GPT Image 2.5 is live — OpenAI's newest image model, targeted edits that leave the rest of the frame alone
All Posts
Best Content Moderation API: Text, Image and Video Compared
2026/10/07

Best Content Moderation API: Text, Image and Video Compared

Best content moderation API by use case: OpenAI, Azure, AWS, Google, Hive, Sightengine and Mistral compared on modality, categories, billing and self-hosting.

The best content moderation API depends less on accuracy claims than on four plain facts: which media it reads, which categories it reports, how it bills, and whether you can run it yourself. OpenAI's omni-moderation-latest is free and reads text and images[1][2], but it scores only 6 of its 13 categories on images and does not classify audio[2]. Amazon Rekognition moderates video by the minute[3]. Mistral publishes open weights for a multimodal moderation model[4]. Those differences decide the pick more often than any benchmark.

This comparison covers eleven options, each checked against the vendor's own documentation and pricing pages in October 2026. It sticks to what each API covers and how it bills. For the actual per-item prices and a worked cost example, see Is OpenAI moderation API free?.

TL;DR

  • Text and images on a budget: OpenAI's moderation endpoint, free for OpenAI API users, 13 categories[1][2].
  • Text with PII and jailbreak detection: Mistral Moderation 2, listed as free, 11 text categories[5][6].
  • Video: Amazon Rekognition (stored and streaming video, billed per minute) or Sightengine (video on paid plans, live streams on Pro)[3][7].
  • Fine-grained image labels: Rekognition's three-level taxonomy with 10 top-level categories, or Sightengine's 120+ image classes[8][7].
  • Self-hosting: Mistral's Shieldstral 1.0 (Apache 2.0 weights, public preview) or Azure's disconnected containers on a commitment plan[4][9].
  • Avoid for new builds: Perspective API, which stops serving after December 31, 2026[10].

What separates one content moderation API from another

Most moderation APIs return the same kind of answer: a set of categories with a score or severity each. The differences that matter in production are elsewhere.

Modality. Text-only APIs cannot help with uploaded photos, and image APIs do not read captions. Video is rarer still: of the options below, only Amazon Rekognition, Sightengine and Hive list video moderation, and Hive enables it through a contact form[3][7][11].

Category design. OpenAI returns 13 harm categories as booleans plus 0 to 1 scores[2]. Azure returns four categories (hate, sexual, violence, self-harm) with a severity level from 0 to 7 for text and 0, 2, 4 or 6 for images[12]. Google Natural Language returns 16 attributes, several of them sensitive topics rather than harms, such as Health, Finance, Politics and Legal[13]. Pick the design that matches how your policy is written.

Billing unit. Vendors count characters, text records, requests, images, minutes or "operations", and the unit changes which inputs are cheap. A 50-character comment is one request at Hive but is billed as 300 characters at Amazon Comprehend[11][14]. The pricing comparison explains each unit.

Where it runs. Every hosted API sends your users' content to a third party. If that is not acceptable, the shortlist is the options with downloadable weights or containers.

Best content moderation API comparison table

APITextImageVideoCategoriesBilling unitFree usageSelf-host
OpenAI omni-moderation-latestYesYes (6 of 13 categories)Not listed13 harm categories, score 0 to 1[2]None (free)Free, within rate limits[1]No
Azure AI Content SafetyYesYesNot listed4 categories, severity 0 to 7[12]Text record (1,000 characters), image5,000 text records + 5,000 images per month[9]Disconnected containers, annual commitment[9]
Amazon RekognitionNoYesYes3-level taxonomy, 10 top-level[8]Image, video minute1,000 images + 60 video minutes per month for 12 months[3]No
Amazon Comprehend toxicityEnglish onlyNoNot listed7 labels[15]100 characters, 3-unit minimumNot on free-tier list[14]No
Google Cloud Vision SafeSearchNoYesNot listed5 likelihoods: adult, spoof, medical, violence, racy[16]Image1,000 units per month[17]No
Google Natural Language Text ModerationYesNoNot listed16 attributes, confidence 0 to 1[13]100 characters50,000 units per month[18]No
HiveYesYesVia contact formNot listed on pricing pageRequest; audio per minute$50+ credits after adding a payment method[11]No
SightengineYesYes (120+ classes)Yes; live streams on Pro120+ image classes[7]Operation2,000 operations per month[7]No
Mistral Moderation 2YesNoNot listed11, incl. PII and jailbreaking[6]None (free)Listed as free[5]No
Mistral Shieldstral 1.0YesYesNot listedYour own yes/no policy questions[4]Your hardwareOpen weightsYes, Apache 2.0
reAPI Content ModerationYesYes (image URLs)NoSame 13 as omni-moderation-latest[19]1,000 words of text or 1 imagePaidNo

Audio is a separate need. OpenAI's model does not classify it[2]; Hive prices audio moderation per minute and Sightengine includes it on the Pro plan[11][7].

