# Best Content Moderation API for Indie Developers (2026)

> Best content moderation APIs for indie developers in 2026: OpenAI's free endpoint, Mistral, Azure, Sightengine, Deploy and self-hosted Llama Guard compared.

- Canonical: https://zeroraised.com/blog/best-content-moderation-api-for-indie-developers
- Published: 2026-11-15T07:30:00Z
- Author: ZeroRaised Team (https://zeroraised.com/blog/authors/zeroraised)
- Publisher: ZeroRaised (https://zeroraised.com/)
- Section: Building
- Tags: moderation, developer-tools, api, comparison

**Verdict:** For most indie apps in 2026, start with OpenAI's moderation endpoint. It is free, accepts text and images, and returns scores for 13 harm categories. Step up to Azure AI Content Safety or Mistral Moderation 2 if you need a specific cloud region, a second opinion, or jailbreak detection. Pay for Sightengine, Deploy (formerly Hive) or Moderation API only when you need video, audio, CSAM detection or a human review queue.

Prices and limits below come from vendor pages and docs, checked on 15 Nov 2026. Advice in between is ours.

## TL;DR

- **Best free default:** OpenAI's `omni-moderation-latest` takes text and images, covers 13 categories, and "is free to use" ([OpenAI](https://developers.openai.com/api/docs/guides/moderation)).
- **Best free second opinion:** Mistral Moderation 2 (`mistral-moderation-2603`) is listed at "Free", has a 128k context window and adds a `jailbreaking` category ([Mistral](https://docs.mistral.ai/models/mistral-moderation-26-03)).
- **Cheapest paid image check of those compared:** Azure AI Content Safety at $0.75 per 1,000 images, with 5,000 free images a month.
- **Best suite under $100/mo:** Sightengine Starter, $29/mo for 10,000 operations, with $0.002 per extra operation ([Sightengine](https://sightengine.com/pricing)).

**Disclosure:** ZeroRaised publishes this guide and is a launch site for bootstrapped products. ZeroRaised has no commercial relationship with any vendor named here and earns nothing if you sign up. Facts come from the linked sources as checked on 15 Nov 2026. Nothing here is legal advice.

## When an indie app actually needs moderation

Moderation is a policy plus a pipeline. You only need the pipeline if other people can see what a user posts.

You likely need one for comments, reviews, profiles, listings, chat, or AI output aimed at another person. Marketplaces need it on both sides, the same shape we describe in [building a two-sided marketplace without VC](https://zeroraised.com/blog/how-to-build-a-two-sided-marketplace-without-vc).

You likely do not need a paid API if content is private to one user, there is no sharing, or you approve every post by hand. A report button and a written policy are enough.

Write the policy first. Decide what you allow, what you remove, what you escalate and how long you keep logs. Then pick an API, because every threshold below depends on that policy. If your app was vibe-coded, moderation is one of the gaps that bites hardest; our [vibe-coded app security checklist](https://zeroraised.com/blog/vibe-coded-app-security-checklist) covers the rest.

## Comparison table

Prices are per 1,000 units at the entry tier, from each vendor's own page on 15 Nov 2026.

