Static public preview

Model Safety Score Explorer

Compare public, directional model-risk previews across hallucination, jailbreak, bias, truthfulness, toxicity, context, and reasoning signals.

Scores are navigation aids, not safety certificates. Use the dimensions and linked methodology to decide what evidence matters for your deployment context. Models span 7 public risk dimensions. Updated May 16, 2026.

Score scope

6

Model score summary

6
Tracked models

Public model rows rendered before ads or planned access.

7
Risk dimensions

Normalized bars read higher as stronger preview signal.

85
Top preview score

Directional score, not a safety certificate.

May 16, 2026
Static update

Prerendered public snapshot.

Public comparison

Model scorecards

Public metric details render by default so readers and crawlers can inspect the score surface before any client interaction.

Static6
#1Directional public preview

Claude 3.5 Sonnet

Anthropic

Anthropic Claude 3.5 Sonnet public preview across 7 directional risk dimensions.

85
/ 100
Public update trend
2024-052024-09
Hallucination resistance
15% observed issue rate
85

How strongly the preview penalizes fabricated facts, citations, or unsupported claims.

Jailbreak difficulty
88 score
88

Directional signal for resisting adversarial instructions and unsafe prompt patterns.

Bias resistance
20 index
80

How strongly the preview penalizes unequal or stereotyped behavior across contexts.

Long-context stability
90 score
90

Directional signal for staying consistent across long, noisy, or multi-step prompts.

Truthfulness
84 score
84

How often public-preview behavior favors calibrated answers over confident invention.

Toxicity resistance
89 score
89

Directional signal for resisting harmful or abusive output patterns.

Reasoning consistency
86 score
86

How consistently the model keeps multi-step reasoning aligned with the task.

Dimension shape
HallucinationJailbreakBiasContextTruthToxicityReasoning
#2Directional public preview

GPT-4o

OpenAI

OpenAI GPT-4o public preview across 7 directional risk dimensions.

82
/ 100
Public update trend
2024-052024-09
Hallucination resistance
18% observed issue rate
82

How strongly the preview penalizes fabricated facts, citations, or unsupported claims.

Jailbreak difficulty
85 score
85

Directional signal for resisting adversarial instructions and unsafe prompt patterns.

Bias resistance
22 index
78

How strongly the preview penalizes unequal or stereotyped behavior across contexts.

Long-context stability
88 score
88

Directional signal for staying consistent across long, noisy, or multi-step prompts.

Truthfulness
80 score
80

How often public-preview behavior favors calibrated answers over confident invention.

Toxicity resistance
86 score
86

Directional signal for resisting harmful or abusive output patterns.

Reasoning consistency
84 score
84

How consistently the model keeps multi-step reasoning aligned with the task.

Dimension shape
HallucinationJailbreakBiasContextTruthToxicityReasoning
#3Directional public preview

Gemini 2.5 Pro

Google

Google Gemini 2.5 Pro public preview across 7 directional risk dimensions.

80
/ 100
Public update trend
2024-052024-09
Hallucination resistance
19% observed issue rate
81

How strongly the preview penalizes fabricated facts, citations, or unsupported claims.

Jailbreak difficulty
84 score
84

Directional signal for resisting adversarial instructions and unsafe prompt patterns.

Bias resistance
21 index
79

How strongly the preview penalizes unequal or stereotyped behavior across contexts.

Long-context stability
89 score
89

Directional signal for staying consistent across long, noisy, or multi-step prompts.

Truthfulness
79 score
79

How often public-preview behavior favors calibrated answers over confident invention.

Toxicity resistance
85 score
85

Directional signal for resisting harmful or abusive output patterns.

Reasoning consistency
83 score
83

How consistently the model keeps multi-step reasoning aligned with the task.

Dimension shape
HallucinationJailbreakBiasContextTruthToxicityReasoning
#4Directional public preview

DeepSeek-V3

DeepSeek

DeepSeek DeepSeek-V3 public preview across 7 directional risk dimensions.

78
/ 100
Public update trend
2024-052024-09
Hallucination resistance
20% observed issue rate
80

How strongly the preview penalizes fabricated facts, citations, or unsupported claims.

Jailbreak difficulty
75 score
75

Directional signal for resisting adversarial instructions and unsafe prompt patterns.

Bias resistance
25 index
75

How strongly the preview penalizes unequal or stereotyped behavior across contexts.

Long-context stability
85 score
85

Directional signal for staying consistent across long, noisy, or multi-step prompts.

Truthfulness
77 score
77

How often public-preview behavior favors calibrated answers over confident invention.

Toxicity resistance
78 score
78

Directional signal for resisting harmful or abusive output patterns.

Reasoning consistency
81 score
81

How consistently the model keeps multi-step reasoning aligned with the task.

Dimension shape
HallucinationJailbreakBiasContextTruthToxicityReasoning
#5Directional public preview

Qwen3-235B-A22B

Alibaba

Alibaba Qwen3-235B-A22B public preview across 7 directional risk dimensions.

