GPT-5.4

+ contribute

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

77.6±0.5#20 of 719 LLMsfrom 54 benchmark results across 20 sources

$ how it compares

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

#	model	score	where	RAM	released
18	Qwen3.8 Max (2.4T-A95B) Alibaba	78±1	api $3.00/M +local	~380GB	2026-08
19	Claude Opus 4.8 Anthropic	78±0	api $10.00/M		2026-05
20	GPT-5.4 OpenAI	78±1	api $5.63/M		2026-03
21	DeepSeek V4.1 Flash DeepSeek	78±1	api $0.26/M +local	~537GB	2026-09
22	Claude Haiku 5.5 Anthropic	77±1	api $0.20/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▁▁▃▆█▆▅▅▅▄▃▁ #33 of 300 · 24% of its score · 16 tests

Trails GLM-5.3-Flash, Grok 4.7, DeepSeek V4 Flash Vision (Exp) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena Agent (October 2026 aggregate IPS)	0.0105 IPS ✎	▁▁▁▄▂▃▇█▂▂▁▃ beats 22/49	avg -0.312 IPS	▇▇ 4.5%	sharp
SWE-bench Verified	76.9% ✎	▁··▁·▁▂▁▄█▄▁ beats 23/29	hard · avg 19%	▇ 3.0%	sharp
Agents' Last Exam overall full-pass rate (fixed OpenClaw + GUI-as-Tool)	20.5% ✎	▄···█▄▄▄█▄·█ beats 9/10	very hard · avg 1.7%	▇ 2.4%	typical
BrowseComp lab	82.7% ✎	▄··█▄··▄·▄█▄ beats 1/8	hard · avg 32%	▇ 2.3%	typical
τ³-bench banking_knowledge (v1.0.1 grading)	39.4% ✎	▃▃▂▃·▆▃▂█▅▅▂ beats 16/25	very hard · avg 4.1%	▇ 1.8%	at its level
Toolathlon (Tool Decathlon) lab	54.6% ✎	▃··▃··▃··▃█▅ beats 6/8	very hard · avg 13%	▇ 1.6%	at its level
Humanity's Last Exam with tools lab	52.1% ✎	▃·▃·▅····▅·█ beats 4/8	very hard · avg 12%	▇ 1.5%	at its level
Terminal-Bench 2.0 lab	75.1% ✎	██···██··███ beats 4/6	very hard · avg 14%	▇ 1.4%	at its level

8 more tests

test	result	vs the others in its run	difficulty	counts	why
SWE Atlas (Codebase Q&A, Test Writing, Refactoring; mini-SWE-agent) pass@1	38.0% ✎	▄·▄··█···▄·█ beats 5/6	very hard · avg 6.1%	▇ 1.3%	at its level
Thinkingbox-bench pass@1 (507 stateful business-workflow agent tasks, N=20)	65.4% ✎	▅··▃·▃·█▃▅██ beats 14/15	very hard · avg 4.9%	▇ 1.1%	at its level
OSWorld-Verified lab	75.0% ✎	█▄·····▄·▄▄▄ beats 5/6	hard · avg 17%	▇ 0.9%	at its level
KlinikeBench strict pass@1 (333 interactive clinical encounters, Terminus-2)	23.4% ✎	▁▁·▁▁▁▁▅▃█▅█ beats 23/27	very hard · avg 2.4%	▇ 0.7%	typical
EEGAgentBench overall (tool-based EEG analysis agents, 6 tasks)	43.8% ✎	▂·▁▁··▇▇█▂·▁ beats 7/25	hard · avg 29%	▇ 0.7%	blunt · at its level
SWE-bench Pro lab	57.7% ✎	▄·····▄·▄▄▄█ beats 5/6	hard · avg 28%	▇ 0.6%	blunt · at its level
τ²-bench Telecom lab	92.8% ✎	▃·▅▅▃▃··█▃▃▅ beats 8/13	hard · avg 16%	▇ 0.4%	sharp · noisy
T1-Bench multi-domain conversational agent pass^3	20.9% ✎	▂▂··▂·▃·▂▂█▄ beats 3/11	hard · avg 25%	▇ 0.0%	blunt · at its level · noisy

