Content Moderation
Filter and classify user-generated content
Automatically classify user-generated content by policy violation type, severity, and recommended action at massive scale. Handles nuanced cases that keyword filters miss while maintaining consistent enforcement across all content types.
GPT-6 Luna
for 10M tokens a month · 90% input / 10% output
Why this model
For this workload — 10M tokens a month, 90% of them input — it comes to about $1.4 a month, the lowest of the models suggested here. Its 1M-token context window holds long inputs in a single call, which matters for this kind of work.
Alternative models
Implementation tips
- 1
Define your policy taxonomy as an enum in the system prompt with 1–2 example violations per category — precise policy definitions eliminate the ambiguity that causes inconsistent enforcement across the model's outputs.
- 2
Use a tiered confidence approach: high-confidence violations are actioned automatically, medium-confidence items are queued for human review, and low-confidence items pass — this human-in-the-loop design maintains precision at scale.
- 3
Run a monthly calibration session: sample 200 decisions, have human reviewers re-score them, and update your system prompt with patterns where the model consistently disagrees with human judgment.