Published 2026-08-16 · 林启明
Direct answer
The live catalog id is z-ai/glm-5-cn. The only CacheSafety row is for base GLM-5: 30% Safe Hit Rate, 0% Bad Hit Rate. The CN variant itself is unmeasured. This guide is written for Canada-based product and platform teams comparing model quality, spend, routing policy, and production rollout risk.
What We Actually Ran
Yes, worth calling for cheap, streaming, low-latency workloads where cache correctness is not load-bearing. But the CN build's own CacheSafety numbers are unmeasured; the only family data belongs to base GLM 5. Verify before you trust. We pulled the live catalog record for z-ai--glm-5-cn on 2026-08-16. We also read the first-party CacheSafety results table in data/cache-safety-results.ts. That bench run used 50 prompts, produced 50 valid responses, and logged 0 upstream errors. The measured row is keyed to modelSlug z-ai--glm-5 with modelVersion z-ai/glm-5 (2026-08 run), evaluated on 2026-08-16. That is the base GLM 5, not the CN variant you would call. The run used exact-cache semantics v1 with promptset v3. We did not run our own prompts against z-ai--glm-5-cn.
What the Numbers Mean
safeHitRate is 0.3, badHitRate is 0, semanticTrapFailureRate is 0. All 10 exact repeats hit the exact cache, but only 3 of 10 replayed the expected answer text. A cache hit therefore does not guarantee a correct replay. Zero trap-pair second calls and zero fresh-question calls hit the cache. That is the correct outcome for exact-cache semantics, where the cache key is the full exact prompt. One-word-different prompts do not collide, which protects against stale answers but also means no semantic caching benefit. costSavedPer1kRequests is null, so the economic value of the cache was not measured. These numbers describe the base GLM 5 build, not the CN build. The CN model's safe hit rate, bad hit rate, and trap failure rate are unmeasured.
What We Did Not Measure
Context window is unpublished in this catalog snapshot. Release date is unpublished. Independent third-party quality scores are absent from these notes. The 1000 to 3000ms latency band is a catalog claim, not an independent measurement. No cost-saved-per-request figure exists for the family. The CN variant has no measured cache safety data of its own.
When This Model Is the Wrong Pick
Pick something else if you need a published context window, because this entry does not provide one. Pick something else if exact-repeat cache correctness is load-bearing, because the sibling build only replayed the correct answer 30% of the time on exact repeats and the CN build has zero measured cache data. Pick something else if you need third-party quality validation, since none is recorded. Pick something else if you want semantic cache matching, since the family bench shows strict exact-cache keying. The price is low at $0.084 per million input tokens and $0.373 per million output tokens. GLM 5.1 CN and GLM 5.2 CN cost more at $0.120/$0.479 and $0.171/$0.596, so GLM 5 CN is the cheapest of the three. Cheap is not verified. If cache behavior matters, run your own eval against the CN endpoint before trusting it.