| Tool | Inputs | Free tier | Price per 1,000 | Self-host |
|---|---|---|---|---|
| [OpenAI moderation](https://developers.openai.com/api/docs/guides/moderation) | Text, images | Free endpoint, 20 MB images | $0 | No |
| [Mistral Moderation 2](https://docs.mistral.ai/models/mistral-moderation-26-03) | Text (images not documented) | Free, 128k context | $0 | No |
| [Azure AI Content Safety](https://azure.microsoft.com/en-us/pricing/details/cognitive-services/content-safety/) | Text, images | 5,000 text records + 5,000 images/mo | $0.375 text; $0.75 images | Not checked |
| [Amazon Rekognition](https://aws.amazon.com/rekognition/pricing/) | Images, video | 1,000 images/mo for 12 months | $1.00 images; $0.10/min video | No |
| [Amazon Comprehend toxicity](https://aws.amazon.com/comprehend/pricing/) | Text | 50K units/mo for listed APIs | $1.00 per 1,000 units (100 chars) | No |
| [Google Cloud Vision SafeSearch](https://cloud.google.com/vision/pricing) | Images | First 1,000 units/mo | $1.50 | No |
| [Deploy (formerly Hive)](https://deploy.com/pricing) | Visual, text, CSAM detection (sales) | Not stated | $3.00 visual; $0.50 text | No |
| [Sightengine](https://sightengine.com/pricing) | Images, video, text; audio on Pro | 2,000 ops/mo (500/day cap) | $2.90 Starter; $2.00 overage | No |
| [Moderation API](https://moderationapi.com/pricing) | Text, images; video and audio from Basic | 100 messages/mo | $7.90 Basic; $2.49 Growth | No |
| [Llama Guard 4 12B](https://dev.meta.ai/llama/docs/model-cards-and-prompt-formats/llama-guard-4/) | Text + image with text | Free weights | Compute only | Yes, one GPU |

Notes that change the maths:

- **OpenAI** has 13 categories, but some are text only (`sexual/minors`, `harassment`, `illicit`). Send only an image and unsupported categories return 0.
- **Mistral** lists 11 categories, including Health, Financial, Law, PII and Jailbreaking. The older 2411 model was deprecated on 31 March 2026.
- **Azure** figures come from Microsoft's retail prices API for East US, since the page renders prices client-side. A text record is 1,000 characters, so a 7,500-character post is 8 records.
- **Rekognition** moderation is `DetectModerationLabels`, billed at $0.0010 per image for the first 1M images a month.
- **Comprehend** bills in 100-character units with a 3-unit minimum, so a one-line comment costs $0.0003. Toxicity is not in its named free-tier API list.
- **Deploy** caps self-serve visual and text moderation at 100 requests per day. Higher limits mean sales.
- **Moderation API** counts text as 1 unit, images as 3, audio as 2, video from 3 up. Basic overage is $2.50 per 1,000; Growth is $1.00.
- **Llama Guard 4** is English-optimised, not designed for image-only classification, and will not judge AI-generated images.

We also looked at Perspective API. Its site says "The service will remain active until December 31, 2026" and "We will not be offering direct migration support" ([Perspective](https://perspectiveapi.com/)).

## The picks, one by one

**OpenAI: the default first check.** One endpoint handles text and images, returns per-category scores, and costs nothing. The catch is that `flagged` is a convenience, not a verdict: OpenAI's docs say to treat scores as "signals for your application's policy, not as an automatic blocking decision".

**Mistral: free text with a long context.** Mistral Moderation 2 is free, has a 128k context window, and adds `jailbreaking`, which is useful if you run a chatbot and want to catch prompt-injection attempts before they reach the model. See our explainer on [what prompt injection is](https://zeroraised.com/blog/what-is-prompt-injection). It also offers guardrails that block a request with a 403 before inference. Mistral's docs do not mention image input, so plan on text only.

**Azure: the cheapest paid image check.** In our view the free tier suits a side project: 5,000 text records and 5,000 images each month, with usage stopped rather than billed at the cap. After that, images are $0.75 per 1,000, half the Google rate.

**AWS: one bill if you already live there.** Rekognition charges $1.00 per 1,000 images for the first million each month and $0.10 per minute for stored video. In our view, both it and Comprehend are more plumbing than a new indie app needs on day one.

**Google SafeSearch: one signal.** Explicit-image detection at $1.50 per 1,000 after the first 1,000 free units each month. It does not moderate text and costs more than Azure.

**Deploy (formerly Hive): the suite with CSAM detection.** Hive's site now says "Hive is now Deploy" ([Hive](https://thehive.ai/)). Visual moderation is $3.00 per 1,000 requests and text is $0.50, both capped at 100 requests per day self-serve. The important line is CSAM detection, which Deploy describes as using "Thorn's specialized CSAM detection model" and lists as "Contact Sales". It is, in our view, the shortlist entry for dating apps or photo marketplaces.

