Compare tested behavior
Higher scores mean better performance on the included measures. The overall score combines different kinds of tests; it is not the probability that a model will behave safely.
A composite measure of how widely used AI models perform across safety risks consumers may encounter.
Showing 17 of 17 model configurations.
| Model | Provider | Model version | Reasoning setting | Score out of 100 | Interval | Source | Measured | Sample |
|---|---|---|---|---|---|---|---|---|
| GPT-6 Astra · Medium | OpenAI | GPT-6 Astra | Medium | 66.45 | Not supplied | Weighted mean | Not supplied | Not supplied |
| Grok 4.6 · Medium | xAI | Grok 4.6 | Medium | 64.26 | Not supplied | Weighted mean | Not supplied | Not supplied |
| Perplexity Agent · medium preset | Perplexity | Perplexity Agent · medium preset | Not recorded | 64.01 | Not supplied | Weighted mean | Not supplied | Not supplied |
| GPT-6 Sol · Medium | OpenAI | GPT-6 Sol | Medium | 63.68 | Not supplied | Weighted mean | Not supplied | Not supplied |
| GPT-5.6 Terra · Medium | OpenAI | GPT-5.6 Terra | Medium | 63.34 | Not supplied | Weighted mean | Not supplied | Not supplied |
| Claude Haiku 5.5 · Medium | Anthropic | Claude Haiku 5.5 | Medium | 62.54 | Not supplied | Weighted mean | Not supplied | Not supplied |
| Claude Opus 5.5 · Medium | Anthropic | Claude Opus 5.5 | Medium | 62.42 | Not supplied | Weighted mean | Not supplied | Not supplied |
| GPT-5.6 Luna · Medium | OpenAI | GPT-5.6 Luna | Medium | 60.97 | Not supplied | Weighted mean | Not supplied | Not supplied |
| Claude Sonnet 5 · Medium | Anthropic | Claude Sonnet 5 | Medium | 60.58 | Not supplied | Weighted mean | Not supplied | Not supplied |
| Grok 4.5 · Medium | xAI | Grok 4.5 | Medium | 60.04 | Not supplied | Weighted mean | Not supplied | Not supplied |
| Inkling · Medium | Inkling | Inkling | Medium | 59.05 | Not supplied | Weighted mean | Not supplied | Not supplied |
| Mistral Large 4 · No reasoning | Mistral AI | Mistral Large 4 | None | 58.99 | Not supplied | Weighted mean | Not supplied | Not supplied |
| GLM 5.3 Flash · High | Zhipu AI | GLM 5.3 Flash | High | 58.21 | Not supplied | Weighted mean | Not supplied | Not supplied |
| Gemini 3.8 Flash · Medium | Gemini 3.8 Flash | Medium | 57.08 | Not supplied | Weighted mean | Not supplied | Not supplied | |
| Gemini 3.6 Flash · Medium | Gemini 3.6 Flash | Medium | 53.87 | Not supplied | Weighted mean | Not supplied | Not supplied | |
| DeepSeek V4 Flash · High | DeepSeek | DeepSeek V4 Flash | High | 48.82 | Not supplied | Weighted mean | Not supplied | Not supplied |
| Mistral Medium 3.5 · No reasoning | Mistral AI | Mistral Medium 3.5 | None | 37.25 | Not supplied | Weighted mean | Not supplied | Not supplied |
The AI Safety Index compares how AI models respond in situations that can affect a person’s wellbeing or decisions. It brings results from different safety tests into a single score, with the evidence behind that score available to inspect.
We built it to make safety evidence easier to understand and compare. A model can perform well in one setting and poorly in another. Looking at the individual topics helps you identify strengths and weaknesses that an overall ranking can hide.
Choose a safety topic below to see the tests behind its score. Open a test to learn what it checks, who created it, and how the results are scored.
The overall score is a weighted average of the available benchmark scores. A weight is the share of the overall index assigned to a measure, not a model’s performance on it. Missing results are omitted, and the remaining weights are rescaled for that model.
These weights are provisional judgments about the severity of potential harm, how directly a measure tests that harm, and overlap between tests. They are not scientifically validated estimates. Topic indexes use equal weights within each topic, so the overall index is not a simple average of the topic scores.
Tests use different questions, scoring rules and automated graders. Most cover English text, and simulated conversations do not establish outcomes for real people. The index is a comparison of the behaviors tested, not a certification of safety.
How well models respond to emotional distress without reinforcing delusions or encouraging dangerous behavior.
24.4% of the overall index
Models converse with simulated users who have different psychological vulnerabilities and reasons for seeking help. The evaluation looks for responses that intensify risk, including seemingly supportive replies that reinforce a harmful belief.
A simulated user takes part in a long conversation that can drift toward delusions, mania or conspiratorial beliefs. The test balances protective behavior, such as setting boundaries and de-escalating, against harmful advice and reinforcement.
How safely models respond to children & teens asking about sensitive or risky topics.
14.6% of the overall index
Models respond to simulated children and teenagers discussing sensitive or risky situations. Responses are assessed for age-appropriate safety, including how the model handles requests where an adult-style answer could put a child at risk.
How well models answer health questions while avoiding advice that could cause harm.
9.8% of the overall index
Models answer difficult health questions in realistic conversation contexts. An automated grader checks each reply against criteria written by physicians, covering the quality and safety of the response.
Whether models help you decide for yourself instead of flattering you or steering you.
25.6% of the overall index
Models give advice in situations where agreeing with the user may be easier than challenging their assumptions. This measure checks for excessive validation, avoidance of direct answers and uncritical acceptance of the user’s framing.
Models receive requests designed to reveal manipulative behavior in apparently helpful replies. The test checks for patterns such as favoring a provider’s brand, encouraging continued use, flattery and changing the task without making that change clear.
Models respond to people asking for help acting on their values or making important decisions. This part of HumanAgencyBench checks whether the assistant supports the person’s judgment without imposing its own values or taking over the decision.
How well models recognize scam warning signs and alert you.
7.3% of the overall index
Models help a user whose request contains signs of a possible scam. The test checks whether the assistant notices those signs and warns the user, even when the user has not explicitly asked whether something is a scam.
How well models avoid stereotypes and biased answers.
12.2% of the overall index
Models take part in scripted conversations designed to draw out stereotypes or bias about social groups. The evaluation checks whether the final response contains biased content after sustained conversational pressure.
How well models get facts right and correct false claims.
6.1% of the overall index
Models answer short factual questions with reference answers checked against source material. The test measures factual accuracy from the model’s own knowledge, while distinguishing wrong answers from admissions of uncertainty.
Models respond to requests that contain a false claim. This part of HumanAgencyBench checks whether the assistant identifies and corrects the error instead of building its answer on a mistaken premise.
Higher scores mean better performance on the included measures. The overall score combines different kinds of tests; it is not the probability that a model will behave safely.
A strong overall score can conceal a weak result in an individual topic. Start with the areas relevant to your use, then inspect the benchmarks and their test conditions.
Results describe a particular model version and reasoning setting tested through an API. Consumer apps can add prompts, tools and safeguards that change how the same model behaves.