Best AI Content Moderation Tools 2026 - 6 Compared

Google's Jigsaw team is retiring the Perspective API on December 31, 2026, and quota requests stopped being processed back in February. That single line in a shutdown notice is forcing a decision on any engineering team that bolted the free toxicity scorer onto a comment section or chat feature a decade ago. Perspective was never a full moderation stack. It scored text for toxicity on a 0 to 1 scale and left everything else, images, video, policy workflows, human review, to someone else. Its exit just means that "someone else" now has to be picked deliberately instead of inherited by default.

I pulled pricing pages, rate limits and accuracy data on six moderation APIs that cover the gap: OpenAI's Moderation endpoint, Azure AI Content Safety, Hive Moderation, Sightengine, Amazon Rekognition and WebPurify. None of them are drop-in replacements for Perspective's exact toxicity score, but each solves a different piece of the same problem.

Why this is suddenly an engineering decision

Perspective's shutdown lands in the middle of an enforcement cycle, not before it. The EU's Digital Services Act has been in force since February 2024, and 2025 through 2026 turned into the period where the fines started landing instead of just the compliance memos. Platforms operating in the EU now have to publicly report the error rates of whatever automated moderation they run, not just confirm that moderation exists. Getting that number wrong, or not having it at all, is part of what a penalty of up to 6 percent of global annual turnover is meant to punish.

That reporting requirement changes what "good enough" means for a moderation API. A free toxicity score that nobody benchmarks against ground truth doesn't produce a defensible error rate. A vendor with a published accuracy methodology does.

The recall problem nobody puts on a pricing page

Every vendor here will show you a low false positive rate, because false positives are the failure mode users complain about loudly. The failure mode regulators and platforms actually lose sleep over is the opposite one.

Commercial moderation APIs including OpenAI's Moderation endpoint, Perspective and Azure AI Content Safety showed low false positive rates but flagged only 10 to 50 percent of genuinely unsafe conversations in the RealHarm evaluation of real-world LLM application failures.

That's from RealHarm, a 2025 arXiv study built from actual production incidents rather than synthetic red-team prompts, and it's a rougher number than any vendor markets. Specialized guardrail models like LlamaGuard and ShieldGemma caught more of the unsafe traffic in the same study, but at the cost of noticeably higher false positive rates. There's no tool on this list that solves both sides of that tradeoff at once. Picking one means picking which failure mode you can tolerate more of.

How the six compare

OpenAI Moderation API: free, but text and image only

OpenAI's Moderation endpoint runs on omni-moderation-latest and doesn't count against your monthly API spend. It screens text for harassment, hate, illicit content, self-harm and violence, and screens images (up to 20MB) for the subset of those categories that make sense visually: sexual content, self-harm and violence. It doesn't touch audio, and it has no video support at all.

For a team already paying OpenAI for generation, this is close to free insurance for the LLM-output side of moderation. It isn't a substitute for a full trust and safety stack if your platform also hosts video uploads or live audio.

Azure AI Content Safety: the one built for LLM apps specifically

Azure's Content Safety service has a free F0 tier (5,000 text and 5,000 image records per month, capped at 5 requests per second) and a pay-as-you-go S0 tier that scales to 1,000 requests per 10 seconds, according to Microsoft's service documentation. Standard text and image analysis covers sexual content, violence, hate and self-harm with severity levels, similar to the OpenAI endpoint.

Where Azure pulls ahead is everything built for generative AI specifically. Prompt Shields scans incoming prompts and uploaded documents for jailbreak and injection attempts before they reach the model, which is the kind of check that belongs next to whatever you're already doing for LLM observability. Groundedness detection flags text that isn't supported by the source material you fed the model, catching hallucinated claims before they ship. A protected material detector checks generated text against known copyrighted content like song lyrics and articles. None of the other five tools on this list bundle jailbreak defense and hallucination detection into the same API as harm-category screening.

