GPT-5.4 mini

+ contribute

GPT-5.4 mini is a closed-weights language model from OpenAI, released March 2026.

71.3±0.7#72 of 719 LLMsfrom 30 benchmark results across 12 sources

$ how it compares

The LLMs ranked either side of GPT-5.4 mini, out of 828 with a score, then more from OpenAI.

#	model	score	where	RAM	released
75	Solar Open 2 (250B-A15B) Upstage	72±1	local	~527GB	2026-07
76	North Small Translate 1.0 Cohere	71±6	api price unknown		2026-08
77	GPT-5.4 mini OpenAI	71±1	api $1.69/M		2026-03
78	Ling 3.0 Flash (124B-A5B) inclusionAI (Ant Group)	71±2	api $0.090/M +local	~75GB	2026-08
79	Mistral Large 4 (public preview) Mistral	71±2	api $2.06/M		2026-10
	more from OpenAI
1	GPT-6 Astra OpenAI	87±1	api $20.00/M		2026-09
3	GPT-6.1 Sol OpenAI	85±1	api $4.00/M		2026-09
7	GPT-6 Sol OpenAI	83±1	api $4.00/M		2026-09

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #61 of 300 · 18% of its score · 10 tests

Beats Gemini 3.6 Flash, GPT-5 Pro, LongCat 2.0 (1.6T-A48B) here, though they rank higher overall. Trails MiniMax-M3, GLM-5.1 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Terminal-Bench 2.0 lab	60.0% ✎	██···██··███ beats 2/6	very hard · avg 14%	▇ 3.2%	at its level
Toolathlon (Tool Decathlon) lab	42.9% ✎	▃··▃··▃··▃█▅ beats 2/8	very hard · avg 13%	▇ 2.9%	at its level
K-BrowseComp-Verified (Korean web-browsing agent questions, pass@1)	30.7% ✎	▃····█··▃█·▃ beats 7/8	very hard · avg 3.4%	▇ 2.9%	typical
Humanity's Last Exam with tools lab	41.5% ✎	▃·▃·▅····▅·█ beats 2/8	very hard · avg 12%	▇ 2.6%	at its level
OSWorld-Verified lab	72.1% ✎	█▄·····▄·▄▄▄ beats 4/6	hard · avg 17%	▇ 1.8%	at its level
τ²-bench Telecom lab	93.4% ✎	▃·▅▅▃▃··█▃▃▅ beats 9/13	hard · avg 16%	▇ 1.4%	sharp · noisy
Legal Research Bench (agentic, all-pass)	12.6% ✎	▄█··▄▄▄██·▄█ beats 0/12	very hard · avg 5.8%	▇ 1.4%	typical
SWE-bench Pro lab	54.4% ✎	▄·····▄·▄▄▄█ beats 2/6	hard · avg 28%	▇ 1.2%	blunt · at its level

2 more tests

test	result	vs the others in its run	difficulty	counts	why
KlinikeBench strict pass@1 (333 interactive clinical encounters, Terminus-2)	15.9% ✎	▁▁·▁▁▁▁▅▃█▅█ beats 15/27	very hard · avg 2.4%	▇ 0.9%	typical
T1-Bench multi-domain conversational agent pass^3	14.1% ✎	▂▂··▃·▂·▂▂█▄ beats 2/11	hard · avg 25%	▇ 0.0%	blunt · at its level · noisy

coding▁▁▁▃▆██▇▆▄▃▁ #86 of 458 * · 8.5% of its score · 2 tests

Beats Grok 4.20 (0309 reasoning) here, though they rank higher overall. Trails Llama 3.1 Nemotron Ultra 253B v1, GLM-5V-Turbo, GPT-5.4 nano here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena WebDev (Code Arena)	1397.0 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 43/119	avg 981	▇▇▇ 7.4%	sharp
SWE-bench Pro lab	54.4% ✎	▄·····▄·▄▄▄█ beats 2/6	hard · avg 28%	▇ 1.2%	blunt · at its level

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #56 of 393 * · 4.1% of its score · 1 test

Beats GPT-6 Luna, Grok 4.7, GPT-5.2 here, though they rank higher overall. Trails Gemma 4 31B, Claude Sonnet 4.5, Gemini 3.5 Flash-Lite here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1447.1 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 310/379	avg 1268	▇▇ 4.1%	blunt

math▁▂▃▅▆▅▄█▄▁▁▁ #56 of 398 · 31% of its score · 3 tests

Beats LongCat 2.0 (1.6T-A48B), Qwen3.6 Max Preview, Claude Sonnet 4.6 here, though they rank higher overall. Trails INTELLECT-3, GLM-4.7, A.X K2 (688B-A33B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
FrontierMath tiers 1–3 v2	51.2% ✎	▂··▁▁▃▄▄█▅▃▃ beats 36/82	very hard · avg 1.2%	▇▇▇▇▇▇▇▇ 20%	sharp · at its level
FrontierMath tier 4 v2	9.8% ✎	▁▂▁▃█▅▂▂▁▁▁▁ beats 11/63	very hard · avg 0.1%	▇▇▇▇ 10%	sharp
OTIS mock AIME 2024–2025	88.9% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 142/193	very hard · avg 8.5%	▇ 1.6%	sharp · noisy

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #43 of 147 * · 6.2% of its score · 2 tests