Picking the best content moderation API by use case

Chat, comments and LLM prompts

For text in a chat app, a comment section or an LLM pipeline, start with a free option. OpenAI's endpoint covers harassment, hate, illicit, self-harm, sexual and violence families with sub-categories[2], and the Responses API can return moderation scores for a model's input and output in the same call[2]. Mistral Moderation 2 adds categories OpenAI does not have: PII, jailbreaking, and requests for tailored health, financial or legal advice[6]. If your concern is prompt injection or users fishing for personal data, that list fits better.

Choose Azure when you want graded severity instead of a probability, or need to define custom categories (in preview)[12][20]. Azure's Content Safety models are trained and tested on eight languages and may work in others at varying quality[20]. Amazon Comprehend's toxicity detection is English only[15].

User-uploaded images

General-purpose models are thinner on images. OpenAI applies only the self-harm, sexual and violence families to an image; harassment, hate and illicit are text only[2]. Azure covers its four categories on images[12].

When your policy names things like swimwear, alcohol, gambling, drugs, rude gestures or hate symbols, use an image specialist. Rekognition has those as top-level categories[8], and Sightengine lists more than 120 image classes[7]. Our image moderation API guide goes deeper on image inputs, thresholds and code.

Video and live streams

Rekognition analyzes stored video from Amazon S3 and streaming video events, billed per minute processed[3]. Sightengine handles video files up to 50 MB on Starter and up to 500 MB plus live streams on Pro[7]. Hive extracts frames at a sampling rate you set, after you enable video through its contact form[11].

Data that cannot leave your infrastructure

Mistral's Shieldstral 1.0 is a 3.8B-parameter model released under Apache 2.0 that classifies text and image inputs against natural-language policy questions, with a 32k-token context. It is in public preview[4]. Azure sells Content Safety as disconnected containers on annual commitment tiers sized in hundreds of millions of records[9], which suits large enterprises rather than small teams.

Already on Perspective API

Jigsaw is sunsetting Perspective. The service stays active until December 31, 2026, new usage and quota requests were accepted only until February 2026, and there is no migration support[10]. Plan the move now. If you want toxicity-style attributes, Google Natural Language Text Moderation returns Toxic, Insult and Profanity among its 16[13].

Moderation next to generation on reAPI

If you already generate images, video or text through reAPI, Content Moderation runs omni-moderation-latest on the same key, credit balance and task contract. It is async: submit to /api/v1/moderations, then poll the task[21]. That differs from OpenAI's synchronous response, so it is not a drop-in for OpenAI SDK integrations. It is billed per moderation unit; the current rate is on the model page[19].

Is OpenAI moderation API good enough?

The question comes straight from developer forums, and the answer is "for many text workloads, yes". It is free, covers 13 categories and accepts batches of strings[1][2]. Its limits are documented, so you can check them against your case:

  • Seven categories ignore images, so an image-heavy platform gets partial coverage[2].
  • It does not classify audio, and OpenAI documents text and image input only[2].
  • OpenAI upgrades the model over time, and thresholds built on category_scores "may need recalibration"[2].
  • It is not designed for CSAM detection; OpenAI says not to send such material to it[2].
  • Throughput is capped by usage tier, starting at 5,000 requests per day on the Free tier[22].

If none of those rules it out, it is the default. If one does, the table above shows which API covers the gap.

FAQ

Is OpenAI moderation API good enough?

For text moderation inside its 13 categories, often yes, and it costs nothing[1]. It falls short when you need audio or video, image checks beyond the self-harm, sexual and violence families, or categories such as PII and jailbreaking[2][6].

Azure content moderation api

Azure's current product is Azure AI Content Safety, with separate Analyze Text and Analyze Image APIs that score hate, sexual, violence and self-harm with severity levels[20][12]. It also offers Prompt Shields, protected material detection and custom categories[20]. The free tier covers 5,000 text records and 5,000 images per month[9].

Content moderation api google

Google splits moderation across two APIs: Cloud Vision SafeSearch for images, with five likelihood categories[16], and Cloud Natural Language Text Moderation for text, with 16 attributes[13]. Both have monthly free allowances[17][18].

Amazon rekognition content moderation api

Rekognition moderates images through DetectModerationLabels and also moderates stored and streaming video[3]. Labels follow a three-level taxonomy with 10 top-level categories, from Explicit and Violence to Gambling and Hate Symbols[8]. For text on AWS, Amazon Comprehend's toxicity detection is the separate, English-only option[15].

