Quality benchmark

MentalHealthBench

How well a model handles realistic mental-health conversations, from everyday well-being to emergencies, graded against rubrics written by licensed clinicians.

“MentalHealthBench is designed to better represent real-world AI mental health usage, spanning the range of acuities, conversational topics, and user profiles such as teens, adults, caregivers, and clinicians across multiple languages and cultural contexts.”

Malik et al., OpenAI

Metric: Clipped task score
Scale: 0–1
Source: Paper ↗

Headline score

GPT-6 AstraOpenAI · Public API
0.578
Claude Opus 5.5Anthropic · Public API
0.524
GPT-6 LunaOpenAI · Public API
0.503
GPT-5OpenAI · Public API
0.444
Claude Sonnet 4.5Anthropic · Public API
0.402
Qwen3.5-122B-A10BQwen · Open weights
0.370
Qwen3.5-27BQwen · Open weights
0.367
GPT-4oOpenAI · Public API
0.326
Gemini 2.5 ProGoogle · Public API
0.295

Bars span the metric's full 0–1 scale; higher is better. Hover the info icon for settings a model ran with.

Acuity

Model
GPT-6 AstraOpenAI · Public API	0.580	0.593	0.571
Claude Opus 5.5Anthropic · Public API	0.537	0.533	0.514
GPT-6 LunaOpenAI · Public API	0.531	0.546	0.474
GPT-5OpenAI · Public API	0.461	0.511	0.412
Claude Sonnet 4.5Anthropic · Public API	0.404	0.448	0.386
Qwen3.5-122B-A10BQwen · Open weights	0.381	0.395	0.357
Qwen3.5-27BQwen · Open weights	0.371	0.402	0.353
Gemini 2.5 ProGoogle · Public API	0.274	0.347	0.289
GPT-4oOpenAI · Public API	0.244	0.374	0.353

User profile

Model
GPT-6 AstraOpenAI · Public API	0.570	0.646	0.640	0.570
Claude Opus 5.5Anthropic · Public API	0.506	0.629	0.511	0.561
GPT-6 LunaOpenAI · Public API	0.495	0.562	0.562	0.501
GPT-5OpenAI · Public API	0.415	0.604	0.505	0.481
Claude Sonnet 4.5Anthropic · Public API	0.385	0.504	0.435	0.426
Qwen3.5-122B-A10BQwen · Open weights	0.346	0.434	0.497	0.401
Qwen3.5-27BQwen · Open weights	0.346	0.440	0.477	0.387
GPT-4oOpenAI · Public API	0.312	0.363	0.416	0.337
Gemini 2.5 ProGoogle · Public API	0.276	0.344	0.466	0.298

Behavior axis (signed score)

Model
GPT-6 AstraOpenAI · Public API	0.133	0.203	0.080	-0.011	0.285	-0.015	0.254	0.090	0.114	0.182
GPT-6 LunaOpenAI · Public API	0.118	0.186	0.076	-0.004	0.183	-0.008	0.255	0.082	0.103	0.151
Claude Sonnet 4.5Anthropic · Public API	0.117	0.056	0.050	-0.103	0.180	-0.213	0.245	-0.179	0.082	-0.144
Claude Opus 5.5Anthropic · Public API	0.108	0.116	0.102	-0.108	0.313	-0.041	0.259	-0.065	0.092	0.007
Qwen3.5-122B-A10BQwen · Open weights	0.100	0.089	0.043	-0.161	0.102	-0.139	0.248	-0.195	0.085	-0.188
Qwen3.5-27BQwen · Open weights	0.099	0.090	0.051	-0.172	0.101	-0.165	0.242	-0.198	0.077	-0.196
GPT-4oOpenAI · Public API	0.086	0.097	0.085	-0.047	0.040	-0.312	0.234	-0.029	0.053	-0.106
Gemini 2.5 ProGoogle · Public API	0.059	0.039	0.018	-0.149	0.058	-0.349	0.230	-0.263	0.088	-0.304
GPT-5OpenAI · Public API	0.030	0.092	0.087	-0.107	0.221	-0.114	0.243	-0.046	0.041	-0.113