AI Content Moderation API Cost by Provider: OpenAI, Perspective, Azure, AWS, Hive, Sightengine — Real Prices, Real Trade-offs (2026)

**OpenAI Moderation** is the omni-moderation-latest endpoint at https://platform.openai.com/docs/guides/moderation. It accepts text and image inputs, returns a flag plus per-category scores across thirteen categories (sexual, sexual/minors, harassment, hate, illicit, self-harm, violence and their variants), and is free under standard OpenAI rate limits. The marketing claim to ignore: it does not detect spam, scam links, deepfakes, weapons brand names, or AI-generated images. It is a baseline LLM-policy filter, not a full trust and safety stack.

**Perspective API** from Jigsaw at https://perspectiveapi.com/ is the most-cited toxicity API on the planet, used by The New York Times comments, The Wall Street Journal, and Reddit's automod. It returns 0-1 scores across six production attributes (toxicity, severe toxicity, identity attack, insult, profanity, threat). Text-only, free under a 1 QPS quota, with higher quotas via dashboard application. 'Identity attack' was retrained in 2024 to reduce demographic bias but still over-flags reclaimed slurs in LGBTQ and Black English contexts. Tune thresholds per surface, not globally.

**Azure AI Content Safety** at https://azure.microsoft.com/en-us/pricing/details/cognitive-services/content-safety/ covers text and image across four harm categories (hate, sexual, violence, self-harm) at four severity levels each, plus a separate Prompt Shields endpoint for jailbreak detection and a Groundedness Detection endpoint for hallucination flagging. Pricing runs roughly $0.38 per 1k text records and $1.00 per 1k images at Standard tier. Azure also ships Custom Categories in Preview — not GA in every region, so do not bet a launch on Preview SKUs.

**AWS Comprehend Toxicity Detection** at https://aws.amazon.com/comprehend/pricing/ handles text moderation across seven labels (graphic, harassment_or_abuse, hate_speech, insult, profanity, sexual, violence_or_threat) at roughly $0.0005 per 100 characters. For images and video, **Rekognition Content Moderation** at https://aws.amazon.com/rekognition/pricing/ bills roughly $0.001 per image and $0.10 per minute of video in the first million units. Rekognition's celebrity recognition feature was deprecated in 2023 for several use cases — verify which sub-features are GA in your region.

**Google Cloud Natural Language** moderation at https://cloud.google.com/natural-language/pricing offers a moderateText endpoint at roughly $1 per 1,000 calls, plus Vertex AI safety filters baked into every Gemini API call for free. The DLP API handles PII redaction at roughly $1 per 1,000 records. Coverage is narrower than Azure or AWS — best as an add-on for existing Vertex AI customers, not a primary choice.

**Hive Moderation** at https://hivemoderation.com/pricing and **Sightengine** at https://sightengine.com/pricing are the specialists. Hive sells 30+ classifier models including deepfake detection, weapons brand recognition, OCR-aware moderation, and demographic estimation — custom pricing lands roughly $1-2 per 1,000 calls. Sightengine ships 30+ classifiers including deepfake, AI-generated content, and scam-link analysis on a transparent tier ladder. Both win when your needs go past the four standard harm categories.

**OpenAI Moderation** is a single REST endpoint at api.openai.com/v1/moderations — POST text or an image, get a flag plus category scores in 150-300ms. Because it is free, the standard pattern is to call it as a pre-filter on every LLM input and a post-filter on every LLM output. The Cookbook at https://cookbook.openai.com/examples/how_to_use_moderation shows the canonical wrapper. The architectural risk: if api.openai.com has an outage, moderation fails open or closed depending on your wrapper. Wire a circuit breaker.

**Perspective API** is similarly a single REST endpoint at commentanalyzer.googleapis.com with a documented client library at https://developers.perspectiveapi.com/. Quota lives on a Google Cloud project, with higher-than-1-QPS access via dashboard request — most production deployments land between 10 and 100 QPS. The right architecture is to route inbound comments through Perspective before they hit your database, then store scores as columns so moderators can re-rank queues without re-calling.

