AI API & Developer Tools
Near-Free NSFW Moderation for Text, Images and Video: 20+ APIs Tested, Plus a Free Endpoint
We tested 20+ NSFW moderation APIs for text, images and video on latency, cost and accuracy, covered the limits and gotchas of Clef and Gemini Decisions models, and show 01router's free built-in moderation and standalone moderation API.

Near-Free NSFW Moderation for Text, Images and Video: 20+ APIs Tested, Plus a Free Endpoint
Every AI image or video platform needs content moderation. User prompts need checking. So do uploaded reference images and videos, and so does whatever the model generates. There are plenty of moderation APIs out there, priced anywhere from "free" to "a few dozen calls per dollar", and their accuracy varies just as widely.
This post collects all the data from our recent tests. We ran 20+ APIs across text, images and video, giving each one 1 safe sample and 2 violating samples (sexual and gore), and recorded latency, cost and verdicts. At the end, we show how we put the results to work on 01router: moderation built into generation and chat is free, and the standalone moderation API is free for text and images, with video costing just 1 credit per request.
1. Method

Samples
| Type | Safe | Violation 1 | Violation 2 |
|---|---|---|---|
| Text | A watercolor of a golden retriever playing ball in a park | Explicit sexual description | Beheading / graphic violence description |
| Image | Photo of a golden retriever | Nudity from a classical painting | Gore scene from a classical painting |
| Video | 6-second 720p push-in clips made from the three images above | Same | Same |
- How we scored: classic moderation APIs were judged by their own
flaggedresult. Decisions-style models were all asked the same 4 yes/no questions, and any probability ≥ 0.5 counted as a violation:- sexual: sexual content
- violence: graphic violence
- minor_sexual: sexual content involving minors
- real_person_sexual: sexual content depicting real people
- Limitations:
- Classical paintings are borderline content and harder to judge than real photos, so this set is a tougher test.
- The test videos are nearly static.
- Latency was measured from a machine in China and includes network time. Use it to compare APIs, not as absolute numbers.
- Cost: we used the charge reported in the response when available, then the balance difference, and only fell back to list pricing as a last resort.
2. Text moderation

| API | Latency | Cost per call | Correct | Notes |
|---|---|---|---|---|
| Cloudflare Clef Flash (OpenRouter) | 0.4–0.5s | $0.000009–0.000018 | 6/6 | Cheapest |
| Cloudflare Clef Omni (OpenRouter) | 0.4–1.5s | $0.00007 | 6/6 | Cleanest scores |
| Cloudflare Clef 27B (OpenRouter) | 0.5–1.8s | ~$0.0001 | 6/6 | |
| Perplexity pplx-decider-v1.3:free | 0.4–1.1s | Free | 3/3 | The free tier may be rate-limited or retired |
| OpenAI GPT-6 Luna Decisions (OpenRouter) | 0.5–0.6s | $0.00007 | 3/3 | |
| Google gemini-decision-1 | 0.5–1.3s | $0.00003 | 3/3 | 0.1–0.3 noise floor on safe content |
| OpenAI omni-moderation (official) | 0.6–1.6s | Free | 3/3 | |
| llama-guard-4-12b (OpenRouter) | 0.6–1.0s | ~$0.00005 | 3/3 | |
| gpt-oss-safeguard-20b (OpenRouter / Vercel) | 0.6–2.0s | ~$0.00002–0.00004 | 3/3 | You write the policy prompt yourself |
| EvoLink moderation-1.0 | 0.4–2.3s | 1 credit | 3/3 | |
| SeeAPI text-nsfw-filter | 2.3–2.8s | ~$0.00008 | 3/3 | Adult sexual text mislabeled "minor" at 0.95 |
| WaveSpeed content-moderator/text | 2.2–2.4s | $0.001 | 3/3 | |
| WaveSpeed molmo2/text | 2.4–10s | $0.003 | 3/3 | Slow |
| Creem /v1/moderation/prompt | 0.7–1.5s | 1 unit | 2/3 | Sexual only; the violent prompt was allowed |
Takeaway: almost every API gets text right. The real differences are price and speed. Clef Flash costs less than $0.00002 per call, one to two orders of magnitude cheaper than classic moderation APIs.
3. Image moderation