coding▁▁▁▃▆██▇▆▄▃▁ #59 of 458 · 8.9% of its score · 4 tests

Trails MiMo-V2.5-Pro (1T-A42B), MiniMax-M3, GLM-5.1 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena WebDev (Code Arena)	1464.9 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 63/119	avg 981	▇▇ 4.0%	sharp
SWE-bench Verified	76.9% ✎	▁··▁·▁▂▁▄█▄▁ beats 23/29	hard · avg 19%	▇ 3.0%	sharp
SWE Atlas (Codebase Q&A, Test Writing, Refactoring; mini-SWE-agent) pass@1	38.0% ✎	▄·▄··█···▄·█ beats 5/6	very hard · avg 6.1%	▇ 1.3%	at its level
SWE-bench Pro lab	57.7% ✎	▄·····▄·▄▄▄█ beats 5/6	hard · avg 28%	▇ 0.6%	blunt · at its level

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #26 of 393 * · 2.2% of its score · 1 test

Beats Claude Sonnet 5.5, GPT-6 Sol, GPT-5.6 Terra here, though they rank higher overall. Trails GLM-5.2, Gemini 3.6 Flash, Gemini 3.5 Flash here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1475.2 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 351/379	avg 1268	▇ 2.2%	blunt

math▁▂▃▅▆▅▄█▄▃▁▁ #18 of 398 · 31% of its score · 10 tests

Beats Qwen3.8 Max (2.4T-A95B) here, though they rank higher overall. Trails GPT-5.6 Luna, GPT-6 Luna here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
FrontierMath tier 4 v2	49.0% ✎	▁▂▁▂█▅▂▂▁▁▁▁ beats 47/63	very hard · avg 0.1%	▇▇▇▇ 10%	sharp · at its level
FrontierMath tiers 1–3 v2	78.6% ✎	▂··▁▁▃▄▄█▅▃▃ beats 66/82	very hard · avg 1.2%	▇▇▇ 8.7%	sharp · at its level
MathArena ArXivMath February 2026 (32 problems)	59.4% ✎	▁▁▁▁█▇▄▃▃▃·▃ beats 20/24	very hard · avg 9.1%	▇ 2.9%	at its level
MathArena ArXivMath March 2026 (30 problems)	67.8% ✎	█▃▅·█▃▃·▅·▃▃ beats 11/14	very hard · avg 15%	▇ 2.3%	at its level
MathArena ArXivMath January 2026 (23 problems)	76.1% ✎	▂··▄▂▄█▆▆▄▄█ beats 23/23	hard · avg 22%	▇ 1.7%	at its level
MathArena BrokenArXiv March 2026 (56 problems)	36.6% ✎	█·▆█▂·▂▃···▂ beats 13/14	very hard · avg 0.8%	▇ 1.4%	sharp · at its level · noisy
MathArena BrokenArXiv February 2026 (31 problems)	37.9% ✎	▃▂▂▃█▃▂·▃··▂ beats 15/16	very hard · avg 0.7%	▇ 1.4%	sharp · at its level · noisy
MathArena ArXivMath December 2025 (17 problems)	60.3% ✎	▃·▃▃█▁·▄·▁▃▁ beats 17/18	hard · avg 19%	▇ 1.3%	at its level

2 more tests

test	result	vs the others in its run	difficulty	counts	why
MathArena ArXivLean March 2026 (41 problems)	17.1% ✎	▂····▂·····█ beats 5/5	very hard · avg 1.0%	▇ 1.0%	typical
OTIS mock AIME 2024–2025	95.3% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 160/193	very hard · avg 8.5%	▇ 0.3%	sharp · maxed out · noisy

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #25 of 147 · 4.0% of its score · 4 tests