Moderation api mistral

Mistral offers mistral-moderation-2603 (Mistral Moderation 2) on /v1/moderations, with a 128k context and a free listed price[5]. It classifies text across 11 categories including PII and jailbreaking[6]. Its open-weight sibling, Shieldstral 1.0, also reads images[4].

Video moderation api

Amazon Rekognition, Sightengine and Hive list video moderation; the OpenAI, Azure Content Safety, Google and Mistral moderation APIs in this comparison document text or image input only[3][7][11]. With a text-and-image API you can sample frames yourself and send them as images, which is how Hive describes its own video processing[11].

Best image moderation api

For broad image policies with labels like alcohol, gambling or hate symbols, Rekognition or Sightengine[8][7]. For sexual, violence and self-harm only, OpenAI's free endpoint covers images too[2]. See the image moderation API guide for details.

Open source moderation api

Mistral's Shieldstral 1.0 is published with Apache 2.0 weights and handles text and images; it is in public preview[4]. If you need Azure's models on your own infrastructure instead, Azure sells Content Safety as disconnected containers on annual commitment tiers[9].

Choosing the best content moderation API for your stack

Start from the media you have to check. Text-only products can begin with OpenAI's free endpoint or Mistral Moderation 2 and move to Azure when they need severity levels or custom categories. Image-heavy products should look at Rekognition or Sightengine for label depth, and anything with video narrows to Rekognition, Sightengine or Hive. If content may not leave your servers, Shieldstral's open weights are the realistic route.

Price usually decides between the survivors, and the moderation API pricing comparison has the numbers. If you want omni-moderation-latest on the same async contract as your other reAPI models, see the Content Moderation model page. Either way, the best content moderation API is the one whose categories match the policy you actually enforce.

References

  1. OpenAI. Is the Moderation endpoint free to use? Retrieved October 2026 from help.openai.com/en/articles/4936833-is-the-moderation-endpoint-free-to-use
  2. OpenAI. Moderation guide. Retrieved October 2026 from developers.openai.com/api/docs/guides/moderation
  3. Amazon Web Services. Amazon Rekognition pricing. Retrieved October 2026 from aws.amazon.com/rekognition/pricing
  4. Mistral AI. Shieldstral 1.0 model card. Retrieved October 2026 from docs.mistral.ai/models/shieldstral-1-0
  5. Mistral AI. Mistral Moderation 2 model card. Retrieved October 2026 from docs.mistral.ai/models/mistral-moderation-26-03
  6. Mistral AI. Moderation and guardrailing. Retrieved October 2026 from docs.mistral.ai/studio/conversations/moderation
  7. Sightengine. Pricing. Retrieved October 2026 from sightengine.com/pricing
  8. Amazon Web Services. Using the image and video moderation APIs, Amazon Rekognition Developer Guide. Retrieved October 2026 from docs.aws.amazon.com/rekognition/latest/dg/moderation-api.html
  9. Microsoft Azure. Content Safety pricing. Retrieved October 2026 from azure.microsoft.com/en-us/pricing/details/content-safety
  10. Jigsaw. Perspective API sunset announcement. Retrieved October 2026 from perspectiveapi.com
  11. Hive. Pricing. Retrieved October 2026 from thehive.ai/pricing
  12. Microsoft Learn. Harm categories in Azure AI Content Safety. Retrieved October 2026 from learn.microsoft.com/en-us/azure/ai-services/content-safety/concepts/harm-categories
  13. Google Cloud. Moderate text, Cloud Natural Language API. Retrieved October 2026 from docs.cloud.google.com/natural-language/docs/moderating-text
  14. Amazon Web Services. Amazon Comprehend pricing. Retrieved October 2026 from aws.amazon.com/comprehend/pricing
  15. Amazon Web Services. Trust and safety, Amazon Comprehend Developer Guide. Retrieved October 2026 from docs.aws.amazon.com/comprehend/latest/dg/trust-safety.html
  16. Google Cloud. Detect explicit content (SafeSearch). Retrieved October 2026 from docs.cloud.google.com/vision/docs/detecting-safe-search
  17. Google Cloud. Cloud Vision API pricing. Retrieved October 2026 from cloud.google.com/vision/pricing
  18. Google Cloud. Cloud Natural Language pricing. Retrieved October 2026 from cloud.google.com/natural-language/pricing
  19. reAPI. Content Moderation model page. reapi.ai/models/content-moderation
  20. Microsoft Learn. What is Azure AI Content Safety? Retrieved October 2026 from learn.microsoft.com/en-us/azure/ai-services/content-safety/overview
  21. reAPI. Content Moderation API docs. reapi.ai/docs/content-moderation
  22. OpenAI. omni-moderation model page. Retrieved October 2026 from developers.openai.com/api/docs/models/omni-moderation-latest
All Posts

More Posts