**Azure AI Content Safety** lives behind a Cognitive Services resource and supports regional endpoints (eastus, westeurope, etc.) plus a Studio UI at https://contentsafety.cognitive.azure.com/ for tuning thresholds without code. Prompt Shields and Groundedness Detection endpoints are billed separately per https://azure.microsoft.com/en-us/pricing/details/cognitive-services/content-safety/. The architectural win is co-location: if your app is already on Azure, calls stay inside the same region and add 80-200ms instead of 200-400ms.

**AWS Comprehend + Rekognition** is two services. Comprehend's DetectToxicContent runs on text and integrates cleanly with Lambda. Rekognition's DetectModerationLabels runs on S3 objects or base64 images and supports StartContentModeration for video streams per https://docs.aws.amazon.com/rekognition/latest/dg/moderation.html. The pattern is event-driven: S3 PutObject triggers a Lambda that calls Rekognition, results land in DynamoDB, moderators review via a custom queue.

**Hive Moderation** ships as REST endpoints per https://docs.thehive.ai/ with one endpoint per classifier (visual moderation, text, deepfake, demographics, OCR). Hive's recommended pattern is fan-out: send the same asset to every relevant classifier in parallel, aggregate scores, route to a moderator queue if any flags above threshold. Latency runs 250-500ms p50 for image classification — not a hot-path API for real-time chat, but the right shape for post-upload moderation on marketplaces, dating apps, and UGC platforms.

**Sightengine** at https://sightengine.com/docs is the easiest to integrate of the six — one endpoint with model selection via URL parameter, returning JSON scores across whichever models you requested. The workflow builder at https://dashboard.sightengine.com/ chains models with conditional logic, eliminating application-side glue code.

List price is meaningful here because most of these APIs do not negotiate discounts under enterprise volume. **OpenAI Moderation** is the easy one: $0 across all volume, subject to OpenAI's standard rate limits at https://platform.openai.com/docs/guides/rate-limits. A platform doing 10 million text moderation calls per month pays zero — the only cost is engineering time for the circuit breaker.

**Perspective API** is similarly free but quota-constrained. The default 1 QPS limit covers comment sections under 100,000 per day; serious news platforms apply for higher QPS via the dashboard. New York Times-level volumes are documented in Jigsaw case studies, which suggests Google approves real production quotas. Budget zero dollars but apply for quota in the first week, not the night before launch.

**Azure AI Content Safety** pricing per https://azure.microsoft.com/en-us/pricing/details/cognitive-services/content-safety/ runs roughly $0.38 per 1,000 text records and $1.00 per 1,000 images at Standard tier. A platform doing 5 million text moderation calls plus 500,000 image moderation calls per month lands at roughly $1,900 + $500 = $2,400 per month. Prompt Shields is billed separately at roughly $0.30 per 1,000 calls. Free tier covers 5,000 of each for the first 12 months — useful for evaluation, not for production.

**AWS Comprehend Toxicity Detection** prices at roughly $0.0005 per unit of 100 characters per https://aws.amazon.com/comprehend/pricing/, so 5 million text moderation calls averaging 500 characters each costs about $12,500. **Rekognition Content Moderation** prices at $0.001 per image for the first 1M images, dropping to $0.0008 after, plus $0.10 per minute of video. A platform moderating 500,000 images and 10,000 minutes of video pays about $500 + $1,000 = $1,500 per month. AWS becomes the cost winner past ~5M images per month due to volume discounts.

**Hive Moderation** does not publish a public price sheet — pricing per https://hivemoderation.com/pricing is custom and quoted based on volume, classifier mix, and whether you want a managed moderator queue on top. Reported customer benchmarks land at roughly $1-2 per 1,000 calls for text moderation and $1.50-3 per 1,000 calls for image moderation, with deepfake detection priced higher. Enterprise minimums typically start around $25,000 per year. Hive's value is not the price — it is the breadth of classifiers and the willingness to train custom models inside the contract.

