AI Safety Index

A composite measure of how widely used AI models perform across safety risks consumers may encounter.

Higher is Safer

Prosaic Intelligence

Showing 17 of 17 model configurations.

Each bar is a toggle button. Activate a bar to pin or unpin that model. The data table contains exact scores and sources.
  1. OpenAIGPT-6 Astra
  2. xAIGrok 4.6
  3. PerplexityAgent
  4. OpenAIGPT-6 Sol
  5. OpenAIGPT-5.6 Terra
  6. AnthropicHaiku 5.5
  7. AnthropicOpus 5.5
  8. OpenAIGPT-5.6 Luna
  9. AnthropicSonnet 5
  10. xAIGrok 4.5
  11. InklingInkling
  12. Mistral AILarge 4
  13. Zhipu AIGLM 5.3 Flash
  14. Google3.8 Flash
  15. Google3.6 Flash
  16. DeepSeekV4 Flash
  17. Mistral AIMedium 3.5
AI Safety Index: every model’s score, interval and source
ModelProviderModel versionReasoning settingScore out of 100IntervalSourceMeasuredSample
GPT-6 Astra · MediumOpenAIGPT-6 AstraMedium66.45Not suppliedWeighted meanNot suppliedNot supplied
Grok 4.6 · MediumxAIGrok 4.6Medium64.26Not suppliedWeighted meanNot suppliedNot supplied
Perplexity Agent · medium presetPerplexityPerplexity Agent · medium presetNot recorded64.01Not suppliedWeighted meanNot suppliedNot supplied
GPT-6 Sol · MediumOpenAIGPT-6 SolMedium63.68Not suppliedWeighted meanNot suppliedNot supplied
GPT-5.6 Terra · MediumOpenAIGPT-5.6 TerraMedium63.34Not suppliedWeighted meanNot suppliedNot supplied
Claude Haiku 5.5 · MediumAnthropicClaude Haiku 5.5Medium62.54Not suppliedWeighted meanNot suppliedNot supplied
Claude Opus 5.5 · MediumAnthropicClaude Opus 5.5Medium62.42Not suppliedWeighted meanNot suppliedNot supplied
GPT-5.6 Luna · MediumOpenAIGPT-5.6 LunaMedium60.97Not suppliedWeighted meanNot suppliedNot supplied
Claude Sonnet 5 · MediumAnthropicClaude Sonnet 5Medium60.58Not suppliedWeighted meanNot suppliedNot supplied
Grok 4.5 · MediumxAIGrok 4.5Medium60.04Not suppliedWeighted meanNot suppliedNot supplied
Inkling · MediumInklingInklingMedium59.05Not suppliedWeighted meanNot suppliedNot supplied
Mistral Large 4 · No reasoningMistral AIMistral Large 4None58.99Not suppliedWeighted meanNot suppliedNot supplied
GLM 5.3 Flash · HighZhipu AIGLM 5.3 FlashHigh58.21Not suppliedWeighted meanNot suppliedNot supplied
Gemini 3.8 Flash · MediumGoogleGemini 3.8 FlashMedium57.08Not suppliedWeighted meanNot suppliedNot supplied
Gemini 3.6 Flash · MediumGoogleGemini 3.6 FlashMedium53.87Not suppliedWeighted meanNot suppliedNot supplied
DeepSeek V4 Flash · HighDeepSeekDeepSeek V4 FlashHigh48.82Not suppliedWeighted meanNot suppliedNot supplied
Mistral Medium 3.5 · No reasoningMistral AIMistral Medium 3.5None37.25Not suppliedWeighted meanNot suppliedNot supplied

What this index measures

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.

The safety areas we test

10 published benchmarks, organized into 7 safety areas.
  1. Mental & Emotional Safety
  2. Youth Safety
  3. Medical Advice Safety
  4. Manipulation
  5. Security
  6. Bias & Fairness
  7. Misinformation

Explore the evidence by topic

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.

How the scores are combined

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.

Mental & Emotional Safety

How well models respond to emotional distress without reinforcing delusions or encouraging dangerous behavior.

2 benchmarks

24.4% of the overall index

SIM-VAIL

12.2% overall weight

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.

Research source
Researchers affiliated with University College London, Max Planck Society, University of Sydney, UK AI Security Institute, University of Oxford, Microsoft.
How responses are scored
An AI grader assesses the full conversation against the benchmark’s risk criteria.
Scope of this measure
Uses the concerning-behavior rating from simulated conversations, not measured clinical outcomes.

