Content Moderation
Run content moderation on Tycho at $0.10 / $1.50 per 1M tokens, with labels returned as JSON schema. Change the base URL, key and model.
What drives the cost
Moderation is the highest-volume, lowest-output work most platforms send to a model. Every comment, message or listing is read against the same policy, and the answer is a label and a short reason. That makes input price the main lever, and the policy itself is most of the input.
The cost of one call is the input tokens multiplied by the input price, plus the output tokens multiplied by the output price, divided by one million. Tycho costs $0.10 per million input tokens and $1.50 per million output tokens. Current rates are on the pricing page.
Two worked examples to replace with your own numbers
The first is checking comments on a community platform before they appear, against a written policy with eight labels.
| Assumption | Value |
|---|---|
| Policy, label definitions and JSON schema, sent on every call | 1,500 tokens |
| The comment, the thread title and the comment it replies to | 200 tokens |
| Input tokens per comment | 1,700 |
| Output tokens per comment, the label, the policy section and the action | 60 |
| Cost per comment on Tycho | $0.00026 |
| Cost per million comments | About $260 before prompt caching |
The same comment costs about $0.0015 on Clarke and $0.0040 on Hal. If 4 percent of comments are borderline and go again to Clarke for a second read, the average is about $0.00032 per comment, or about $322 per million.
The second is reviewing marketplace listings, where the text is longer and the policy covers more categories.
| Assumption | Value |
|---|---|
| Policy, label definitions and JSON schema, sent on every call | 2,500 tokens |
| Listing title, description and seller notes | 700 tokens |
| Input tokens per listing | 3,200 |
| Output tokens per listing, a label per policy area and a reason | 150 |
| Cost per listing on Tycho | $0.00055 |
| Cost per 100,000 listings | About $55 before prompt caching |
In the comment example input is 65 percent of the cost, and the policy is 88 percent of the input. It is identical on every call, so prompt caching applies to it, and Radium's model pages put the saving on repeated prompts at up to 80 percent. Listing photos are not in these numbers. They go to Clarke, and image tokens depend on size, so count them on your own images.
Count the tokens in a sample of your own content before you plan around any of these numbers.
Which model to use
Tycho is the default for content moderation. Radium built it for classification, routing and triage at high volume and positions it as Haiku-class. It reads text, returns labels that follow a JSON schema and costs the least per token of the three models, which matters when every post on a platform goes through it.
Clarke takes the second read. Send it the items Tycho labels as borderline, the items a user reports after they were allowed, and anything with an image, because Tycho reads text only. Clarke also has a context window of 200,000 tokens against Tycho's 128,000, so a long thread or a user's recent history fits in one call when the context decides the label.
Hal is for the cases a moderator would escalate, such as reviewing a group of accounts for coordinated behavior or drafting the summary a person reads before deciding an appeal. Most platforms send it a small share of traffic, if any.
| Step | Model | Model ID |
|---|---|---|
| First pass on every comment, message and listing | Tycho | tycho-1.0 |
| Borderline labels, reported items, images, long threads | Clarke | clarke-1.0 |
| Escalations, coordinated behavior, appeal summaries for a person | Hal | hal-1.0 |
Keep people on appeals and on every label your policy says a person must decide. The model sorts the queue. Your team owns the outcome.
What this workload needs from the API
A moderation pipeline depends on a short list of API capabilities. Confirm each one against your own client and a sample of your own content before production traffic moves.
| Capability | Why moderation needs it | On Radium |
|---|---|---|
| Structured output | Returning a label from a fixed list that your code can act on without parsing free text | JSON mode and JSON schema on all three models, through response_format |
| Image input | Checking profile photos, listing photos and memes | Hal and Clarke. Tycho reads text only |
| Context window | Reading a whole thread or a user's recent history when context decides the label | Clarke 200,000 tokens, Hal 128,000, Tycho 128,000 |
| Prompt caching | The policy is the same on every call and is most of the input | Up to 90 percent saving on repeated prompts |
| Repeatable runs | Giving the same item the same label twice, which appeals depend on | temperature 0 with a fixed seed |
| Concurrency | Absorbing a spike in posts, or re-reviewing a backlog after a policy change | No published rate limits. Send requests in parallel and retry 429 and 503 responses with backoff |
Three things do not carry over unchanged. Radium does not offer a dedicated moderation endpoint with fixed categories, so the policy lives in your prompt and your label list. Radium does not offer a batch endpoint, so re-reviewing old content after a policy change runs as ordinary requests sent in parallel. Radium does not publish latency figures, so if a post waits for the label before it appears, measure time to the full response at the 95th percentile on your own traffic.
Compare all three models on the model overview. The switching guide walks through a four-step evaluation against Anthropic and OpenAI. Labeling content is close to the classification in Document Extraction and Classification, and triaging the reports users send is close to the routing in Customer Support Chat. More workloads are in the use cases index.