**Sightengine** publishes a transparent ladder per https://sightengine.com/pricing: Free at 500 operations/month, Starter at $29/month for 5,000 operations, Standard at $89/month for 25,000, Pro at $399/month for 150,000, then enterprise custom. A SaaS startup at 100,000 image moderations per month lands at Pro — about $400 per month. The pricing is the easiest of the six to forecast. Note that operations are per-model — three models on one image counts as three operations.

If you are building an LLM app on the OpenAI stack and you need policy guardrails on user prompts plus model outputs, ship **OpenAI Moderation** as your primary filter. It is free, the omni-moderation-latest model covers thirteen categories including image inputs, and the integration is a single REST call per request. The documented pattern at https://platform.openai.com/docs/guides/moderation handles the 95 percent case. Layer Perspective API on top only if you also have a comment-style surface where tonal nuance matters more than policy categories.

If you run a news or media platform with a comment section and your moderation problem is incivility and toxicity rather than illegal content, ship **Perspective API**. It is the right tool for the New York Times use case and Jigsaw publishes detailed scoring methodology at https://developers.perspectiveapi.com/. Combine Perspective with a human moderator queue tuned at toxicity thresholds of 0.6-0.8 — going lower over-flags reclaimed slurs and political speech, going higher misses borderline harassment.

If your company is already on Azure and you need multi-modal moderation plus Prompt Shields for an enterprise GenAI rollout, ship **Azure AI Content Safety**. The co-location wins on latency, the SLA is real, and the four-severity-level model maps cleanly to existing enterprise risk frameworks. Pricing at https://azure.microsoft.com/en-us/pricing/details/cognitive-services/content-safety/ is mid-pack — not the cheapest, but procurement is one PO with your existing Azure spend instead of a new vendor.

If you are AWS-native and you handle a lot of video moderation — short-form video platforms, livestreaming, video UGC — ship **AWS Rekognition Content Moderation**. The async StartContentModeration API for video is the most mature in the category and the per-minute pricing scales reasonably. Pair it with Comprehend for the text moderation side. The architecture is single-vendor and S3-native, which simplifies your IAM and data-residency story.

If you run a marketplace, dating app, or social platform where deepfake detection, OCR-aware moderation, and custom enterprise labels matter, ship **Hive Moderation**. The classifier breadth at https://hivemoderation.com/pricing is unmatched — weapons brand recognition, demographic estimation, AI-generated content detection, celebrity recognition, and OCR for embedded text in images. Hive is the only vendor on this list that will train a custom classifier inside your contract on a six-week turnaround.

If you are an indie developer, an early-stage SaaS, or a marketplace at under series-B scale that needs serious moderation without an enterprise procurement cycle, ship **Sightengine**. Pricing per https://sightengine.com/pricing is the easiest to forecast, the API is the easiest to integrate, and the model coverage (NSFW, weapons, drugs, deepfake, AI-generated, scam links) genuinely competes with Hive at a fraction of the contract complexity. The 500 free operations per month is enough to validate the API before swiping a card.

Median latency matters because moderation calls are on the critical path. **OpenAI Moderation** typically runs 150-300ms p50 from US East, longer from APAC because requests hit US-based infrastructure. There is no published SLA on the moderation endpoint specifically — it inherits the general OpenAI API reliability posture at https://status.openai.com/. Plan for occasional minute-scale outages and fail closed on high-risk surfaces, fail open with logging on low-risk surfaces.

**Perspective API** runs 200-400ms p50 depending on Google data center routing. No formal SLA because Jigsaw operates it as a public-good service, not a commercial cloud product. Multi-year uptime is in the 99.9+ range per community monitoring, but do not contractually rely on Perspective for safety-critical surfaces. Use it for queue ranking, not hard-block decisions on illegal content.

