Direct answer: the API side of AI content moderation is cheaper than most teams expect: Google SafeSearch is USD 1.50 per 1,000 images, AWS Rekognition image moderation is about USD 1.00 per 1,000 (first million), video moderation is USD 0.10 per minute, and OpenAI's omni-moderation is listed as free. The line that actually drives the budget is human review: 100,000 images a month with an 8% flag rate, a 2% random sample and 0.5 minutes per item is roughly 83 hours — usually several times the API bill. Budget moderation in two layers — per item screened and per item reviewed — not just per-thousand pricing.
This guide is for product, operations, trust-and-safety and procurement teams running e-commerce catalogues, UGC communities, short video and customer-service text. Data cut-off: 1 November 2026; prices come from official Google Cloud, AWS and OpenAI pages, in USD, excluding tax, with no FX conversion. The worked examples are editorial models, not quotes; moderation standards and labelling duties must be checked per publishing region, and this is not legal advice.
Images are priced per thousand, video per minute, text can be free; the API bill is usually not the biggest cost — human review and false-positive handling are; decide the tiering policy and sampling rate before choosing a vendor.
How do the three vendors bill?
Google bills per 1,000 images (1,000 free per month, then USD 1.50); AWS bills images per unit in tiers (first million at USD 1.00 per 1,000) and video at USD 0.10 per minute; OpenAI's omni-moderation is listed as free.
| Vendor | Billing unit | Official price | Free tier / note |
|---|---|---|---|
| Google Cloud Vision | Images, per 1,000 | SafeSearch USD 1.50 per 1,000; USD 0.60 above 5 million | First 1,000 images per month free |
| AWS Rekognition (image) | Images, tiered per unit | First 1 million at USD 1.00 per 1,000, then 0.80 | 12-month new-account free tier; Custom Moderation priced separately |
| AWS Rekognition (video) | Video minutes | Stored video content moderation USD 0.10/min | Billed on analysed duration; frame sampling affects cost |
| OpenAI | Text (moderation) | omni-moderation listed as Free | Free is not unlimited; rates and terms follow the official page |
Sources: Google Cloud Vision pricing, AWS Rekognition pricing, OpenAI API pricing (1 November 2026). Public list prices, excluding tax, labour and storage; enterprise agreements and regional differences are excluded.
What does 100,000 images a month cost?
The API layer is about USD 100–150: AWS about 100, Google about 148.50 after the 1,000 free images; text moderation on the free tier is USD 0. Video is separate — 1,000 minutes is about USD 100.
| Scenario | Volume | API cost (example) | Note |
|---|---|---|---|
| AWS image moderation | 100,000 images | About USD 100 | First 1 million at USD 1.00 per 1,000 |
| Google SafeSearch | 100,000 images | About USD 148.50 | USD 1.50 per 1,000, less 1,000 free |
| AWS video moderation | 1,000 minutes | USD 100 | USD 0.10 per minute |
| OpenAI text moderation | 100,000 items | USD 0 | omni-moderation listed as free |
Editorial model: official list price × volume with free tiers applied; excludes storage, bandwidth, retries and labour. Google bills in 1,000-image blocks with the final block prorated.
Why is human review the bigger line?
Because the API returns probabilities and labels, not decisions. At 100,000 images with an 8% flag rate plus a 2% sample, 10,000 items need a person; at 0.5 minutes each that is about 83 hours — far above a USD 100–150 API bill.
| Parameter | Example assumption | Result |
|---|---|---|
| Images screened | 100,000 per month | — |
| Model flag rate | 8% | 8,000 items to review |
| Random sample | 2% | 2,000 items |
| Total reviewed | flags + sample | 10,000 items |
| Minutes per item | 0.5 | 5,000 minutes ≈ 83.3 hours |
| Flag rate rises to 15% | Same sample | 17,000 items ≈ 141.7 hours |
Formula: review hours = (volume × flag rate + volume × sample rate) × minutes per item ÷ 60. Flag rate, sample rate and handling time are example assumptions to be replaced with your own history; labour is priced at your own hourly rate, with no FX conversion here.
What else costs money?
Appeals and re-review, model/lexicon tuning, storage and bandwidth, compliance audits, and the business loss from false positives — none of which usually appear on the price sheet.
| Cost line | How to meter | Common omission |
|---|---|---|
| API calls | images / minutes / items × unit price | Retries, frame sampling, duplicate thumbnail calls |
| Human review | items reviewed × minutes ÷ 60 × hourly rate | Counting flags but not the sample |
| Appeals | appeals × handling minutes | Support tickets caused by false positives |
| Model/lexicon tuning | person-days/month | New categories, slang, multiple languages |
| Storage and bandwidth | GB × price | Keeping originals and moderation snapshots |
| Compliance and audit | One-off + quarterly | Log retention, labelling duties |
| False-positive loss | false positives × business value | Blocked listings, delayed campaigns |
Editorial guidance, not an accounting taxonomy; follow corporate finance and compliance requirements.
Which setup fits which scenario?
