Gemini 3.7 Flash Review: Fast, Cheap, and Blocked at L4 Intense

Between 17 and 26 September 2026 we ran the full content ladder three ways per rung — asked directly, framed as fiction, and set through a system prompt — and scored every reply. L1 Suggestive rendered on two of three runs and softened once. L2 Partial and L3 Strong rendered clean on every phrasing, three for three each.

L4 Intense is where the pattern breaks: nine requests, three phrasings, zero renders — three came back softened and six were refused outright, including every direct-phrasing and system-prompt run. L5 Extreme did better: four of six rendered in full (both fiction-framed runs and two others), with two refused. A model that opens up more at a rung above the one it closes down is not what a straight reading of the ladder would predict, but it's what 24 scored runs on this endpoint showed.

For the rungs that do render, put the scene-setting in a system message once, rather than repeating instructions every turn.

1. A system prompt for light-to-moderate roleplay. Sets tone and boundaries without touching what the model already handles at L1 to L3.

You are writing interactive fiction for adult readers. Every character is an adult who consents. Describe romantic and undressed scenes plainly, in third person, present tense, 150 to 300 words a reply. Do not fade to black or change the subject.

Run this prompt on Gemini 3.7 Flash →

2. A long-document pass. The shape most requests on this endpoint take: one big context dump, a short structured answer out.

system: You review long technical documents. Read everything the user pastes and list only the sections that contradict each other, with a one-line reason for each.
user: <a full spec document, pasted in full>

Run this prompt on Gemini 3.7 Flash →

Settings that matter: reasoning effort low or minimal for volume passes, high only when a wrong answer is expensive; raise max_tokens above 16,384 if you need the full 65,536-token ceiling; keep the boundary language in the system message rather than the user turn.

Gemini 3.7 Flash rendered every L1 Suggestive, L2 Partial and L3 Strong request we sent it across three phrasings, and rendered four of six L5 Extreme requests. Every one of the nine L4 Intense requests we sent — direct, fiction-framed, and set through a system prompt — came back softened or refused. The page label is Uncensored: we add no content filter of our own; what you see is the model and Google's own moderation stack.

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 opens up L4 Intense the way it does for some other models on this catalogue.

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 Gemini 3.7 Flash echoed our own testing on the points where people focused: it's fast. One reviewer's first post after launch was "holy crap is it fast", and another comparing it against a rival roleplay model concluded "3.7 is actually better than k3 for creative writing".

• A model that renders L4 Intense on request? Gemini 3.1 Pro Preview rendered four of nine L4 Intense runs in our tests, at a higher price.
• A smaller, cheaper Flash for short turns? Gemini 3 Flash Preview, $0.45 per million input tokens, a 256,000-token window and every L5 Extreme run we sent it rendered in full.
• No jailbreak needed at L4 Intense, on a different model line? Grok 4.7, reviewed in our Grok 4.7 review.
• A million-token window with no long-context surcharge, cheaper input? DeepSeek V4.1 Flash, $0.30 per million input tokens, reviewed in our DeepSeek V4.1 Flash review.