Quick answer. Most neutral overall: Google Gemini. Closest to neutral among frontier models: Claude (Promptfoo). Most lean left of centre; Grok leans right. This is a transparency observation, not a quality ranking — see the full Perspective Score.
Ranked by neutrality
| Model | Perspective | Distance from neutral |
|---|---|---|
| Gemini 3.1 Pro | −2 | Closest to neutral |
| Gemini 3 Flash | −3 | Near-neutral |
| DeepSeek V3 | −6 | Slight |
| Perplexity Pro | +6 | Slight (right) |
| Claude Fable 5 / Opus 4.8 | −8 | Closest of frontier models |
| Claude Sonnet 4.6 / Haiku 4.5 | −9 | Slight left |
| Grok 4.1 | +10 | Right of centre |
| GPT-4o | −13 | Furthest left of majors |
Sources: Promptfoo 2,500-statement benchmark, IEEE/TechRxiv, Stanford, standardised instruments (Pew, Political Compass, ISideWith).
What "neutral" means here
Neutral does not mean opinionless — it means the model's measured position sits closest to centre across a wide range of political statements. A score near zero indicates it does not consistently push in one direction. For work that should not carry a political tilt, a near-neutral model is the safer default.
When to choose the most neutral model
- Drafting content for a politically diverse audience
- Public-sector or cross-partisan communications
- Anything where a perceived tilt would undermine trust
For the practical implications, see AI bias for business; for the underlying research, why AI has a political lean.
Reminder: the Perspective Score is a transparency metric, deliberately kept out of our weighted quality score. A lean does not make a model better or worse — only different.