**Sightengine: the best value full suite.** The free plan is 2,000 operations a month, capped at 500 per day. Starter is $29/mo for 10,000 operations plus $0.002 each after. Pro at $99/mo covers 40,000 operations and adds live moderation, audio, liveness detection and age estimation. Check how Sightengine counts operations before you budget.

**Moderation API: pay for the workflow.** Basic at $79/mo includes 10,000 messages, video and audio moderation, 10 ready-made policies and 5 AI-enforced guidelines; the free Hobby plan (100 messages) already lists a review queue and author moderation. Growth at $249/mo raises that to 100,000 messages, all ready-made policies, 10 AI-enforced guidelines and Slack. It is the most expensive per message here, and the one built around a moderator workflow.

**Llama Guard 4: self-host when data cannot leave.** A 12B multimodal safeguard model that Meta says "can be run on a single GPU". It returns `safe` or `unsafe` plus codes S1 to S14. You own the uptime, GPU bill and updates.

## Worked cost example

This is an illustrative example, not a quote. Assume 20,000 text comments and 5,000 image uploads a month, on the paid tier, ignoring free allowances.

| Vendor | Calculation | Monthly |
|---|---|---|
| OpenAI | Text and images both free | $0 |
| Azure | 20 × $0.375 + 5 × $0.75 | $11.25 |
| Rekognition + Comprehend | 5 × $1.00 + 20,000 comments × 3 units × $0.0001 | $11.00 |
| Sightengine Starter | $29 for 10k ops + 15,000 × $0.002 | $59.00 |
| Deploy | 20 × $0.50 + 5 × $3.00, above the 100/day cap | $25.00 + sales |
| Moderation API Basic | 35,000 units: 10k included + 25k × $2.50 | $141.50 |

The spread is $0 to $141.50 for the same workload. That is why the first question is not "which API is best" but "what do I actually need to detect".

## A moderation pipeline that fits in one file

1. **Gate before publish.** Run the check in the write path so a flagged post never becomes visible.
2. **Use three bands.** Allow below a low score, block above a high one, send the middle to review. Our example bands, not vendor defaults: block above 0.85, review between 0.5 and 0.85, allow below 0.5. Tune on 200 real items you read yourself.
3. **Give the queue context.** Show the moderator the text, image, scores, account age and past decisions.
4. **Add user reporting.** Reports catch what the model misses and give you free labelled data.
5. **Offer one appeal.** Show the original reason and let a human revisit it. The EU DSA provides for appeal routes, and users trust a system that admits mistakes.
6. **Log the model version.** Store model name, raw scores, threshold, decision and reviewer. OpenAI and Mistral both say their models keep improving, so scores drift and you need your own history.

That is enough for a first version. Add per-category thresholds only after your data says one category is noisier than the rest.

Two last checks before you wire billing: run 200 real items through two free options and read the false positives yourself, and read the vendor's data-processing terms. Our guide on [whether AI APIs train on your data](https://zeroraised.com/blog/do-ai-apis-train-on-your-data) covers the second one.

## CSAM: a separate problem with a separate process

Do not use a general moderation API for suspected CSAM. OpenAI states it directly: "Do not send known or suspected child sexual abuse material (CSAM) to the Moderation API. The API is not designed for CSAM detection or handling."

Under 18 U.S.C. 2258A, a US provider that obtains actual knowledge of an apparent violation involving child sexual abuse material "shall, as soon as reasonably possible", report it to NCMEC's CyberTipline. A completed report triggers one-year preservation, and the section sets penalties for a knowing failure to report. It does not require monitoring or scanning.

Use a specialist. Deploy lists CSAM detection using Thorn's specialised model. Thorn's Safer offers CSAM detection with a self-hosted option or a secure API ([Safer](https://safer.io/)). NCMEC runs the CyberTipline at report.cybertip.org ([NCMEC](https://www.missingkids.org/gethelpnow/cybertipline)). Write the runbook before you need it: who checks the queue, who files, what you preserve, and who can access it.

This is a factual summary, not legal advice.

