Model

GPT-5.4

CLOSED

OpenAI · GPT-5.4 family · released Mar 5, 2026

Flagship/'Thinking' tier of the GPT-5.4 family, confirmed real via TechCrunch and The Decoder launch coverage. GPT-5.4 Pro ($30/$180 per M, ARC-AGI-2 83.3%, GDPval 82%) and cheaper GPT-5.4 mini/nano also shipped the same month but are out of scope for this pass (not requested).

ReasoningCodingVisionFunction callingTool useAgentic

2473.4

Elo · rank #22

Parameters

Undisclosed

Active params

—

Context

1.05M tokens

Architecture

Undisclosed architecture (presumed dense transformer, reasoning-tuned); first general-purpose OpenAI model with native computer-use; ships in Thinking/Pro/mini/nano tiers

License

Proprietary (OpenAI API Terms)

Languages

—

API price (in/out)

$2.5 / $15

Modalities

text · vision

Benchmark results

Bar shows position within the tracked field; marker = field best

Agents' Last ExamAgents37.3%#9

best: GPT-5.6 · 52.7%

AIMEMath95.3%#14

best: GPT-5.2 · 100.0%

ARC-AGI-1Reasoning93.7%#4

best: GPT-5.6 · 96.5%

ARC-AGI-2Reasoning73.3%#6

best: GPT-5.6 · 92.5%

BrowseCompAgents82.7%#14

best: Kimi K3 · 91.2%

FrontierMath (Tiers 1-3)Math47.6%#6

best: GPT-5.6 · 89.0%

FrontierMath (Tier 4)Math27.1%#6

best: GPT-5.6 · 83.0%

GDPvalAgents83.0%#3

best: Seed 2.1 Pro · 87.9%

GPQA DiamondReasoning92.8%#12

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning39.8%#27

best: Claude Sonnet 5 · 57.4%

LiveCodeBenchCoding84.1%#27

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MMMUVision87.5%#2

best: Claude Fable 5 · 89.3%

OSWorld-VerifiedAgents75.0%#9

best: Claude Fable 5 · 85.0%

SWE-bench ProCoding57.7%#16

best: Claude Fable 5 · 80.0%

SWE-bench VerifiedCoding76.9%#33

best: Claude Fable 5 · 95.0%

Run it locally

Closed weights — available via API only. No local deployment.

Input / M tok

$2.5

Output / M tok

$15

GPT-5.4 family

Elo progression across releases

GPT-5.4Mar 20262473.4 GPT-5.4 ProMar 20262786.0 GPT-5.4 miniMar 20262098.8 GPT-5.4 nanoMar 20261965.6

Resources

APIAvailable via OpenAI API

API price $2.5/$15 · each benchmark row carries its own source badge (see methodology)