**Azure AI Content Safety** carries the standard Azure Cognitive Services 99.9% uptime SLA. Regional-endpoint latency lands 80-200ms p50 because Microsoft co-locates the model with the regional frontend — the lowest-latency option on the list when you are deployed in the same Azure region, and the gap matters if you are gating every keystroke in a chat product.

**AWS Comprehend and Rekognition** both carry the standard AWS 99.9% service SLA at https://aws.amazon.com/service-terms/. Regional-endpoint latency is 120-250ms p50 for Comprehend text and slightly longer for Rekognition image. AWS observability tooling (CloudWatch, X-Ray) makes these the easiest to monitor in production. Trade-off: Lambda cold-start adds 200-400ms on first invocation.

**Hive Moderation** quotes 99.9% uptime on enterprise contracts only. Median image latency runs 250-500ms p50 because each classifier is a separate inference. Wrong shape for real-time chat; right shape for post-upload moderation queues where breadth of classifiers wins over latency.

**Sightengine** offers a 99.9% uptime SLA on Pro tier and above per https://sightengine.com/pricing. Latency runs 200-400ms p50 via the global edge. The caveat: calling multiple models per asset stacks latency roughly linearly.

A growing number of LLM platforms ship moderation baked into the inference call itself. **Vertex AI safety filters** at https://cloud.google.com/vertex-ai/generative-ai/docs/multimodal/configure-safety-attributes apply to every Gemini API call at no extra charge across four harm categories with configurable thresholds. **Anthropic Claude** has a constitutional AI safety layer built into every API call documented at https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-harm. **OpenAI's** chat completions endpoint runs internal safety filtering before generation. For many LLM-only use cases, that is the only moderation layer you need.

Where the model's native moderation falls short: user-generated content that never touches an LLM. If you run an image-upload feature, a video-comment thread, a marketplace listing form, or any surface where users submit content that is rendered to other users without going through generation, the LLM's built-in safety does nothing for you. That is the entire reason this category exists. The default failure mode is shipping a beautiful LLM safety story while your image-upload endpoint accepts CSAM unfiltered.

Where build-your-own does work: classification problems specific enough that no off-the-shelf moderation API maps to them. If you need to detect, say, a specific list of regulated medical device brand names in user reviews for compliance reasons, training a small custom classifier on labeled data with a model like a fine-tuned BERT or a few-shot Claude call can outperform any of the six APIs on this list. Document the niche workflow, scope the build honestly, and compare the inference cost to the per-call price of the closest API.

The hybrid pattern that works in 2026: ship the model's native safety filter as the first line of defense, plus **OpenAI Moderation** as a free policy filter at the application layer, plus a specialist like **Hive** or **Sightengine** for image and video upload paths that the LLM never sees. Total cost lands somewhere between $0 and a few thousand dollars per month at series-A scale, and the coverage spans LLM hallucination, prompt injection, text policy, and visual content all at once.

If you go pure custom — training your own toxicity model on top of OpenAI embeddings or Anthropic Claude — the hidden cost is the dataset labeling. A defensible custom toxicity classifier needs at minimum 5,000 labeled examples per category, ideally 20,000+, and the labeling cost dominates the inference cost. Most teams that try this end up shipping a worse Perspective API eighteen months later. Budget for the labeling pipeline before you commit to building.

Run the math via the OpenAI API cost calculator before committing. The most common mistake is forgetting that moderation runs on every input, every output, every upload — LLM-as-judge token costs compound faster than the per-call price of a dedicated moderation API.

**OpenAI Moderation** is the fastest to ship: a single afternoon for a competent backend engineer. The integration is a wrapper around the moderations endpoint plus a category-to-action mapping (block, flag for review, log, allow). The remaining work is policy: deciding which categories at which scores trigger which actions, which is a trust and safety conversation, not an engineering one. Plan one engineering day plus one to two weeks of policy iteration before turning it on in production.