Run your policy on 1,000 items your moderators have already decided before the cutover.
How to switch
Radium serves OpenAI-compatible and Anthropic-compatible APIs, so an existing moderation service keeps its SDK, its policy prompt, its label schema and its retry logic. Three values change, which are the base URL, the API key and the model name.
For the OpenAI SDKs, set the base URL to https://api.radium.cloud/v1 and the model to tycho-1.0. For the Anthropic SDKs, set the base URL to https://api.radium.cloud, because the SDK adds the /v1 path itself.
Steps
- Create an account at deploy.radium.cloud and generate a key at deploy.radium.cloud/keys. New accounts get $25 in credit with no card required.
- Store the key as a secret in the service that runs moderation, and keep it out of the repository and out of browser code.
- Change the base URL and model name, then run Radium in shadow mode on a copy of the day's content while your current provider keeps making the decisions.
- Compare the two sets of labels item by item, then move one surface at a time, starting with the one where a wrong label costs least.
Switching back means pointing the base URL at your previous provider. Guides for the OpenAI and Anthropic SDKs, LangChain, LlamaIndex and PydanticAI are in the Radium docs.
How to test it on decisions your moderators already made
Moderation is easy to test, because your team has already made thousands of decisions. Take 1,000 items your moderators labeled, spread across every label in your policy, and run each one through your current model and through Tycho with the same policy and schema.
- Score each label on its own. Overall accuracy hides the label that fails. Most content is allowed, so a model that allows everything scores well on average and misses every serious case. Make sure the rare, severe labels have enough examples to measure.
- Count both kinds of error. A missed violation reaches your users. A wrong removal reaches a user who did nothing wrong and becomes an appeal. Decide which one costs you more before you pick a threshold.
- Check the reasons. Read a sample of the policy sections the model cites. A right label with the wrong reason makes the appeal harder to answer.
- Test what your users actually write. Include every language you serve, slang, misspellings meant to slip past filters and quotes of rule-breaking content posted to report it.
Use temperature 0 and a fixed seed on both sides, and decide the pass mark for each label before you look at the results.
When Radium is the wrong choice
Radium fits when you moderate against your own policy, the content reaches the model as text or images, and moderation is a meaningful share of your AI spend. There are cases where switching is the wrong move.
- A fixed set of categories covers your policy. OpenAI's Moderation API is free and covers common harm categories in text and images. If its categories match your rules, use it. Many teams run it as a first filter and send policy-specific labels to a model.
- You need to detect known illegal imagery. That work runs on hash matching against industry databases and on reporting duties set by law. A language model does not replace either.
- Most of your content is video or audio. Radium reads text and, on Hal and Clarke, images. A platform built on video keeps a moderation service built for video.
- Your volume is small. If the bill is a few hundred dollars a month, the evaluation will cost more than the switch saves.
- Your contracts require a certification Radium does not hold yet. Radium is SOC 2 compliant, and ISO 27001, HIPAA and GDPR work is in progress, so request the SOC 2 report and check status before user content flows through the pipeline.
If none of these apply, run the 1,000 items your moderators have already decided.
Frequently asked questions
Can I use an LLM to moderate content against our own policy?
Yes. Put your policy and label definitions in the system prompt, and ask for the label as JSON with a schema that lists the allowed values. All three Radium models support JSON mode and JSON schema through response_format, on the OpenAI-compatible and the Anthropic-compatible endpoints. Keep people on appeals and on the labels your policy says a person must decide.
Which Radium model is best for content moderation?
Use Tycho (tycho-1.0) for the first pass on every comment, message and listing. Use Clarke (clarke-1.0) for borderline labels, reported items, images and long threads. Use Hal (hal-1.0) for escalations such as coordinated behavior and appeal summaries.
How much does LLM content moderation cost on Radium?
Tycho costs $0.10 per million input tokens and $1.50 per million output tokens. Checking a comment against a policy, at about 1,700 input tokens and 60 output tokens, costs about $0.00026, or about $260 per million comments before prompt caching. The policy is the same on every call, and Radium's model pages put the prompt caching saving on repeated prompts at up to 80 percent.
Can Radium moderate images and video?
Clarke and Hal accept image input, so profile photos and listing photos can go to Clarke. Tycho reads text only. Radium's models read text and images, so a platform built on video or audio keeps a moderation service built for those formats.
Is Radium a replacement for OpenAI's Moderation API?
They do different jobs. OpenAI's Moderation API is free and returns fixed harm categories. Radium runs a model against the policy you write, with your own labels and a reason for each decision. Many teams use a free filter first and send policy-specific labels to a model.
Does Radium train on the content we send for moderation?
No. Radium does not train on customer data, and prompts and completions are used only to service the request. Radium is SOC 2 compliant, and ISO 27001, HIPAA and GDPR work is in progress.