GPT-5.4

Released

GPT-5.4 is an AI model released by OpenAI on Thursday, Mar 5 2026, 21 days after GPT-5.3-Codex-Spark. Benchmark results (shown below) cover BullshitBench v2, ProgramBench, Next.js Evals, Terminal-Bench 2.0, Expert-SWE (Internal), Toolathlon, and 8 more.

API pricing

Input
$2.50
Cached input
$0.25
Output
$15.00
  • Reasoning or thinking is supported.
USD per 1M tokens

Available from

Azure$2.50$15.001M
Azureeu$2.75$16.501M
Azureus$2.75$16.501M
Amazon Bedrockus-east-1$2.75$16.501M
USD per 1M tokens

Benchmarks

Coding

ProgramBench
#5 of 17
Next.js Evals
65%
#16 of 28
Expert-SWE (Internal)
68.5%
#2 of 2

Terminal & CLI

Terminal-Bench 2.0
75.1%
#3 of 14

Agentic & tool use

Toolathlon
54.6%
#3 of 5
BrowseComp
82.7%
#19 of 28
CyberGym
79%
#7 of 10
OSWorld-Verified
75%
#13 of 28

Reasoning & science

Humanity's Last Exam
52.1%
with tools
#20 of 42
ARC-AGI-2
73.95%
#10 of 21
FrontierMath
47.6%
Tier 1–3
27.1%
Tier 4
Tier 1–3
#4 of 6
Tier 4
#4 of 6
GPQA Diamond
92.8%
#7 of 59

Knowledge work

GDPval (win/tie rate)
83%
#2 of 6

Robustness

BullshitBench v2
48%
#31 of 77

Source: ProgramBench, retrieved 11 September 2026 · Source: BenchLM, retrieved 31 August 2026. Every other score here is the figure the lab published at launch.

Compare GPT-5.4 with

GPT-5.4

Suggested comparisons

Frequently asked questions

GPT-5.4 was released by OpenAI on Thursday, Mar 5 2026.

All OpenAI releases

44 tracked

2022

1 release