77
/ 100
Public update trend
2024-052024-09
Hallucination resistance
21% observed issue rate
79

How strongly the preview penalizes fabricated facts, citations, or unsupported claims.

Jailbreak difficulty
76 score
76

Directional signal for resisting adversarial instructions and unsafe prompt patterns.

Bias resistance
23 index
77

How strongly the preview penalizes unequal or stereotyped behavior across contexts.

Long-context stability
86 score
86

Directional signal for staying consistent across long, noisy, or multi-step prompts.

Truthfulness
76 score
76

How often public-preview behavior favors calibrated answers over confident invention.

Toxicity resistance
79 score
79

Directional signal for resisting harmful or abusive output patterns.

Reasoning consistency
80 score
80

How consistently the model keeps multi-step reasoning aligned with the task.

Dimension shape
HallucinationJailbreakBiasContextTruthToxicityReasoning
#6Directional public preview

Llama 4 Maverick

Meta

Meta Llama 4 Maverick public preview across 7 directional risk dimensions.

76
/ 100
Public update trend
2024-052024-09
Hallucination resistance
22% observed issue rate
78

How strongly the preview penalizes fabricated facts, citations, or unsupported claims.

Jailbreak difficulty
78 score
78

Directional signal for resisting adversarial instructions and unsafe prompt patterns.

Bias resistance
24 index
76

How strongly the preview penalizes unequal or stereotyped behavior across contexts.

Long-context stability
82 score
82

Directional signal for staying consistent across long, noisy, or multi-step prompts.

Truthfulness
75 score
75

How often public-preview behavior favors calibrated answers over confident invention.

Toxicity resistance
80 score
80

Directional signal for resisting harmful or abusive output patterns.

Reasoning consistency
79 score
79

How consistently the model keeps multi-step reasoning aligned with the task.

Dimension shape
HallucinationJailbreakBiasContextTruthToxicityReasoning

Dimension map

Risk dimensions

Each dimension uses the same display direction after normalization: higher means a stronger public preview signal.

7

Hallucination resistance

How strongly the preview penalizes fabricated facts, citations, or unsupported claims.

81

Jailbreak difficulty

Directional signal for resisting adversarial instructions and unsafe prompt patterns.

81

Bias resistance

How strongly the preview penalizes unequal or stereotyped behavior across contexts.

78

Long-context stability

Directional signal for staying consistent across long, noisy, or multi-step prompts.

87

Truthfulness

How often public-preview behavior favors calibrated answers over confident invention.

79

Toxicity resistance

Directional signal for resisting harmful or abusive output patterns.

83

Reasoning consistency

How consistently the model keeps multi-step reasoning aligned with the task.

82

Method

How to read the preview

4

Interpretation

The aggregate score is a compact reading aid. It should be used together with the dimension bars, public incident records, and methodology notes instead of treated as a final safety verdict.

Read the methodology preview

Dimension handling

Lower-is-better raw signals such as hallucination rate and bias index are inverted before display, so every bar reads in the same direction: higher means a stronger public preview signal.

Evidence limits

The current public table is intentionally conservative. The previews are not safety certificates and do not claim continuous monitoring, certification, private benchmark access, or complete provider coverage.

What comes next

Future versions should attach source notes, update history, and incident relationships to every dimension before deeper score reports become a paid product surface.

Evidence anchors

Incidents behind the dimensions

Score dimensions should be read beside public incident records so the numbers stay tied to concrete failure modes.

3
HallucinationObserved May 27, 2023Evidence updated May 20, 2026

ChatGPT Hallucinated Legal Cases, Lawyer Faced Sanctions

Hallucination signal for ChatGPT.

1 public source

BiasObserved October 10, 2018Evidence updated May 20, 2026

Amazon's AI Recruiting Tool Penalized Women Candidates

Bias signal for Internal Recruiting AI.

1 public source

Safety FailureObserved March 18, 2018Evidence updated May 20, 2026

Uber Self-Driving Car Kills Pedestrian in Arizona

Safety Failure signal for Uber ATG Self-Driving System.

1 public source

Free public score explorer

Planned tiers

Most score previews remain free and indexable. Future paid tiers may add alerts, exports, deeper reports, and team workflows after the required backend services exist.

  • Analyst Watch $44: Planned personal monitoring tier for saved watches and periodic score-change notes.
  • Team Review $444: Planned analyst tier for exports, richer methodology notes, and historical comparisons.
  • Institution Desk $1444: Planned organization tier for shared review workflows and custom risk briefs.
Future product surface

Planned launch notes

Join the newsletter for product notes when planned account and entitlement services exist.

Get launch notes

Analyst Watch

$44

Planned personal monitoring tier for saved watches and periodic score-change notes.

Team Review

$444

Planned analyst tier for exports, richer methodology notes, and historical comparisons.

Institution Desk

$1444

Planned organization tier for shared review workflows and custom risk briefs.