by Microsoft
Microsoft's moderation API scores text and images for hate, sexual content, violence and self-harm on severity levels, trained on 8 languages.
Pick Azure AI Content Safety if your app runs on Azure and you moderate text and images in major European or Asian languages with graded severity. Skip it as your only tool for video, audio or child-safety detection, which its docs exclude. Its free tier stops when the monthly limit is reached.
Severity levels per harm category instead of a single flag.
Azure customers get it on their existing bill and region, and Microsoft Foundry offers the same service as guardrails for the models it hosts (Microsoft Learn, read 22 Nov 2026), while its lack of video, audio and CSAM detection sends media-heavy platforms elsewhere.
Main trade-off
Text and images only, and no child exploitation detection.
| Plan | Monthly price | What it includes |
|---|---|---|
| Free (F0) | Free | 5,000 text records a month, 5,000 images a month, Usage stops at the limit, no overage, 5 requests a second |
| Commitment, image 250K | $169/mo | 250,000 images a month, Overage $0.675 per 1,000 images |
| Commitment, text 1M | $338/mo | 1 million text records a month, Overage $0.338 per 1,000 text records |
| Standard (S0) | Custom | $0.375 per 1,000 text records (East US), $0.75 per 1,000 images (East US), 1,000 requests per 10 seconds for text and image analysis |
Free tier (F0): 5,000 text records and 5,000 images a month, with usage stopped at the limit. Standard (S0) pay as you go in East US: $0.375 per 1,000 text records and $0.75 per 1,000 images. A text record is up to 1,000 characters, so a 7,500-character input counts as 8. Commitment tiers: 1 million text records for $338 a month, 250,000 images for $169 a month.
Azure AI Content Safety is Microsoft's content moderation API. It rates user text and pictures in four harm categories, giving each a severity level instead of a single flag, and adds blocklists, Prompt Shields and custom categories. Microsoft says its models were trained and tested on eight languages.
Analyze text
Scores text for hate, sexual content, violence and self-harm and returns a severity level per category, so you can set different cut-offs for each.
Analyze image
Scores images in the same four categories with severity levels, using the same thresholds and blocklists as text.
Blocklists
Custom lists of banned words and phrases run alongside the models for terms the categories do not cover.
Prompt Shields and protected material
Checks user prompts for attacks on a language model and checks generated text for known lyrics, articles and recipes.
Custom categories (preview)
Define and train your own content category, such as a platform-specific scam pattern, and scan text for matches.
Content Safety Studio
A browser tool to test sensitivity levels on sample or uploaded text and images, watch block rates and export working code.
| Tool | Price from | Free plan | Runs on | Rating |
|---|---|---|---|---|
| $169/mo | Yes | Web | 3.9 | |
| Free | Yes | Web | 3.8 | |
| $29/mo | Yes | Web | 3.9 | |
| Quote | No | Web | 3.9 |
Proprietary
Pay-as-you-go pricing in East US is $0.375 for every thousand text records and $0.75 for every thousand images. Committed monthly bundles bring that down, at $338 for a million text records or $169 for 250,000 images. Each text record covers up to a thousand characters.
Yes, the F0 tier is free each month up to its fixed allowance of text records and images. When you reach it, calls stop rather than running up overage, so it suits testing and small apps.
Every category the service checks comes back with a graded level rather than a yes or no. You choose the level at which to block in each one, for example blocking only severe violence in a game chat while letting mild cartoon violence through.
Microsoft says the models were trained and tested on Chinese, English, French, German, Spanish, Italian, Japanese and Portuguese. Other languages can work with varying quality, and protected material and groundedness detection are English only.
No. Microsoft's overview states the service cannot be used to detect illegal child exploitation images. Platforms need a dedicated child-safety tool for that.
Azure AI Content Safety is the moderation service in Microsoft's Azure cloud. It is sold as a standalone API and Studio, and the same service sits behind the guardrails Microsoft Foundry applies to the models it hosts. A team that already runs on Azure creates a Content Safety resource in a supported region and calls it with its existing account and billing.
The core is two calls. Analyze text and Analyze image each score content in four harm categories, hate, sexual, violence and self-harm, and return a severity level for each rather than a yes or no, so a platform can allow mild content and block severe content in the same category. Blocklists add your own banned terms. Around that sit features aimed at AI apps: Prompt Shields for injection attacks on a language model, protected material detection for lyrics and articles in generated text, a task adherence check for agents, and custom categories you define and train, currently in preview. The Studio lets a team test sensitivity levels on sample or uploaded content and export the code.
Its docs list eight trained languages, Chinese and Japanese among them alongside the main Western European ones, and warn that quality varies elsewhere; protected material, groundedness and standard custom categories are English only. That suits social, gaming and marketplace teams on Azure who moderate user text and images in those languages, and teams building chat apps who want one vendor for both user content and model guardrails.
Its limits are clear in its own docs. Child exploitation imagery is explicitly out of scope, so a platform needs a specialist such as Safer by Thorn, or the free CSAM image classifier that approved organisations can get alongside the OpenAI Moderation API. It has no video, audio or live-stream API, where Sightengine or Deploy Content Safety go further. Two preview features, rapid custom categories and groundedness detection, are scheduled to retire on 30 April 2027. Hokai lists the category in AI governance tools.
Product facts on this page (pricing, plans, integrations, compliance) are read from Microsoft's own pages and documentation. The verdict is Hokai's reading of them, not the vendor's claim.