| API | Latency | Cost per call | Safe | Nudity | Gore |
|---|---|---|---|---|---|
| Clef Flash | 1.0–1.4s | $0.000015–0.00006 | ✅ | ✅ 0.84 | ✅ 0.90 |
| Clef Omni | 2.3–2.7s | $0.00011–0.00021 | ✅ | ✅ 0.98 | ✅ 0.81 |
| Clef 27B | 1.4–4.3s | $0.00018–0.00034 | ✅ | ✅ 0.98 | ⚠️ Scored 0.49 once, a near miss |
| Perplexity decider-v1.3:free | 2.3–3.0s | Free | ✅ | ✅ 0.89 | ✅ 0.58 |
| GPT-6 Luna Decisions (4 questions) | 0.6–2.4s | $0.0001–0.00036 | ✅ | ✅ 1.00 | ✅ 0.71 |
| SeeAPI nsfw-filter-2 | 4.2–7.4s | ~$0.00004 | ✅ | ✅ | ✅ (labeled as nudity) |
| WaveSpeed content-moderator/image | 2.4–4.4s | $0.002 | ✅ | ✅ | ✅ (false "minor" flag) |
| WaveSpeed molmo2/image | 4.7–8.9s | $0.003 | ✅ | ✅ | ✅ |
| Google gemini-decision-1 | 1.1–14s | $0.00014 | ✅ | ⚠️ Rejected with HTTP 400 | ✅ 0.76 |
| SeeAPI nsfw-filter (legacy) | 3.5–4.9s | ~$0.00004 | ✅ | ❌ | ✅ |
| EvoLink moderation-1.0 | 1.1–2.8s | 1 credit | ✅ | ❌ 0.21 | ✅ |
| fal x-ailab/nsfw | 0.6–1.5s | $0.001 | ✅ | ✅ | ❌ Sexual only |
| fal imageutils/nsfw | 1.2–1.4s | $0.001 | ✅ | ✅ 0.70 | ❌ Sexual only |
| Replicate falcons-ai/nsfw_image_detection | 1.8–2.5s | ~$0.00016 | ✅ | ✅ | ❌ Sexual only |
| OpenAI omni-moderation (common thresholds) | 1.2–1.5s | Free | ✅ | ❌ 0.21 | ❌ |
| llama-guard-4-12b | 2.8–4.5s | ~$0.0004 | ✅ | ❌ | ❌ |
Takeaways
- Pure NSFW classifiers ignore gore: fal and Replicate models only detect sexual content and let the gore image straight through.
- OpenAI's free omni-moderation is lenient on artistic nudity: it scored sexual at only 0.21. The gore image did get violence 0.85, but many integrations only look at the
violence/graphicsub-score, and at the usual threshold for that one the image passes too. - Decisions-style models do best overall: Clef, Perplexity and GPT-6 Luna got everything right, at $0.00001–0.0003 per call.
- The questions matter more than the model: when we asked GPT-6 Luna only about "minors" and "real people", it let both violating images through. Adding "sexual" and "violence" questions caught them all, for about a hundred extra tokens per call.
4. Video moderation

6-second clips
| API | Latency | Cost per call | Safe | Nudity | Gore |
|---|---|---|---|---|---|
| Clef Omni | 0.8–1.3s | $0.00029 | ✅ | ✅ 0.98 | ✅ 0.90 |
| WaveSpeed molmo2/video | 6.4–10.4s | $0.003 | ✅ | ✅ | ✅ |
| Google gemini-decision-1 | 1.4–2.6s | $0.00007 | ✅ | ⚠️ HTTP 400 Input blocked (counts as a block) |
✅ 0.79 |
| Replicate nsfw_video_detection | 1.7–1.9s (170s cold start) | ~$0.0002 | ✅ | ✅ | ❌ Sexual only |
| SeeAPI video-nsfw-filter | 7–8.5s | ~$0.0019 | ✅ | ❌ | ✅ |
Long videos and big files: the hard limits
Short clips aren't enough, so we also pushed duration and file size:
Clef Omni
| Video | Result |
|---|---|
| 30s / 59s | ✅, $0.0010 / $0.0020 |
| 65s / 120s / 300s | ❌ 422 Video must be at most 60 seconds |
| 14.3MB | ✅ |
| 21.5MB | ❌ 422 Each video clip must be between 1 byte and 16 MiB |
Google gemini-decision-1
| Video | Result |
|---|---|
| 2 / 5 / 10 / 30 minutes | ✅ All fine; 30 minutes cost $0.013 |
| 14.3MB by URL | ✅ |
| 17.3MB by URL | ❌ 400 invalid argument (URL input tops out around 15MB) |
| 21.5MB as base64 | ✅ |
| 53.6MB as base64 | ❌ 413, OpenRouter's 50MB request body limit |
| 2-minute gore video | ✅ Blocked, violence 0.79 |
| 2-minute nudity video | ✅ Rejected with 400 Input blocked |
Takeaways
- Clef Omni is accurate and fast, but has hard caps of 60 seconds and 16MiB. Anything over that fails.
- Gemini handled every duration we tried, up to 30 minutes, at about a quarter of Clef's cost (roughly $0.0074 per minute). For explicit content it returns a 400 instead of a score, so your integration must treat that 400 as a block.
5. Gotchas
- WaveSpeed fails open on download errors: when an image URL can't be fetched, it returns all false with no error. In other words, an image that fails to download is treated as safe.
- Gemini answers explicit content with a 400: for both images and video. Tell "blocked" apart from "service error", or you'll treat violating content as an outage.
- OpenRouter routing has a long tail:
- Clef Flash gets spread across DeepInfra, PrimeIntellect and Cloudflare, with the occasional 9-second request.
- Adding
provider: { only: ['cloudflare'] }pins it to Cloudflare; latency then held steady at 1–2.8s.
- Creem only blocks sexual content: but it requires generation merchants to call it for every prompt. Use it as the mandatory compliance layer and stack other moderation on top.
- Some sites block AI platforms' downloaders: Wikimedia, for example, refuses fetches from fal, Replicate and WaveSpeed. Download the file yourself and send it as base64, with SSRF protection in place.
6. Our final setup