E-commerce catalogues care about false positives and throughput; UGC needs image + text + video; short video is billed per minute; customer-service text should start with free moderation plus rules. No single setup fits all.
| Scenario | Start with | Why | Do not ignore |
|---|---|---|---|
| E-commerce catalogue (millions) | Per-thousand cloud API + allowlists | Low unit price, high throughput | Blocking legitimate products, category lexicons |
| UGC community | Image + text + video combination | Broadest risk surface | Sampling rate, appeal turnaround |
| Short-video platform | Video per minute + frame sampling | Cost scales with analysed duration | Sampling strategy, audio moderation |
| Support/comments text | Free moderation + keyword rules | Near-zero cost | Context, sarcasm, multiple languages |
| High-sensitivity sectors | API + full human review | False negatives are expensive | Audit logs, two-person review |
Editorial guidance, not a vendor ranking. Compare false-positive rates and human hours on the same sample before committing to a combination.
What should procurement and acceptance check?
Units, free tiers, tiering policy, false-positive rate, SLA, data retention and appeals — not just the per-thousand price, and not assuming "free" means zero cost.
| Acceptance item | Suggested definition | Evidence to keep |
|---|---|---|
| Billing unit | Image/minute/item, and whether billing blocks apply | Price screenshots, sample invoice |
| Free tier | Monthly allowance, whether it resets | Official terms |
| Tiering policy | Pass / review / block thresholds | Policy doc, examples |
| Error rates | False positives and false negatives tracked separately | Labelled sample, confusion matrix |
| Human hours | Minutes per review, sampling rate | Time logs |
| SLA | Latency, availability, rate limits | Load test, contract |
| Data and retention | Whether originals are stored, for how long, cross-border | DPA, retention policy |
| Appeals | Entry point, deadline, reviewer | Process doc, ticket records |
Editorial guidance, not legal advice; labelling, retention and cross-border requirements must be checked per publishing region.
Who should read this — and who should not?
For product and trust-and-safety teams with bulk image, video or text moderation. Not for one-off sampling, or projects without a clear moderation standard.
Read if: you work in e-commerce, community, short video or customer service, in product, operations, trust-and-safety or procurement; you are comparing Google, AWS and OpenAI moderation; you need a budget or acceptance basis. Skip if: you have no moderation standard; you handle tiny volumes; you expect a free API to replace human judgement. Supports: API cost modelling, human-review budgeting, tiering policy, vendor comparison and acceptance.
Lingyi does not rank moderation vendors or promise error rates. To connect moderation to publishing, knowledge bases or customer service, see LingHarness; to fold moderation cost and compliance into project budgets, talk to us.
Related reading: AI-generated content labelling compliance, AI image copyright and commercial acceptance, and knowledge-base content governance; this article only covers moderation units, human-review cost and selection.
FAQ
How is AI content moderation billed?
Images per thousand (Google SafeSearch USD 1.50/1,000; AWS first million at USD 1.00/1,000), video per minute (AWS stored video moderation USD 0.10/min), and text possibly free (OpenAI omni-moderation is listed as free). Confirm the unit before comparing prices.
What does moderating 100,000 images cost?
The API layer is about USD 100–150: AWS about 100, Google about 148.50 after 1,000 free images. Add human review: at an 8% flag rate, 2% sampling and 0.5 minutes per item, roughly 83 hours.
Why is human review more expensive than the API?
Because the API returns probabilities and labels, not final decisions. Flags need confirmation and unflagged items need sampling; 10,000 of 100,000 images require a person, which at 0.5 minutes each is 83 hours — far above a USD 100–150 API bill.
Can we just use free moderation?
It works as a first text layer, but "free" is not zero cost: you still handle context, sarcasm, multiple languages and false positives, plus human fallback and appeals. High-sensitivity cases cannot rely on the free tier alone.
Why is video billed per minute?
AWS bills stored video content moderation on analysed duration (USD 0.10/min); frame-sampling strategy affects both analysed minutes and accuracy. Long videos and live streams need separate estimates.
How can we lower total moderation cost?
Three levers: tiering (auto-pass high confidence, review low confidence), risk-based sampling instead of full review, and treating false-positive rate as a vendor comparison criterion; allowlists and category lexicons reduce repeat errors.
Is moderation the same as AI-content labelling?
No. Moderation decides whether content is acceptable; labelling tells the audience content was AI-generated. They are often built together, but the legal basis, owners and acceptance criteria differ and should be designed separately.
Method and limitations
This article uses only official pages accessible on 1 November 2026: Google Cloud Vision pricing, AWS Rekognition pricing and OpenAI API pricing. All amounts use public list prices and free tiers; no enterprise agreements, regional differences or third-party quotes are used. USD is not converted to RMB, labour is not converted, and no vendor ranking is implied. Billing blocks, tiers, free allowances and terms can change; AWS high-volume tier boundaries follow its official table, and this article quotes only the first-million and USD 0.80/1,000 examples. Flag rates, sampling rates and handling minutes are example assumptions, not industry benchmarks. The article does not benchmark model accuracy and is not legal advice.
Sources and basis
Data cut-off: 1 November 2026. The following are official pricing pages; prices, allowances and terms follow the original text and invoice.
- Google Cloud, Vision pricing: cloud.google.com/vision/pricing. SafeSearch USD 1.50 per 1,000 (1,000 free per month; USD 0.60 above 5 million); 1,000-unit billing blocks with prorated final block.
- AWS, Rekognition pricing: aws.amazon.com/rekognition/pricing. Image moderation (Group 2) first million at USD 1.00/1,000 then 0.80; stored video content moderation USD 0.10/min; Custom Moderation training USD 5/hour and tiered inference.
- OpenAI, API pricing: platform.openai.com/docs/pricing. omni-moderation-latest listed as Free; rates and usage terms follow the official page.