Beats MiMo-V2.6-Pro, MiMo-V2.6-Flash (309B-A15B), DeepSeek V4 Flash Vision (Exp) here, though they rank higher overall. Trails Qwen3.6 Flash, Gemini 3.5 Flash-Lite, Kimi K2.5 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
MMMU Pro lab	76.6% ✎	▅·····▃·▃·█▃ beats 2/7	hard · avg 40%	▇ 3.3%	at its level
Arena vision (style controlled)	1250.8 ✎	▁▁▂▂▂▂▄▅▅▆██ beats 103/145	avg 1068	▇ 2.8%	blunt

web search▂▁▂▆▄█▃▃▁▁·▁ #61 of 80 · 6.9% of its score · 3 tests

Beats DeepSeek V4 Flash (Preview, Apr 2026) here, though they rank higher overall. Trails GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503), Llama 4 Scout (109B-A17B) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
K-BrowseComp-Verified (Korean web-browsing agent questions, pass@1)	30.7% ✎	▃····█··▃█·▃ beats 7/8	very hard · avg 3.4%	▇ 2.9%	typical
Humanity's Last Exam with tools lab	41.5% ✎	▃·▃·▅····▅·█ beats 2/8	very hard · avg 12%	▇ 2.6%	at its level
Legal Research Bench (agentic, all-pass)	12.6% ✎	▄█··▄▄▄██·▄█ beats 0/12	very hard · avg 5.8%	▇ 1.4%	typical

other tests13 not in a task · 32% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA Diamond	86.9% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 154/208	hard · avg 40%	▇▇ 4.5%	sharp
Humanity's Last Exam lab	28.2% ✎	▃··▃·▃▃···▃█ beats 2/7	very hard · avg 4.4%	▇▇ 4.0%	at its level
ARC-AGI-2 public evaluation set (ARC Prize model_baseline, pass@2)	0.8% ✎	█··▂▅▁▃▁▃·▁▃ beats 6/24	very hard · avg 0.3%	▇▇ 4.0%	sharp · noisy
Mystery Game Puzzles (Epoch)	7.0% ✎	█··▂▃▇▄█▂▃▁▁ beats 9/74	very hard · avg 9.4%	▇ 3.2%	sharp · noisy
DSEffi-Bench Pass (1,000 data-science tasks, greedy)	62.2% ✎	▅··▃▅·█·▃██▃ beats 14/15	hard · avg 24%	▇ 2.9%	at its level
ARC-AGI-1 public evaluation set (ARC Prize model_baseline, pass@2)	66.3% ✎	▁·▇▃▁·█▇▁·▇▃ beats 15/27	very hard · avg 4.1%	▇ 2.8%	sharp · at its level · noisy
Chess Puzzles (Epoch)	24.0% ✎	█▁▁▁▁▂▁▄▂▂▁▁ beats 94/135	very hard · avg 5.9%	▇ 2.6%	sharp · noisy
The Token Games Elo (10-model puzzle-duel tournament)	1068.0 ✎	█··▄█··▄·▄·▄ beats 4/7	avg 815	▇ 2.4%	typical

5 more tests

test	result	vs the others in its run	difficulty	counts	why
SimpleQA Verified	29.4% ✎	▁▁▃▂▃▃▆▆█▃▂▅ beats 18/78	very hard · avg 8.0%	▇ 1.6%	at its level
Legal Research Bench (single-shot, no tools, all-pass)	2.2% ✎	▃▃█▃·▅···█▃▃ beats 0/12	very hard · avg 0.8%	▇ 1.5%	too hard for it
SkillsBench 87-task subset	44.8% ✎	▄··▄▄▄···█·█ beats 5/7	very hard · avg 8.2%	▇ 1.5%	at its level
WideSearch corrected dataset	70.2% ✎	▃······▃▃▃█▃ beats 3/7	hard · avg 20%	▇ 1.3%	at its level · noisy
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	94.5% ✎	▁▂▁▂▅█▂▃▄▃▁▁ beats 81/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

Through an API

api · $1.69/M · $0.75 in / $4.5 out per 1M tokens · 400k context · platform.openai.com ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

$ how the score was worked out · Epoch AI — internal evaluatio… 43%, OpenAI vendor-reported benchm… 20%, LMArena leaderboard dataset 14% and 9 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
Epoch AI — internal evaluations	7	▇▇▇▇▇▇▇▇▇▇▇▇▇ 43%	CC BY 4.0
OpenAI vendor-reported benchmark scores for its own models (openai.com announcement posts and Deployment Safety Hub system cards) · lab-reported	8	▇▇▇▇▇▇ 20%	Factual vendor-reported score (not licensed content)
LMArena leaderboard dataset	3	▇▇▇▇ 14%	CC BY 4.0
ARC Prize Foundation model_baseline outputs on ARC-AGI-1/2 public evaluation sets	2	▇▇ 7%	MIT
DSEffi-Bench: Demystifying Large Language Models' Capability in Efficient Data Science Code Generation — authors' own model runs	1	▇ 3%	CC BY 4.0
Legal Research Bench (Vals AI, arXiv:2610.00609v1)	2	▇ 3%	CC-BY-4.0
K-BrowseComp: A Web Browsing Agent Benchmark Grounded in Korean Contexts (arXiv:2606.02404v1)	1	▇ 3%	CC BY 4.0
inclusionAI/Ling-3.0-flash model card	2	▇ 3%	MIT
The Token Games: Evaluating Language Model Reasoning with Puzzle Duels — authors' own tournament	1	▇ 2%	CC BY-SA 4.0
KlinikeBench: Evaluating Language Models Beyond Diagnostic Accuracy (arXiv:2609.38480v2)	1	▇ 1%	CC BY 4.0
T1-Bench: Benchmarking Multi-Scenario Agents in Real-World Domains (arXiv:2606.11070v1)	1	▇ 0%	CC BY 4.0
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 71.3