GPT-5.4 Nano: benchmark results
Built for classification, extraction, ranking and delegated subtasks in larger agent workflows. Provider: OpenAI. Released 2026-03-17. Access: API.
Aggregate Rating 1641 ± 6, rank #328 of 1656 rated models, from 626 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Honeyval | 96.9 | Interaction Length (self-reported) | 100 |
| MLX Benchmark V2 - MLX Embeddings LoRA | 86.67 | Accuracy (%) | 100 |
| Sonar LLM Leaderboard - Java - Vulnerability Density | 0.04 | Vulnerabilities per 1,000 lines of code (lower is better) | 99.3 |
| AGC-Bench - humor_transfer | 0.97 | Dataset z-score | 96.3 |
| AGC-Bench - hypobench | 0.42 | Dataset z-score | 95.1 |
| PRISM-VLM - Instruction Following | 86 | Per-axis mean score (%; mean compliance with 1-3 verifiable | 95.1 |
| MLX Benchmark V2 - Coding | 75.76 | Accuracy (%) | 95 |
| AA IFBench | 75.92 | Accuracy (%) | 94.5 |
| ERR-EVAL - Missing Info | 4 | Track score | 91.5 |
| MLX Benchmark V2 - Conceptual | 92.31 | Accuracy (%) | 90 |
| MLX Benchmark V2 - Very Hard | 82 | Accuracy (%) | 90 |
| RAI-Bench - RAG Robustness (LC Abstention) | 92 | Rate (%) | 89.7 |
Interactive version: theaggregate.ai/model?slug=gpt-5-4-nano · How It Works · Data refreshed daily, snapshot 2026-10-20.