name = "GPT-5.4 mini"
description = "Strong small GPT for coding subagents, quick tool use, and high-volume work"
family = "gpt-mini"
release_date = "2026-03-17"
last_updated = "2026-03-17"
attachment = true
reasoning = true
temperature = true
tool_call = true
structured_output = true
knowledge = "2025-08-31"
open_weights = false

[limit]
context = 400_000
input = 272_000
output = 128_000

[modalities]
input = ["text", "image"]
output = ["text"]

[[benchmarks]]
name = "SWE-Bench Pro"
score = 54.4
metric = "resolve rate"
variant = "reasoning effort xhigh"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "Terminal-Bench"
score = 60.0
metric = "accuracy"
variant = "reasoning effort xhigh"
version = "2.0"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "MCP Atlas"
score = 57.7
metric = "score"
variant = "reasoning effort xhigh"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "Toolathlon"
score = 42.9
metric = "score"
variant = "reasoning effort xhigh"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "τ²-Bench Telecom"
score = 93.4
metric = "accuracy"
variant = "reasoning effort xhigh"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "GPQA Diamond"
score = 88.0
metric = "accuracy"
variant = "reasoning effort xhigh"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 41.5
metric = "accuracy"
variant = "with tools"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "Humanity's Last Exam"
score = 28.2
metric = "accuracy"
variant = "without tools"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "OSWorld-Verified"
score = 72.1
metric = "success rate"
variant = "reasoning effort xhigh"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "MMMU Pro"
score = 78.0
metric = "accuracy"
variant = "with Python"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "MMMU Pro"
score = 76.6
metric = "accuracy"
variant = "without tools"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "OmniDocBench"
score = 0.1263
metric = "overall edit distance"
variant = "reasoning effort none"
version = "1.5"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "OpenAI MRCR"
score = 47.7
metric = "accuracy"
variant = "8-needle, 64K-128K"
version = "v2"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "OpenAI MRCR"
score = 33.6
metric = "accuracy"
variant = "8-needle, 128K-256K"
version = "v2"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "Graphwalks"
score = 76.3
metric = "accuracy"
variant = "BFS, 0-128K"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"

[[benchmarks]]
name = "Graphwalks"
score = 71.5
metric = "accuracy"
variant = "parents, 0-128K"
source = "https://openai.com/index/introducing-gpt-5-4-mini-and-nano/"
date = "2026-03-17"