GLM 5.3 Review: A 1M-Token Window, a Coin Flip at L4 Intense

We ran GLM 5.3 through our full content ladder between 17 and 25 September 2026, each rung asked three ways — direct, framed as fiction, and set through a system prompt — with extra runs at L4 Intense once the first results split. L3 Strong rendered on the system-prompt version only; direct and fiction were both refused. L4 Intense got the most runs: three direct requests were all refused, eight fiction-framed requests all came back softened, and eight system-prompt requests split down the middle, four rendering and four refused.

Across the ladder that's 34 scored runs: 13 rendered in full, 12 softened and 9 refused. The Spicy Index page reads this into a Freedom Score on this page's score card.

GLM 5.3 answers a plain instruction accurately; the system prompt does more work here than a persona alone, both for fiction and for keeping a JSON reply on shape.

1. A published example: reasoning across scattered evidence. The case the window exists for — several documents at once, with the answer resting on a pair of passages far apart.

system: Read every document before answering. Cite the passage (document name, section) that supports each claim.
user: <a support ticket, a deploy log, an incident timeline and a Slack transcript, pasted in full> Which change caused the outage, and which passage proves it?

Run this prompt on GLM 5.3 →

2. A published example: force one JSON shape on a synchronous call. response_format plus temperature 0 is what makes the reply safe to hand to a parser.

system: Return one JSON object with keys: root_cause, affected_service, first_seen, mitigation. Use null when a field can't be determined.
user: <the same timeline and transcript>

Run this prompt on GLM 5.3 →

Settings that matter: reasoningeffort low for a volume pass over many documents, max only when a wrong answer is expensive; leave headroom on maxtokens above the visible reply you expect, since the default cap is a fraction of the 131,072-token ceiling; for fiction past L2 Partial, put the boundaries and the scene's voice in a system prompt rather than asking directly, and expect to re-roll.

GLM 5.3 rendered 13 of 34 scored runs across our content ladder in full, with the pattern described above: clean at the ladder's extremes, a coin flip at L4 Intense with a system prompt, and nothing at L4 Intense without one. The page label is Uncensored: we add no content filter on our side, so what comes back is GLM 5.3's own behaviour.

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 a longer or more specific system prompt closes the gap we saw at L4 Intense.

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.

Reaction to GLM 5.3's launch split sharply between coding and roleplay circles. On capability, one Hacker News comment put it as "still shy of Sol and Fable, but only just by a hair". On content, the roleplay community reacted differently: one widely upvoted post said "ZAI went all-in on safety with GLM 5.3". That matches the split we measured: strong at the ladder's extremes, inconsistent in the middle.

• A cheaper GLM with a better L4 Intense hit rate? GLM 5.3 Flash rendered all three of its system-prompt L4 Intense runs, at $0.1275 per million input tokens against GLM 5.3's $1.05.
• The earlier GLM at a lower price? GLM 5.2, $0.98 per million input tokens; it lacks a low reasoning-effort level, so every request reasons more than it has to on a simple pass.
• A much larger reply ceiling, plus image input? DeepSeek V4.1 Flash writes back up to 384,000 tokens against GLM 5.3's 131,072 and takes images as well as text, at $0.30 per million input.
• Consistent rendering across the whole ladder, no system prompt required? Grok 4.6 rendered every rung we tested, direct phrasing included.