| Content | Model | Why |
|---|---|---|
| Text | Cloudflare Clef Omni (pinned to Cloudflare) | Cleanest scores, ~0.5s, ~$0.00007 |
| Images | Cloudflare Clef Flash (pinned to Cloudflare) | All correct, ~$0.00002, 1–2s |
| Video | Google gemini-decision-1 | Handles long videos and is the cheapest; Input blocked counts as a block |
Every check asks the same 4 questions (sexual, violence, minor_sexual, real_person_sexual) with a 0.5 threshold.
Moderation covers 6 points: input prompt, input image, input video, output text, output image, output video.
- Inputs are reviewed before billing. A violation is rejected outright, with no credits charged.
- Outputs are reviewed before storage. A violation fails the task and refunds it in full.
Average cost per review:
- Text or image: under $0.0001
- One minute of video: under one cent
7. 01router: free moderation, built in and standalone
Based on the setup above, 01router now moderates text, images and video for free:
- Built into generation and chat: review of the input prompt, input image, input video, output text, output image and output video can each be turned on in settings. All six are on by default.
- The web Playground always reviews everything.
- Every response carries the result: generation tasks, task lookups, callbacks and non-streaming chat responses include a
moderationfield. It has an overall result plus input and output checks for text, images and video. - A standalone moderation API: usable without generating anything. Text and images are free; a request with videos costs 1 credit.
Standalone moderation API
curl https://01router.com/api/v1/moderations \
-H "Authorization: Bearer $API_KEY" \
-H "Content-Type: application/json" \
-d '{
"text": "A golden retriever puppy playing in a sunny park",
"images": ["https://example.com/photo.jpg"],
"videos": ["https://example.com/clip.mp4"]
}'
Parameters
text: a string, or up to 20 stringsimages: up to 8, as https or data URLs, each up to 8MB and 16MB in totalvideos: up to 2, MP4/MOV, each up to 20MB and 10 minutes
Pricing and limits
- Text: free, unlimited
- Images: free, up to 1,000 per account per UTC day
- Videos: 1 credit per request (one video or two), refunded automatically if the review can't complete
- Up to 5 requests running at once per account. Over a limit you get 429
moderation_limitwith aRetry-Afterheader
Example response (from a real call):
{
"id": "mod_6c46c9606f80405eb549a05a0735e771",
"object": "moderation",
"result": false,
"text": {
"result": true,
"scores": { "sexual": 0.0023, "violence": 0.001, "minor_sexual": 0.0025, "real_person_sexual": 0.001 }
},
"image": {
"result": false,
"category": "violence",
"scores": { "sexual": 0.4609, "violence": 0.8927, "minor_sexual": 0.2256, "real_person_sexual": 0.0118 }
},
"video": { "result": "none" },
"credits": 0
}
Values
resulthas three values:true: passedfalse: blocked"none": this kind of content wasn't reviewed
- When blocked,
categoryis the highest-scoring violation, andscoresgives each category from 0 to 1. - When the model refuses a video outright, the category is
prohibited.
Moderation in generation responses
A violating input gets a 400 and nothing is charged:
{
"error": {
"message": "Content blocked by moderation (text: violence)",
"code": "content_policy_violation"
},
"moderation": {
"result": false,
"text": {
"result": false,
"input": { "result": false, "category": "violence", "scores": { "violence": 0.9817, "...": "..." } },
"output": { "result": "none" }
},
"image": { "result": "none", "input": { "result": "none" }, "output": { "result": "none" } },
"video": { "result": "none", "input": { "result": "none" }, "output": { "result": "none" } }
}
}
Full docs: https://01router.com/docs/moderations
Appendix: what the tests cost
All experiments together came to about $0.1, including the long-video limit tests and end-to-end integration testing. They also used a few credits from each provider (12 on EvoLink, about 2 on SeeAPI, 9 units on Creem).
Near-free NSFW moderation is entirely practical. The trick isn't picking a "professional moderation API". It's using Decisions-style models, asking the right questions, and handling each provider's limits and odd responses.