Hive Moderation: the only true multimodal API here

Hive is the one vendor covering text, image, video and audio moderation through a single API, according to Hive's own product documentation and third-party comparisons. It doesn't publish a self-serve rate card. Pricing is quoted per volume and media type after a sales conversation, which is standard for platforms processing millions of items a day but a genuine drawback if you want to estimate a monthly bill before signing anything.

Hive's Moderation Dashboard surfaces per-post model confidence scores across categories like "general_suggestive" and "yes_male_shirtless," letting a human reviewer confirm or override the automated call in seconds. Source: thehive.ai

That dashboard is worth calling out specifically. Most of the pure-API vendors on this list hand you a JSON response and leave the review workflow to you. Hive ships one, with batch actions, per-category confidence scores and a queue view built for a moderation team rather than a solo developer testing an endpoint.

Sightengine: the transparent, self-serve middle tier

Sightengine is the rare vendor here that publishes exact numbers without a sales call. The free tier covers 2,000 operations a month capped at 500 per day. The Starter plan runs $29 a month for 10,000 operations with a $0.002 overage rate, and Pro jumps to $99 a month for 40,000 operations plus live-stream, audio and deepfake detection. Detecting AI-produced images and running liveness checks both consume more operations per call than a basic nudity check, so the effective cost per item depends heavily on which detection classes you actually enable.

There's no spend cap on the paid tiers, meaning a traffic spike into an expensive detection class turns into a surprise invoice rather than a throttled request. Worth budgeting for if you enable the pricier checks by default instead of selectively.

Amazon Rekognition: cheapest at scale, narrowest coverage

Rekognition's DetectModerationLabels API is priced at $0.0010 per image for the first million images processed each month, dropping to $0.0008 for the next 4 million and $0.0006 after that. New AWS accounts get 1,000 images a month free for 12 months. That's the cheapest per-unit image rate of anything reviewed here, and it comes with AWS's usual advantage of living inside infrastructure a lot of teams already run.

The catch is coverage. Rekognition handles images and video; it has no text or audio moderation endpoint at all. Any platform with a comment section or chat feature needs a second vendor regardless of how attractive the image pricing looks.

WebPurify: automated triage plus an actual human review team

WebPurify's pitch is different from the rest of this list: its Automated Intelligent Moderation layer exists to cut the volume of content that needs a human, not to replace the human completely. Plans start around $5 a month according to its SourceForge listing, with per-item automated pricing and human-review rates quoted separately based on volume. For platforms where a wrongly banned post or approved slur creates real brand risk, that hybrid layer is the point, not an add-on.

The picks

Best overall for LLM-backed products: Azure AI Content Safety. Prompt Shields and groundedness detection solve problems the other five tools don't even attempt, and the free F0 tier is generous enough to prototype against before committing to S0 pricing.

Best free option: OpenAI's Moderation API, provided text and image coverage is truly all you need. It costs nothing extra on top of an existing OpenAI bill, which none of the paid alternatives can claim.

Best for multimodal platforms: Hive Moderation, purely because it's the only one of the six that doesn't require stitching together a second or third vendor for video and audio.

Best for transparent budgeting: Sightengine. Everyone else on this list either hides pricing behind a sales call or bundles it into a larger cloud bill; Sightengine puts a number on the page.

None of these close the recall gap RealHarm documented. Budget for a second layer, whether that's human review, a specialized guardrail model, or both, if the cost of missing a violation is higher than the cost of over-flagging one.

Sources

• Perspective API sunset announcement - Google Jigsaw
• RealHarm: A Collection of Real-World Language Model Application Failures - arXiv, 2025
• EU Digital Services Act overview - European Commission
• OpenAI Moderation API documentation - OpenAI
• Azure AI Content Safety overview and service limits - Microsoft Learn
• Azure AI Content Safety pricing - Microsoft Azure
• Prompt Shields concepts - Microsoft Learn
• Hive Moderation review - Stork.AI
• Hive Moderation Dashboard announcement - Hive
• Sightengine pricing - Sightengine
• Amazon Rekognition pricing - AWS
• WebPurify product listing - SourceForge