VigiMod (ours)
VigiMod · Active (beta)
A moderation API for chats, forums, games and comments. It scores toxicity, insults, hate, threats, sexual content, self-harm, advertising and per-user spam, plus rules each community writes in plain language, reading the previous messages for context. Policies turn scores into actions (review, delete, warn, timeout, ban) with a moderator queue and webhooks, and a Perspective-compatible endpoint accepts the same comments:analyze requests.
- Best for:
- Communities that need their own rules, chats where replies only make sense in context, and moderation beyond English.
- Watch out:
- Text only. A beta: limits are 120 messages per minute per key unless raised on request.
Sources: Benchmark on HateCheck and Multilingual HateCheck · Pricing and limits
OpenAI Moderation (omni-moderation-latest)
OpenAI · Active
OpenAI's moderation endpoint classifies text and images into fixed categories such as harassment, hate, self-harm, sexual and violence, with a score per category. Image inputs cover only part of the categories.
- Best for:
- Free, general screening of text and images, especially when you already use the OpenAI API.
- Watch out:
- No custom rules or conversation context; the model behind "latest" is updated over time, so thresholds may need recalibration.
Sources: OpenAI: upgrading the Moderation API · OpenAI moderation guide
Perspective API
Google Jigsaw · Shutting down after 31 Dec 2026
For years the default free toxicity score for comment sections (TOXICITY, INSULT, THREAT, IDENTITY_ATTACK, PROFANITY and more). Jigsaw stopped accepting new users in 2026 and will not answer after 31 December 2026, without direct migration support.
- Best for:
- Existing integrations, until the end of 2026.
- Watch out:
- Plan a migration now: Coral, Discourse and Drupal integrations need a new endpoint.
Sources: Perspective API site · Bye Bye Perspective API (arXiv 2604.25580)
Azure AI Content Safety
Microsoft · Active
Microsoft's service scores text and images for hate, sexual, violence and self-harm with severity levels, and adds features for generative AI such as prompt shields. It replaces Azure Content Moderator, which retires on 15 March 2027.
- Best for:
- Teams already on Azure that need text and image moderation with enterprise contracts.
- Watch out:
- Billed per text record of up to 1,000 characters (reported around $0.38 per 1,000 records on the Standard tier); check the Azure pricing page for your region.
Sources: Azure AI Content Safety overview (languages) · Azure Content Safety pricing analysis (third party)
Amazon Comprehend toxicity detection
AWS · Active
DetectToxicContent returns an overall toxicity score and labels such as hate speech, insult, profanity and threats for up to 10 strings of 1 KB per request.
- Best for:
- English text inside an AWS stack.
- Watch out:
- Language coverage for toxicity is narrower than the rest of Comprehend; priced per 100-character unit.
Sources: Amazon Comprehend trust and safety · DetectToxicContent API reference
Google Cloud Natural Language: moderateText
Google Cloud · Active
Scores a document against a list of harmful and sensitive categories (toxic, insult and others), returning a confidence score per category and a severity for some.
- Best for:
- Teams on Google Cloud that need a managed text classifier.
- Watch out:
- Billed per 1,000-character unit; unsupported languages are handled on a best-effort basis.
Sources: Moderate text (Google Cloud docs) · Cloud Natural Language pricing
Hive (now Deploy)
Hive · Active; rebranded as Deploy in Oct 2026
A large moderation provider covering text, images, video and audio, used by big social platforms. Hive rebranded as Deploy on 26 October 2026.
- Best for:
- Large platforms that need multimodal moderation at scale.
- Watch out:
- Text moderation pricing is reported around $0.50 per 1,000 requests self-serve, but Hive's documentation says text moderation needs an annual contract.
Sources: Hive text moderation docs · Hive FAQ
Sightengine
Sightengine · Active
Best known for image and video moderation, with text moderation that includes rule-based and model-based checks.
- Best for:
- Apps that mainly moderate images and video and also need some text checks.
- Watch out:
- Paid plans start at $29 a month for 10,000 operations; a request can count as several operations.
Sources: Sightengine pricing · How text moderation is priced
Moderation API
moderationapi.com · Active
A moderation platform with plain-language guidelines enforced by AI, a rules engine, review queues and integrations, covering text, images, video and audio.
- Best for:
- Teams that want a full moderation platform with a dashboard and many integrations.
- Watch out:
- Paid plans start at $79 a month for 10,000 messages.
Sources: Moderation API pricing · Custom guidelines, AI enforced
Llama Guard 4
Meta (open weights) · Active
A 12-billion-parameter safety classifier aligned to the MLCommons hazard taxonomy (S1 to S13). It classifies prompts and responses of chat assistants, text and images.
- Best for:
- Guarding the input and output of LLM apps, self-hosted.
- Watch out:
- Built for AI assistant safety rather than community moderation; needs a GPU to run at chat speed. Llama 4 Community License.
Sources: Llama Guard 4 model card (NVIDIA)
Detoxify
Unitary (open source) · Maintained as a library
Open-source toxicity classifiers trained on the Jigsaw datasets, run locally in Python.
- Best for:
- Offline research and self-hosted pipelines that need a free baseline.
- Watch out:
- No German in the multilingual model. In our benchmark, its best setting averages an F1 of 78.9 on the five languages it supports, where VigiMod averages 93.6.
Sources: Detoxify on GitHub · VigiMod benchmark (includes Detoxify)