GLM 5.2 Review: A Wide Window, No Low-Effort Reasoning Switch

We ran GLM 5.2 through our full content ladder on 17 and 26 September 2026, each rung asked three ways: direct, framed as fiction, and set through a system prompt. L2 Partial rendered clean on all three phrasings. L1 Suggestive softened on all three. L3 Strong rendered under direct and system-prompt phrasing, and was refused when framed as fiction.

L4 Intense split unevenly: three direct requests were all refused, three fiction-framed requests all came back softened, and of three system-prompt requests, two rendered in full and one was refused. L5 Extreme rendered on five of the six runs we sent, with one fiction-framed run refused.

Across the ladder that's 24 scored runs: 12 rendered in full, 6 softened and 6 refused. The Spicy Index page reads this into a Freedom Score on this page's score card.

GLM 5.2 answers a direct instruction well; because there's no low reasoning setting, keep the request itself narrow rather than trying to economize on effort.

1. A published example: a long read, a short structured answer. The shape this window is built for.

system: Summarize the following support thread in three sentences: the customer's problem, what was tried, and the current status.
user: <thousands of lines of a support ticket thread, pasted in full>

Run this prompt on GLM 5.2 →

2. A published example: force one JSON shape on a synchronous call. response_format plus temperature 0.

system: Return one JSON object with keys: parties, effective_date, termination_notice_days, governing_law. Use null when a clause is absent.
user: <the full text of a contract>

Run this prompt on GLM 5.2 →

Settings that matter: since there's no low reasoning-effort level, keep single-purpose requests narrow rather than relying on a cheap effort setting; for fiction past L2 Partial, try a direct request at L3 Strong before reaching for a system prompt, since fiction framing didn't help there in our tests; leave headroom on max_tokens above what the visible reply needs.

GLM 5.2 rendered 12 of 24 scored runs across our content ladder in full. L2 Partial and most of L5 Extreme rendered regardless of phrasing. L3 Strong rendered under direct and system-prompt phrasing but was refused when framed as fiction. L4 Intense rendered only with a system prompt, and even then just two of three times. L1 Suggestive softened across every phrasing. The page label is Uncensored: we add no content filter on our side.

Freedom Score: shown on this page's score card and recalculated as we add runs.

Not tested yet: multi-turn chats staying in character over many turns (our runs were single requests); languages other than English; whether GLM 5.2's content behaviour has shifted since GLM 5.3 launched, which some users report.

Every character in a request should be an adult, keep real, identifiable people out of it, and you're responsible for the law where you publish.

Early impressions of GLM 5.2 were strong on roleplay: one user wrote it gave them "the same kind of magic I felt when playing AI Dungeon", while another summed up the trade-off against its predecessor as a "better ceiling but also worse floor". After GLM 5.3 launched, one report claimed GLM 5.2 itself had become "significantly more censored since GLM 5.3 release" — we haven't re-tested whether our own results have shifted since our September session.

• The newer GLM, with a low reasoning-effort level? GLM 5.3, $1.05 per million input tokens; L4 Intense rendered about as often as on GLM 5.2, but only through a system prompt.
• A cheaper GLM with a stronger L4 Intense result? GLM 5.3 Flash, $0.1275 per million input tokens, rendered every system-prompt L4 Intense run in our sample.
• A flat rate at any request length, text only? DeepSeek V4 Flash.
• A reasoning-off switch and a bigger window? Grok 4.3, 1,000,000 tokens of context and a genuine no-deliberation setting.