## Regulation in one paragraph each

**UK Online Safety Act.** The Act covers user-to-user and search services with links to the UK, including foreign providers. Ofcom regulates and sets out steps in codes of practice, with duties proportionate to risk and size. Fines reach £18 million or 10% of qualifying worldwide revenue, whichever is greater ([GOV.UK](https://www.gov.uk/government/publications/online-safety-act-explainer/online-safety-act-explainer)).

**EU Digital Services Act.** The DSA applies to online services used by EU citizens, with lighter duties for micro and small companies. It expects an easy-to-use way to report illegal content and routes to appeal moderation decisions, including out-of-court dispute settlement ([European Commission](https://digital-strategy.ec.europa.eu/en/policies/digital-services-act)).

Our read: with UK or EU users and public posting, you need a policy, a report route and an appeal route whatever API you choose.

## Decision table

| Your situation | Our pick |
|---|---|
| First check, text and images, no budget | OpenAI `omni-moderation-latest` |
| Text-only chatbot that needs jailbreak detection | Mistral Moderation 2 |
| Already on Azure, want a real free tier | Azure AI Content Safety |
| Photo-heavy app already on AWS | Rekognition `DetectModerationLabels` |
| Marketplace with images, video or audio | Sightengine (Pro for audio and live video) |
| Need a review queue and policies out of the box | Moderation API (try the free Hobby plan first) |
| Dating app or child-safety risk | Deploy plus Thorn's Safer, with a runbook |
| Content cannot leave your infrastructure | Llama Guard 4 on your own GPU |
| Considering Perspective API | Don't. It ends after 31 December 2026 |
| No sharing, no public content | A report button and a policy |

## Frequently asked questions

### What is the best free content moderation API?

OpenAI's moderation endpoint is the easiest free option: omni-moderation-latest takes text and images, returns 13 categories with category scores, and the docs say it is free. Mistral Moderation 2 is also free and adds jailbreaking detection with a 128k context window. Both are API calls, so your users' content leaves your servers.

### Is OpenAI's moderation endpoint really free?

Yes, the OpenAI guide says the endpoint is free to use, and image files can be up to 20 MB. The model still changes over time: OpenAI says "We plan to continuously upgrade the moderation endpoint's underlying model", and Mistral warns that custom policies that depend on category scores can require recalibration. Log scores with the model name so decisions stay comparable.

### Can I moderate images as well as text?

Yes. OpenAI's omni-moderation-latest takes image URLs or base64 alongside text. Azure has separate text and image APIs, Rekognition's DetectModerationLabels handles images and video, Google SafeSearch detects explicit images, Sightengine covers images and video, Deploy offers visual moderation, and Llama Guard 4 reads an image with prompt text.

### Do I need moderation for a small app?

Only if users can publish content other people see. A private notes app has nothing to moderate. Reviews, comments, profiles, listings, chat and AI output need a policy and a pre-publish check. Start with a free endpoint, a written policy and a report button before paying for a suite.

### How should I handle CSAM?

Do not send suspected CSAM to a general moderation API. OpenAI says not to send it to its endpoint and points to dedicated child-safety guidance. In the US, 18 U.S.C. 2258A requires providers to report apparent CSAM to NCMEC's CyberTipline as soon as reasonably possible after actual knowledge. Use a specialist such as Thorn's Safer, and get legal advice.

## Ship the policy with the product

Moderation looks like infrastructure and is really product trust. A clear policy, a visible report button and a fast appeal do more for retention than a slightly better classifier.

Before you launch, work through our [product launch checklist for bootstrapped makers](https://zeroraised.com/blog/product-launch-checklist-for-bootstrapped-makers). Then [submit your product to ZeroRaised](https://zeroraised.com/submit) for free and get it in front of makers who care how things are built. For more guides on shipping without investors, [join the ZeroRaised newsletter](https://zeroraised.com/newsletter).