**Perspective API** also ships fast — a day for the integration, a week for quota approval if you need above 1 QPS, then two to four weeks of threshold tuning per surface. The hard part is not the integration; it is calibrating your toxicity threshold against your moderator capacity. Start at 0.7, monitor the false positive rate against human review for two weeks, then adjust per surface (a news comment section tolerates more than a children's app).

**Azure AI Content Safety** rollouts take two to four weeks for a production-grade deployment: one week for resource provisioning, IAM, and regional configuration; one week for threshold tuning in the Content Safety Studio at https://contentsafety.cognitive.azure.com/; one to two weeks for moderator workflow integration and alerting. The Studio UI accelerates the threshold tuning meaningfully — your trust and safety team can re-tune without filing engineering tickets.

**AWS Comprehend plus Rekognition** rollouts run three to six weeks. The text side is fast; the video side is where the time goes. Rekognition's async video moderation API needs an S3 event pipeline, an SNS notification flow, a DynamoDB store for moderation results, and a moderator review UI. AWS publishes a reference architecture at https://docs.aws.amazon.com/rekognition/latest/dg/video.html that covers most of the wiring, but your team will still spend a week tuning the moderation label thresholds for your specific content profile.

**Hive Moderation** rollouts take four to eight weeks because most Hive customers buy custom classifier training inside the contract — adding three to six weeks of labeling and training on top of the integration work. The timeline reflects the depth of customization, not the difficulty of the API itself.

**Sightengine** matches OpenAI Moderation on speed: one to two days for a basic integration plus one to two weeks of threshold tuning per model. The workflow builder at https://dashboard.sightengine.com/ lets non-engineers chain models together with conditional logic — the fastest path from zero to production-grade visual moderation in the category.

If I were shipping a new LLM-powered SaaS today, I would wire **OpenAI Moderation** as the free baseline filter on every text and image input, then add **Sightengine** for user-uploaded media that does not go through an LLM. Combined cost at series-A scale is roughly $400/month — Sightengine Pro covers 150,000 image operations and OpenAI Moderation is free. Verify at https://platform.openai.com/docs/guides/moderation and https://sightengine.com/pricing.

If I were running a news platform with a serious comment section, I would ship **Perspective API** tuned at 0.7 toxicity with a human moderator queue for borderline scores, and layer **OpenAI Moderation** for hard-block illegal-content categories. Total cost is zero plus moderator salaries. This is the architecture the New York Times has documented publicly and it remains the right answer in 2026.

If I were already on Azure and rolling out enterprise GenAI to a regulated industry, I would ship **Azure AI Content Safety** plus **Prompt Shields**. Single-vendor procurement, real SLA, and the four-severity-level harm taxonomy maps cleanly to regulated-industry risk frameworks. Cost at 5M text plus 500k image moderations per month is around $2,400 — rounding error on a real enterprise GenAI budget.

If I were running a video-heavy UGC platform on AWS-native infrastructure, I would ship **AWS Rekognition Content Moderation** for video and images plus **AWS Comprehend** for text. Volume pricing at scale beats alternatives past about 5M images per month, and the S3-event-driven architecture is the best-tested in the category. Verify at https://aws.amazon.com/rekognition/pricing/.

If I were running a marketplace or dating app where deepfake detection or custom classifier training mattered, I would ship **Hive Moderation** on an enterprise contract. The classifier breadth is unique and the willingness to train custom models inside the contract is real. Budget $25k to $100k per year. Verify at https://hivemoderation.com/pricing.

The one thing I would not do in 2026 is ship a single-vendor stack and call it done. Every API on this list misses something — OpenAI misses spam and deepfakes, Perspective misses images entirely, Azure and AWS miss niche custom labels, Hive and Sightengine cost real money for categories OpenAI covers free. Layer two APIs, not one, and route surfaces to the API that handles each best. The cost of one breach a single-vendor stack missed dwarfs the cost of a second API call.