Atlas

Models

← All models

GPT-5.4 Nano

OpenAI · 2026-03-17

GPT-5.4 Nano is the smallest and cheapest tier of OpenAI's GPT-5.4 generation, optimized for simple, high-volume tasks such as classification, extraction, and ranking. It is built for low-latency subagent and pipeline workloads.

Benchmark scores

BenchmarkScore
AA-Briefcase Elo670
AA-Briefcase v1.1 Elo674
AA-LCR66.0
AA-LCR v1.176.7
AA-Omniscience Index-18.0
Agent Red Teaming k=100 (CAIS)81.7
Aikido Known-CVE Detection (Pass@3 Union)69.2
Aikido Known-CVE Detection (Single-Run Average)57.7
AIME 2024 (avg@8)85.4
AIME 2024-202588.8
AIME 2025 (avg@8)92.1
ALE-Bench1312
APEX-Agents-AA (avg@3)24.9
ARC-AGI-1 Semi-Private51.5
ARC-AGI-25.7
Arena Text — Business, Management & Financial Ops (No Style Control)1366
Arena Text — Chinese (No Style Control)1391
Arena Text — Creative Writing (No Style Control)1314
Arena Text — English (No Style Control)1381
Arena Text — Entertainment, Sports & Media (No Style Control)1324
Arena Text — Exclude Ties (No Style Control)1345
Arena Text — Exclude Ties (Style Controlled)1389
Arena Text — Expert (No Style Control)1396
Arena Text — French (No Style Control)1395
Arena Text — French (Style Controlled)1428
Arena Text — German (No Style Control)1368
Arena Text — Hard Prompts (No Style Control)1381
Arena Text — Hard Prompts English (No Style Control)1387
Arena Text — Instruction Following (No Style Control)1362
Arena Text — Japanese (No Style Control)1343
Loading Atlas data…