LLM safety and moderation — refusals, flags, and evals
Safety is not a single slider, even though our leaderboard has a safety weight for moderation metadata. A model can refuse a lot and still leak PII, or refuse little and still be fine behind your own filter. Treat catalog flags as one layer.
OpenRouter exposes whether a top provider marks a route as moderated. We surface that on model pages. It is not a red-team report. Pair it with your policy tests, especially if you build consumer chat.
Popular frontier models by quality
Live OpenRouter pricing and Arena quality, cached about an hour. Not a static blog table.
| Model | Provider | Blend $ / 1M | Quality | Context |
|---|---|---|---|---|
| Claude Opus 5 | Anthropic | 15.00 | 93.4 | 1,000,000 |
| Claude Fable 5 | Anthropic | 30.00 | 91.1 | 1,000,000 |
| Qwen3.8 27B | Qwen | 1.71 | 89.0 | 1,000,000 |
| Gemini 3.7 Flash | 2.25 | 88.4 | 1,048,576 | |
| DeepSeek V4 Pro 0423 | DeepSeek | 1.12 | 88.2 | 1,024,000 |
Refusals are a product feature
A medical chatbot that refuses diagnosis may be correct. A coding agent that refuses to write an exploit may still write insecure code. Define the policy, then test. Do not assume Claude is “the safe one” without measuring the harms you care about.
Jailbreaks
Public models get jailbroken. Your app needs output filters and tool allowlists regardless of the base model. Switching from GPT to Llama does not remove that work; it may add it.
Leaderboard safety weight
We do not have a universal 0–100 safety eval for every row. The safety dimension is a weak signal from moderation metadata. If you need true safety ranking, bring your own eval — then still use our cost and quality columns.
Open weights and safety
Uncensored fine-tunes exist on purpose. Hosting them is a policy choice. Our open-weight filter is not a safety filter.
What to do this week
Write 15 prompts that must refuse and 15 that must answer. Run two candidate models. Keep the one that matches the spec. Use this site only to ensure the winner is also affordable and fast enough.
FAQ
Which LLM is the safest?
- Safest for whose policy? Measure refusals and leaks on your tests. Catalog moderation is a hint.
Does quality index include safety?
- No. It is mapped Arena Elo. Safety is a separate, limited signal.
Should I add a second moderation API?
- For consumer products, usually yes — independent of which LLM generates the draft.
Where is the moderated flag?
- On each model detail page under Provider.
Related guides
Also try: LLM leaderboard, compare tool, live arena.