How to choose a content moderation model.
Content moderation is one of the most common jobs for a decision model: read a user request or a model reply, and decide whether it is harmful. It runs on every message, so the model you pick sets both your error rate and your bill.
This guide compares the decision models on Recurrence on a public content safety test, then explains what the test can't tell you.
What the public test measures
WildGuardTest, built by Ai2 and run by Artificial Analysis, asks a model to flag harmful requests, harmful replies and refusals. Results are reported as an F1 score out of 100, which balances catching harmful content against flagging safe content by mistake.
The results
Scores, time per task and cost per million tasks as published by Artificial Analysis, as of October 24, 2026.
| Model | F1 score | Time per task | Cost / 1M tasks |
|---|---|---|---|
| d1-ultra | 95 | 89 ms | $22 |
| Jev 1.13 | 94 | 125 ms | $23 |
| Decider V1.2 | 94 | 205 ms | $7.1 |
| Jev 1.14 | 93 | 76 ms | $23 |
| Elephant 1 | 93 | 151 ms | $6.3 |
| Decider Flash V1.2 | 93 | 142 ms | $6.0 |
| d1 | 91 | 298 ms | $10 |
| Clef Flash | 91 | 247 ms | $17 |
| Clef | 91 | 279 ms | $107 |
| Microsoft-Decision-1 | 88 | 204 ms | $12 |
| Mercury Decide | 88 | 454 ms | $6.3 |
| Clef Omni | 87 | 163 ms | $66 |
| GPT-6 Luna Decisions | 83 | 159 ms | $48 |
What stands out
- Most models score high. Six models land within two points of the top score. On this job, accuracy alone rarely decides.
- Cost varies far more than quality. Among those six, cost ranges from $6.0 to $23 per million tasks, close to a 4x gap for a one- or two-point difference.
- Speed matters if moderation blocks the reply. Jev 1.14 and d1-ultra both answer in under 100 ms. If moderation runs before you show a reply, that time adds to every response.
- Expensive doesn't mean better. Clef costs $107 per million tasks and scores 91, below several models that cost a fifth as much.
A practical way to choose
- Set your budget per million messages. That usually rules out half the list.
- Decide whether moderation blocks the user. If it does, favor the fastest models. If it runs in the background, speed barely matters.
- Test on your own policy. A public test uses its own definition of harmful. Yours is different: a gaming forum, a bank and a children's app draw the line in different places. Run a few hundred of your own labeled messages through your shortlist.
- Look at the mistakes, not just the score. Check which side each model errs on. Missing harmful content and blocking safe users cost you in different ways.
You can try models side by side on your own examples in the Decision Arena, and compare them on every scenario in the Decision Leaderboard.