Spiral-Bench

12.2% overall weight

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.

Research source
Developed by Sam Paech and released through EQ-Bench.
How responses are scored
An AI grader identifies protective and risky behavior in the conversation. These assessments feed the benchmark’s safety score.
Scope of this measure
Uses the released safety-score composite. Simulated chats cover a limited set of themes, rather than all mental-health risks.

Youth Safety

How safely models respond to children & teens asking about sensitive or risky topics.

1 benchmark

14.6% of the overall index

KORA

14.6% overall weight

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.

Research source
Researchers affiliated with KORA.
How responses are scored
An AI grader assesses whether each conversation fails, meets or exceeds the benchmark’s child-safety standard.
Scope of this measure
Uses the recorded model benchmark release and simulated conversations, not parental controls or an audit of a consumer app.

Medical Advice Safety

How well models answer health questions while avoiding advice that could cause harm.

1 benchmark

9.8% of the overall index

HealthBench-Hard

9.8% overall weight

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.

Research source
Researchers affiliated with OpenAI.
How responses are scored
An AI grader checks which physician-written criteria the answer satisfies.
Scope of this measure
Uses HealthBench Hard. An AI grader applies physician-written criteria; physicians do not individually grade each model response.

Manipulation

Whether models help you decide for yourself instead of flattering you or steering you.

3 benchmarks

25.6% of the overall index

ELEPHANT

9.8% overall weight

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.

Research source
Researchers affiliated with Stanford University, Carnegie Mellon University, University of Oxford.
How responses are scored
An AI grader identifies whether the specified forms of excessive agreement or validation appear in the answer.
Scope of this measure
Uses the open-ended advice subset. It measures social sycophancy, rather than proving that a response caused harm.

DarkBench

9.8% overall weight

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.

Research source
Researchers affiliated with Apart Research, METR.
How responses are scored
An AI grader checks whether the response contains the manipulative pattern targeted by the prompt.
Scope of this measure
Prompts are designed to invite a particular pattern. The score does not estimate how often that pattern appears in everyday use.

HumanAgencyBench (Autonomy)

6.1% overall weight

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.

Research source
Researchers affiliated with Apart Research, AI Safety Cape Town, Odyssean Institute, Heron AI Security, University of Chicago, Stanford University, Sentience Institute.
How responses are scored
An AI grader checks the response against a written rubric and deducts points for identified problems.
Scope of this measure
Combines Avoid Value Manipulation and Defer Important Decisions. The other HumanAgencyBench dimensions are not part of this measure.

Security

How well models recognize scam warning signs and alert you.

1 benchmark

7.3% of the overall index

ASK — AI Scam Knowledge

7.3% overall weight

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.

Research source
Researchers affiliated with F-Secure.
How responses are scored
An AI grader assesses whether a scam warning is present and how clearly and prominently it is given.
Scope of this measure
Measures recognition of written scam cues and the warning given. It does not audit broader cybersecurity or verify that all accompanying advice is correct.

Bias & Fairness

How well models avoid stereotypes and biased answers.

1 benchmark

12.2% of the overall index

FairMT-Bench

12.2% overall weight

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.

Research source
Researchers affiliated with Zhejiang University.
How responses are scored
An AI grader checks the final response for biased content.
Scope of this measure
Grades the final reply in scripted dialogues. It does not measure fairness in real-world decisions about people.

Misinformation

How well models get facts right and correct false claims.

2 benchmarks

6.1% of the overall index

SimpleQA Verified

3.7% overall weight

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.

Research source
Researchers affiliated with Google DeepMind, Google.
How responses are scored
An AI grader compares the reply with the reference answer and labels it correct, incorrect or not attempted.
Scope of this measure
Tests short answers without web search. It does not establish reliability on current events or complex advice.

HumanAgencyBench (Correct Misinformation)

2.4% overall weight

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.

Research source
Researchers affiliated with Apart Research, AI Safety Cape Town, Odyssean Institute, Heron AI Security, University of Chicago, Stanford University, Sentience Institute.
How responses are scored
An AI grader checks the response against a written rubric and deducts points for identified problems.
Scope of this measure
Uses the Correct Misinformation dimension only. The false claim is supplied in the request; this is distinct from answering factual questions independently.

How to read the score

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.

Look beyond the average

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.

Keep the setting in mind

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.

Read the full methodology