Beats Claude Sonnet 5.5, MiMo-V2.6-Pro, GPT-5.6 Terra here, though they rank higher overall. Trails Gemini 3.6 Flash, Gemini 3.5 Flash, GLM-5.3-Flash here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena vision (style controlled)	1283.8 ✎	▁▁▂▂▂▂▄▅▅▆██ beats 128/145	avg 1068	▇ 1.5%	blunt
MMMU Pro lab	81.2% ✎	▅·····▃·▃·█▃ beats 5/7	hard · avg 40%	▇ 1.5%	typical
Arena document (style controlled)	1470.1 ✎	▁▁▃▃▄▃▇█▅▇▄▅ beats 24/37	avg 1335	▇ 0.7%	blunt
SpaceCast-Bench predictive spatial reasoning (overall, 16 task types)	37.7% ✎	▃·▅▃▅▃··▅··█ beats 4/11	very hard · avg 32%	▇ 0.3%	blunt · at its level · noisy

web search▂▁▂▆▄█▃▃▁▁·▁ #24 of 80 · 4.7% of its score · 3 tests

Beats Claude Opus 4.8 here, though they rank higher overall. Trails GPT-4o mini (2024-07-18), Mistral Small 3.1 24B (2503), GPT-4o (2024-08-06) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
BrowseComp lab	82.7% ✎	▄··█▄··▄·▄█▄ beats 1/8	hard · avg 32%	▇ 2.3%	typical
Humanity's Last Exam with tools lab	52.1% ✎	▃·▃·▅····▅·█ beats 4/8	very hard · avg 12%	▇ 1.5%	at its level
Arena search systems (style controlled)	1194.5 ✎	▁·▁▁▂▂▂█▅▂▁▁ beats 18/30	avg 1055	▇ 0.9%	blunt

other tests21 not in a task · 33% of its score

test	result	vs the others in its run	difficulty	counts	why
EBR-bench (Epoch)	25.4% ✎	▄▂·▆█▂▆▄▆▆·▄ beats 12/23	very hard · avg 0.3%	▇▇ 5.2%	sharp · at its level
Furniture Assembly (Epoch)	37.5% ✎	█▆▆▆▆▂▅▂▂▂·▅ beats 14/30	very hard · avg 1.2%	▇▇ 4.0%	sharp · at its level
Mystery Game Puzzles (Epoch)	37.0% ✎	█··▂▃▇▄█▂▃▁▁ beats 60/74	very hard · avg 9.4%	▇ 3.7%	sharp · at its level · noisy
ARC-AGI-2 public evaluation set (ARC Prize model_baseline, pass@2)	75.8% ✎	█··▂▅▁▃▁▃·▁▃ beats 22/24	very hard · avg 0.3%	▇ 2.8%	sharp · at its level · noisy
Humanity's Last Exam lab	39.8% ✎	▃··▃·▃▃···▃█ beats 4/7	very hard · avg 4.4%	▇ 2.6%	at its level
Chess Puzzles (Epoch)	44.0% ✎	█▁▁▁▁▂▁▄▂▃▁▁ beats 123/135	very hard · avg 5.9%	▇ 2.1%	sharp · at its level · noisy
KINA (Knowledge Index of Noah's Ark) accuracy, 899 ten-option items	48.5% ✎	▁▁▂▁▃▄█▆▄▅▂▃ beats 37/39	very hard · avg 22%	▇ 1.8%	blunt · at its level
DSEffi-Bench Pass (1,000 data-science tasks, greedy)	66.9% ✎	▅··▃▅·█·▃██▃ beats 15/15	hard · avg 24%	▇ 1.4%	at its level

