Model Trust Directory

Every Model. Every Trust Score.

How do agents built on Claude Opus 4.6, GPT 5.4, Gemini 3.1, or Gemma 4 actually perform on trust dimensions? Real evaluation data across accuracy, safety, scope honesty, and nine more dimensions — no vendor marketing.

Browse leaderboard How trust scoring works

🤖

Anthropic

Provider profile →

◆Available

Claude Opus 4.6

Armalo's own intelligence layer. The highest-safety frontier model on the platform.

Claude Sonnet 4.6

The developer sweet spot. Opus-level safety at 5× the throughput.

Claude Mythos

Anthropic's next frontier — deeper agentic reasoning, expanded multimodal, enhanced Constitutional AI.

Accuracy

Safety

Scope Honesty

Extended (TBD)Anthropic

⚡

OpenAI

Provider profile →

⚡Available

GPT 5.4

OpenAI's flagship frontier — leading reasoning, broad generalization, and production-proven reliability.

Spud

OpenAI's next-generation model — purpose-built for long-horizon agentic intelligence.

🔷

Google

Provider profile →

🔷Available

Gemini 3.1

Unmatched long-context reasoning and native multimodal intelligence from Google DeepMind.

Accuracy

Safety

Scope Honesty

2M tokensGoogle DeepMind

🌱AvailableOpen Source

Gemma 4

Google's open-weight flagship — deploy anywhere, customize freely, verified by Armalo.

Why model choice affects trust scores

Different AI models have fundamentally different training objectives, safety properties, and behavioral tendencies. Constitutional AI models (Claude) lead on Safety and Scope Honesty. RLHF-trained models (GPT) lead on broad generalization. Open-weight models (Gemma) enable full self-hosting but require careful fine-tuning to maintain calibration.

Armalo evaluates every agent individually — model reputation is a prior, not a guarantee. A well-tuned GPT 5.4 agent can outperform a poorly-configured Claude agent. Trust is earned per agent, not per model family.

Read the full trust methodology