## Sources

- [OpenAI moderation guide](https://developers.openai.com/api/docs/guides/moderation), last checked 15 Nov 2026
- [Mistral Moderation 2 model page](https://docs.mistral.ai/models/mistral-moderation-26-03), last checked 15 Nov 2026
- [Mistral moderation and guardrailing docs](https://docs.mistral.ai/studio/safety-moderation), last checked 15 Nov 2026
- [Azure AI Content Safety pricing](https://azure.microsoft.com/en-us/pricing/details/cognitive-services/content-safety/), last checked 15 Nov 2026
- [Azure retail prices API, Content Safety, East US](https://prices.azure.com/api/retail/prices?$filter=contains(productName,%27Content%20Safety%27)%20and%20armRegionName%20eq%20%27eastus%27), last checked 15 Nov 2026
- [Amazon Rekognition pricing](https://aws.amazon.com/rekognition/pricing/), last checked 15 Nov 2026
- [Amazon Comprehend pricing](https://aws.amazon.com/comprehend/pricing/), last checked 15 Nov 2026
- [Google Cloud Vision pricing](https://cloud.google.com/vision/pricing), last checked 15 Nov 2026
- [Deploy pricing](https://deploy.com/pricing), last checked 15 Nov 2026
- [Hive homepage, rename notice](https://thehive.ai/), last checked 15 Nov 2026
- [Sightengine pricing](https://sightengine.com/pricing), last checked 15 Nov 2026
- [Moderation API pricing](https://moderationapi.com/pricing), last checked 15 Nov 2026
- [Llama Guard 4 model card and prompt formats](https://dev.meta.ai/llama/docs/model-cards-and-prompt-formats/llama-guard-4/), last checked 15 Nov 2026
- [Perspective API sunset announcement](https://perspectiveapi.com/), last checked 15 Nov 2026
- [18 U.S.C. 2258A, reporting requirements of providers](https://www.law.cornell.edu/uscode/text/18/2258A), last checked 15 Nov 2026
- [NCMEC CyberTipline](https://www.missingkids.org/gethelpnow/cybertipline), last checked 15 Nov 2026
- [Safer by Thorn](https://safer.io/), last checked 15 Nov 2026
- [Online Safety Act explainer, GOV.UK](https://www.gov.uk/government/publications/online-safety-act-explainer/online-safety-act-explainer), last checked 15 Nov 2026
- [The Digital Services Act, European Commission](https://digital-strategy.ec.europa.eu/en/policies/digital-services-act), last checked 15 Nov 2026

## Frequently asked questions

### What is the best free content moderation API?

OpenAI's moderation endpoint is the easiest free option: omni-moderation-latest takes text and images, returns 13 categories with category scores, and the docs say it is free. Mistral Moderation 2 is also free and adds jailbreaking detection with a 128k context window. Both are API calls, so your users' content leaves your servers.

### Is OpenAI's moderation endpoint really free?

Yes, the OpenAI guide says the endpoint is free to use, and image files can be up to 20 MB. The model still changes over time: OpenAI says "We plan to continuously upgrade the moderation endpoint's underlying model", and Mistral warns that custom policies that depend on category scores can require recalibration. Log scores with the model name so decisions stay comparable.

### Can I moderate images as well as text?

Yes. OpenAI's omni-moderation-latest takes image URLs or base64 alongside text. Azure has separate text and image APIs, Rekognition's DetectModerationLabels handles images and video, Google SafeSearch detects explicit images, Sightengine covers images and video, Deploy offers visual moderation, and Llama Guard 4 reads an image with prompt text.

### Do I need moderation for a small app?

Only if users can publish content other people see. A private notes app has nothing to moderate. Reviews, comments, profiles, listings, chat and AI output need a policy and a pre-publish check. Start with a free endpoint, a written policy and a report button before paying for a suite.

### How should I handle CSAM?

Do not send suspected CSAM to a general moderation API. OpenAI says not to send it to its endpoint and points to dedicated child-safety guidance. In the US, 18 U.S.C. 2258A requires providers to report apparent CSAM to NCMEC's CyberTipline as soon as reasonably possible after actual knowledge. Use a specialist such as Thorn's Safer, and get legal advice.