GPT-5.4 (Thinking): benchmark results

GPT-5.4 evaluated with thinking enabled. Provider: OpenAI. Released 2026-03-05. Access: API.

Aggregate Rating 1770 ± 12, rank #198 of 2168 rated models, from 74 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
MMCL-Bench26.5Overall (self-reported)100
MMCL-Bench - Empirical Discovery and Induction23.8Strict task pass rate (%) on the 21 empirical discovery task100
MMCL-Bench - Rubric Accuracy76.6Rubric-level accuracy (%): share of all rubric items satisfi100
MMCL-Bench - Rule System Application32.7Strict task pass rate (%) on the 49 rule-system tasks (a tas100
OpenAI GPT-5.5 System Card - Hard Negative Protein Binding Prediction3.46pass@4 (%)100
OpenAI GPT-5.6 System Card - Adversarial User Simulations - Emotional Reliance98.5not_unsafe message rate (%)100
OpenAI GPT-5.6 System Card - Adversarial User Simulations - Self-Harm97.7not_unsafe message rate (%)100
OpenAI GPT-5.6 System Card - Biology Refusal - Dual Use95.5Not unsafe rate (%)100
OpenAI GPT-5.6 System Card - Biology Refusal - Severe96.1Not unsafe rate (%)100
OpenAI GPT-5.6 System Card - Challenging Prompts - Non-Violent Illicit Behavior100not_unsafe rate (%)100
OpenAI GPT-5.6 System Card - Challenging Prompts - Self-Harm98.7not_unsafe rate (%)100
OpenAI GPT-5.6 System Card - Image Input - Extremism99.5not_unsafe rate (%)100

Interactive version: theaggregate.ai/model?slug=gpt-5-4-thinking · How It Works · Data refreshed daily, snapshot 2026-10-20.