Atlas

Models

← All models

GPT-5.4 Pro

OpenAI · 2026-03-05

GPT-5.4 Pro is OpenAI's frontier professional-tier model released on March 5, 2026, available to Pro and Enterprise plan subscribers. It integrates the coding strengths of GPT-5.3-Codex with improved agentic tool-use, long-context reasoning up to 1 million tokens, and reduced factual error rates compared to prior generations. It targets complex professional workflows spanning coding, document analysis, and multi-step agentic tasks.

Benchmark scores

BenchmarkScore
Apex (MathArena)69.8
ARC-AGI-1 Semi-Private94.5
ArXivMath 02/202675.8
BrowseComp89.3
Capability161.6
Chess Puzzles (task version 2.0)59.0
CritPt (avg@5)30.0
Epoch Capabilities Index (ECI)160
Finance Agent v1.1 - release-reported score61.5
FORTRESS Average Risk Score (ARS)14.8
Frontier Science Research - release-reported score36.7
FrontierMath Tier 4 (v1) - Private38.0
FrontierMath Tier 4 (v2) - Private58.5
FrontierMath Tiers 1-3 (v1) - Private50.0
FrontierMath Tiers 1-3 (v2) - Private82.5
GDPval82.0
GeneBench-Pro16.3
GeneBench-Pro Artificial Analysis Subset9.2
GeneBench-Pro Externally Reviewed15.4
GeneBench-Pro Not Externally Reviewed17.9
GeneBench-Pro Public Release Subset20.0
Humanity's Last Exam44.3
Humanity's Last Exam (Text Only)45.3
Humanity's Last Exam (with tools)58.7
MASK91.7
MineBench — Bradley-Terry Rating1809
MineBench Conservative Rating1671
MultiChallenge69.2
MultiNRC62.3
OpenAI investment-banking modeling - internal release score83.6
Loading Atlas data…