

Speculative decoding is spreading

vLLM now has native DeepSpark speculative decoding for DeepSeek: ~250 tokens/sec on 8xB300, beating MTP on acceptance rate. DSpark preview on GLM 5.2 claims ~1.5x faster decode. A separate DFlash drafter on Qwen 3 32B reports ~50% higher throughput via diffusion-based drafting. Different mechanisms, same goal: speed up open inference without retraining the target.