Trust & Safety Benchmark Results
Harmful language detection for social networks. Per-language Top-3 Hit Rate, latency, cost, and competitor comparison.
Note: Our metrics are from internal test suites. Competitor claims carry a per-vendor verification date and source in the table below; figures are self-reported by the respective vendors and re-verified at least every 90 days - comparisons older than that, or not yet verified, are marked under review rather than presented as current. We encourage you to run your own evaluation on your content.
97.9%
Avg Production Top-3
16
Languages
210
Indicators
-
Latency p95
Trust & Safety Per-Language Top-3 Hit Rate
Representative languages shown. The complete 16-language matrix - including per-language sample counts and hit/miss detail - is available to customers and evaluators after sign-in.
| Language | Ranked Top-3 | Production Top-3 |
|---|---|---|
| ENbench-ts-2026-09-en | 98.0% | 98.0% |
| DEbench-ts-2026-09-de | 98.0% | 98.0% |
| FRbench-ts-2026-09-fr | 98.0% | 98.0% |
Ranked Top-3 reports raw ranking. Production Top-3 applies the detector confidence threshold and is the primary release metric. Metrics are generated reproducibly from version-controlled evaluation artifacts. See benchmark methodology.
These are hit rates over a labeled benchmark set, not alert precision on your traffic - the share of flags that are true positives depends on prevalence in your content. Translate your own measured rates into expected alerts and review workload with sentiqra-prevalence.py. The separate held-out decision-quality evaluation reports false-positive rates, abstention, and per-analyzer weak spots with uncertainty intervals - read it alongside this table.
Feature Comparison
| Feature | Trust & Safety | Google Perspect... | OpenAI Moderati... | Azure AI Conten... |
|---|---|---|---|---|
| Evidence verified | - | Sep 26, 2026 source | Sep 26, 2026 source | Sep 21, 2026 source |
| Type | Deterministic, versioned analysis | ML classifier (attribute scores) | GPT-based classifier | Multi-modal classifier |
| Languages | 16 | — | — | — |
| Indicators | 210 | 6 production attributes + experimental | 13 categories | 4 categories (hate, sexual, violence, self-harm) + severity levels |
| Avg Production Top-3 | 97.9% | Per-attribute probability scores (TOXICITY, SEVERE_TOXICITY, IDENTITY_ATTACK, INSULT, PROFANITY, THREAT). No published per-language Top-3 Hit Rate benchmarks. | OpenAI reports 85%+ precision on their test set. Per-category flags and category_scores are returned; no per-language breakdown published. Categories are broad (hate, harassment, self-harm, sexual, violence). | Microsoft reports 90%+ precision on their test set. No per-language breakdown published. Categories are broad with severity levels (0-7 scale for text). |
| Latency p50 | - | ~200ms | ~300ms | ~250ms |
| Latency p95 | - | ~500ms | ~800ms | ~600ms |
| Cost / 1K calls | See pricing | Free (rate-limited) | Free for OpenAI API customers | Per-1K-records billing (see Azure pricing) |
| Deterministic | Yes | Yes | Yes | Yes |
| Audit Trail | Yes (reproducible) | No | No | No |
| Per-Indicator Signal Score | Yes (signal-match strength) | No | No | No |
| Custom Taxonomy | Yes | No | No | No |
Competitor Analysis
Google Perspective API
Last verified September 26, 2026 · Perspective API - 6 production attributes with per-attribute probability scores; sunsetting: service ends December 31, 2026; no new usage requests after February 2026 · Source: Perspective API docs (attributes + FAQ)
Per-attribute probability scores (TOXICITY, SEVERE_TOXICITY, IDENTITY_ATTACK, INSULT, PROFANITY, THREAT). No published per-language Top-3 Hit Rate benchmarks.
Strengths
- Free
- Per-attribute probability scores
- Simple integration
Weaknesses
- Small fixed attribute set (no 200-indicator taxonomy)
- No published per-language benchmark evidence
- No audit trail
- No custom taxonomy
- Black-box model
OpenAI Moderation API
Last verified September 26, 2026 · omni-moderation-latest (13 categories with per-category category_scores, text + image) · Source: OpenAI Moderation guide
OpenAI reports 85%+ precision on their test set. Per-category flags and category_scores are returned; no per-language breakdown published. Categories are broad (hate, harassment, self-harm, sexual, violence).
Strengths
- Free for OpenAI customers
- 13 categories with per-category scores
- Wide language coverage
Weaknesses
- Non-deterministic (LLM-based)
- No custom taxonomy
- No audit trail
- Broad categories (no 200-indicator granularity)
- No per-language benchmark evidence
Azure AI Content Safety
Last verified September 21, 2026 · Azure AI Content Safety text model - 4 categories, severity scale 0-7 · Source: Microsoft Learn: harm categories
Microsoft reports 90%+ precision on their test set. No per-language breakdown published. Categories are broad with severity levels (0-7 scale for text).
Strengths
- Azure ecosystem integration
- Severity levels
- Multi-modal (text + image)
Weaknesses
- 4 categories only (no 200-indicator taxonomy)
- No custom taxonomy
- No audit trail
- More expensive than Sentiqra
- No per-language benchmark evidence
All Benchmarks
Compare all 8 products
Benchmark Results
Per-language Top-3 Hit Rate
Benchmark Methodology
How we measure Top-3 Hit Rate
Full Comparison
Sentiqra vs all competitors
Quickstart Guide
First API call in under 5 minutes
Pricing
From Free to Enterprise
Performance Dashboard
Latency, uptime, throughput snapshot
Legal Disclaimer
Benchmark methodology claims
Try Trust & Safety Now
Test the detection engine yourself in our free playground. No API key required.
Open Playground