Content Moderation API: Detect NSFW & Unsafe Content in 2026

Generative AI pipelines ship fast. Safety layers often don't. If your app lets users submit free-form text or generates images on demand, you need a reliable content moderation API sitting in the hot path — one that catches NSFW material, hate speech, and policy violations before they reach storage, other users, or a compliance audit. This guide covers the concepts, the options, and the practical code to get it wired up.

What is a content moderation API

A content moderation API is an endpoint you call with a piece of content — text, image URL, or base64 payload — and receive back a structured judgment: safe or not, and why. The response typically includes category labels (sexual, violent, self-harm, hate speech, spam) and confidence scores per category, so you can tune your own threshold rather than accepting a hard binary.

In a generative pipeline there are two places to apply it:

• Ingress (user input): Check the prompt before you ever forward it to a model. Blocks prompt-injection attacks and policy-violating requests before they cost you a single API credit.
• Egress (model output): Check the generated image or text before you persist it or return it to the end user. Catches the cases where a compliant-looking prompt still produces unsafe output.

Both gates together give you defense-in-depth. Either gate alone leaves a hole.

Best content moderation API 2026

There are a handful of serious options in 2026:

Option	Modality	Notes
OpenAI Moderation (omni-moderation-latest)	Text + image	Free with an OpenAI key; solid coverage across 11+ categories
AWS Rekognition	Image + video	Strong for visual nudity/violence; no native text
Google Cloud Vision SafeSearch	Image	Five-label scale; fast and cheap at volume
Azure AI Content Safety	Text + image	Fine-grained category scores; enterprise SLA
Open-source (NudeNet, Detoxify)	Depends	Self-hosted; no latency overhead; maintenance burden

For teams already running on a unified AI gateway: the easiest path is to call the OpenAI-compatible moderation endpoint through Hypereal, keep the same auth header and base URL as the rest of your pipeline, and pay a fraction of the official rate. No separate account, no second set of credentials.

Hypereal's API base URL is https://api.hypereal.cloud/v1 — the same endpoint you use for image generation and LLM calls. Pricing for moderation calls is a fraction of official provider rates; check hypereal.cloud for live numbers.

NSFW detection with a content moderation API

NSFW detection is the most common use case — especially for apps that let users upload avatars, generate product images, or feed content into a social feed.

Most moderation APIs return a score per category. A typical response for an image check looks like:

{
  "id": "modr-abc123",
  "results": [
    {
      "flagged": false,
      "categories": {
        "sexual": false,
        "sexual/minors": false,
        "violence": false,
        "hate": false,
        "self-harm": false
      },
      "category_scores": {
        "sexual": 0.04,
        "violence": 0.01,
        "hate": 0.00
      }
    }
  ]
}

A flagged: true on sexual with category_scores.sexual > 0.7 is a reliable soft-block threshold for most consumer apps. You can tune this: stricter for under-18 audiences, more lenient for adult platforms that require age verification.

Common pitfall: using flagged as a hard gate without checking the raw scores. The default flagged threshold is conservative. If you're rejecting content at too high a rate, read the raw scores and set your own threshold.

How to add a content moderation API to your pipeline

Here is a complete example. It calls the Hypereal-proxied moderation endpoint to check a user's text prompt, then only fires the image generation if the prompt is clean.

cURL (quick test):

curl -X POST https://api.hypereal.cloud/v1/moderations \
  -H "Authorization: Bearer $HYPEREAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "omni-moderation-latest",
    "input": "A sunny beach with kids playing volleyball"
  }'

Python (production pattern):

import os
import httpx

HYPEREAL_BASE = "https://api.hypereal.cloud/v1"
HEADERS = {
    "Authorization": f"Bearer {os.environ['HYPEREAL_API_KEY']}",
    "Content-Type": "application/json",
}

def is_safe(text: str, threshold: float = 0.7) -> bool:
    resp = httpx.post(
        f"{HYPEREAL_BASE}/moderations",
        headers=HEADERS,
        json={"model": "omni-moderation-latest", "input": text},
        timeout=5,
    )
    resp.raise_for_status()
    result = resp.json()["results"][0]
    scores = result["category_scores"]
    # Reject if any category score exceeds threshold
    return not any(v >= threshold for v in scores.values())

def generate_image(prompt: str) -> dict:
    if not is_safe(prompt):
        raise ValueError("Prompt flagged by content moderation — request rejected.")
    resp = httpx.post(
        f"{HYPEREAL_BASE}/images/generate",
        headers=HEADERS,
        json={"model": "gpt-image-2", "prompt": prompt, "size": "1024x1024"},
        timeout=60,
    )
    resp.raise_for_status()
    return resp.json()

# Usage
image_data = generate_image("An oil painting of a mountain lake at sunrise")

This pattern adds roughly 150–300 ms of latency per request — fast enough for interactive products and cheap enough to run on every request.

Get set up in three steps:

1. Sign up at hypereal.cloud
2. Dashboard → API Keys → Create Key
3. export HYPEREAL_API_KEY=sk-... and drop the code above into your pipeline

FAQ

Is a content moderation API the same as a classifier? Functionally, yes — it's a classifier tuned for policy categories. The difference is that moderation APIs are pre-trained on policy-relevant labels (NSFW, hate, self-harm) rather than arbitrary classes, and they return calibrated scores rather than raw logits.

Should I moderate prompts, outputs, or both? Both, for any app that stores or surfaces generated content. Prompt moderation is cheaper (text is smaller than images); output moderation catches jailbreaks and unexpected model behavior. Skip either gate only if you have a clear reason.

Can I use Hypereal's moderation endpoint for image inputs? Yes. The omni-moderation-latest model accepts both text and image URLs in the input field. Pass an array with {type: "image_url", image_url: {url: "..."}} items alongside your text.

What threshold should I use? Start at 0.7 for general consumer apps. Move to 0.5 for stricter environments (schools, under-18 apps). For adult platforms where some content is permitted, inspect per-category scores and only block sexual/minors and self-harm unconditionally.

How does Hypereal price moderation calls? Moderation is billed in credits like every other call (100 credits = $1 USD). New accounts receive free trial credits — enough to test the full moderation + generation loop before spending anything. See hypereal.cloud for the current rate card.