13 more tests

test	result	vs the others in its run	difficulty	counts	why
EQ-Bench 3 (emotional intelligence, judge Elo)	1328.0 ✎	▁·▂▁▄▅▅▃█▄▁▁ beats 57/62	avg 1044	▇ 1.4%	typical
GPQA Diamond	93.3% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 195/208	hard · avg 40%	▇ 1.3%	sharp
WildClawBench	50.3% ✎	▃·▅▅█·▅█▅▃▃▃ beats 14/17	very hard · avg 11%	▇ 1.3%	at its level
AgentWorldBench (open-ended, overall)	58.3% ✎	▄▄··▄▄·▄█▄██ beats 11/11	hard · avg 25%	▇ 1.2%	blunt · at its level
SimpleQA Verified	45.1% ✎	▁▁▃▂▃▃▆▆█▃▂▅ beats 41/78	very hard · avg 8.0%	▇ 0.9%	at its level
ARC-AGI-1 public evaluation set (ARC Prize model_baseline, pass@2)	95.6% ✎	▁·▇▃▁·█▇▁·▇▃ beats 25/27	very hard · avg 4.1%	▇ 0.7%	sharp · noisy
CyberGym lab	79.0% ✎	▄···▄··▄█▄█▄ beats 5/8	hard · avg 31%	▇ 0.6%	small run
HealthBench Hard lab	30.3% ✎	▂··▂····▃▄·█ beats 2/8	very hard · avg 9.7%	▇ 0.5%	blunt · at its level
HealthBench lab	55.7% ✎	▃·▃·····█▃▃▅ beats 4/8	hard · avg 24%	▇ 0.5%	blunt · at its level
MirrorCode (Epoch)	15.6% ✎	█·▄▄·▄·▄·▄·█ beats 2/8	very hard · avg 1.5%	▇ 0.5%	sharp · at its level · noisy
FrontierScience-Research lab	33.0% ✎	██·······█·█ beats 2/3	very hard · avg 7.4%	▇ 0.4%	at its level · small run
DiagnosticIQ industrial maintenance MCQ (macro accuracy, 4 options)	73.3% ✎	▁··▃▁▁▁▅█▅▁▅ beats 23/24	medium · avg 52%	▇ 0.4%	blunt · at its level
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	93.0% ✎	▁▂▁▂▅█▂▃▄▃▁▁ beats 70/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

Through an API

api · $5.63/M · $2.5 in / $15 out per 1M tokens · 1.1M 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… 40%, OpenAI vendor-reported benchm… 15%, LMArena leaderboard dataset 14% and 17 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	11	▇▇▇▇▇▇▇▇▇▇▇▇ 40%	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	13	▇▇▇▇ 15%	Factual vendor-reported score (not licensed content)
LMArena leaderboard dataset	6	▇▇▇▇ 14%	CC BY 4.0
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	7	▇▇▇▇ 12%	CC BY-SA 4.0
ARC Prize Foundation model_baseline outputs on ARC-AGI-1/2 public evaluation sets	2	▇ 3%	MIT
Agents' Last Exam — authors' own fixed-harness model comparison	1	▇ 2%	CC BY 4.0
Knowledge Index of Noah's Ark (arXiv:2606.05104v2)	1	▇ 2%	CC BY 4.0
τ²/τ³-bench leaderboard: Sierra's own evaluation runs	1	▇ 2%	MIT
DSEffi-Bench: Demystifying Large Language Models' Capability in Efficient Data Science Code Generation — authors' own model runs	1	▇ 1%	CC BY 4.0
EQ-Bench 3 canonical leaderboard results	1	▇ 1%	MIT
WildClawBench: A Benchmark for Real-World, Long-Horizon Agent Evaluation (arXiv:2605.10912v1)	1	▇ 1%	CC BY 4.0
SWE Atlas: Benchmarking Coding Agents Beyond Issue Resolution (arXiv:2605.08366v1)	1	▇ 1%	CC BY 4.0
Qwen/Qwen-AgentWorld-35B-A3B model card	1	▇ 1%	Apache-2.0
One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows (arXiv:2608.19741v4)	1	▇ 1%	CC BY 4.0
KlinikeBench: Evaluating Language Models Beyond Diagnostic Accuracy (arXiv:2609.38480v2)	1	▇ 1%	CC BY 4.0
EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)	1	▇ 1%	CC BY 4.0
DiagnosticIQ: A Benchmark for LLM-Based Industrial Maintenance Action Recommendation from Symbolic Rules (arXiv:2605.08614v1)	1	▇ 0%	CC BY 4.0
SpaceCast-Bench: Evaluating Predictive Spatial Reasoning in Vision-Language Models (arXiv:2610.12402v1)	1	▇ 0%	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		= 77.6