GPT-5.4 Pro

+ contribute

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

Catalogue identity notes

Comparison model from the Epoch AI internal-evaluation CSVs (Epoch release date 2026-03-05). Epoch label gpt-5.4-pro-2026-03-05 (default and only snapshot). Distinct from gpt-5-4.

80.7±0.9#13 of 719 LLMsfrom 12 benchmark results across 5 sources

$ how it compares

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

#	model	score	where	RAM	released
11	Claude Fable 5 Anthropic	81±1	api $20.00/M		2026-06
12	Claude Mythos Preview (limited preview) Anthropic	81±1	limited preview		2026-04
13	GPT-5.4 Pro OpenAI	81±1	api $67.50/M		2026-03
14	GPT-5.5 OpenAI	80±1	api $11.25/M		2026-04
15	MiMo-V2.6-Pro Xiaomi	79±1	api $0.54/M +local	~443GB	2026-09
	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▁▁▃▆█▆▅▅▅▄▃▁ #16 of 300 * · 12% of its score · 2 tests

Trails GLM-5.3, Claude Haiku 5.5, DeepSeek V4.1 Flash here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
BrowseComp lab	89.3% ✎	▄··█▄··▄·▄█▄ beats 5/8	hard · avg 32%	▇▇▇ 6.6%	typical
Humanity's Last Exam with tools lab	58.7% ✎	▃·▃·▅····▅·█ beats 8/8	very hard · avg 12%	▇▇ 4.9%	at its level

math▁▂▃▅▆▅▄█▄▃▁▁ #14 of 398 · 62% of its score · 3 tests

Trails GPT-5.6 Terra, GPT-5.5 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
FrontierMath tier 4 v2	58.5% ✎	▁▂▁▂█▅▃▂▁▁▁▁ beats 50/63	very hard · avg 0.1%	▇▇▇▇▇▇▇▇▇▇▇▇▇ 32%	sharp · at its level
FrontierMath tiers 1–3 v2	82.5% ✎	▂··▁▁▃▄▄█▅▃▃ beats 70/82	very hard · avg 1.2%	▇▇▇▇▇▇▇▇ 21%	sharp
MathArena ArXivMath February 2026 (32 problems)	75.8% ✎	▁▁▁▁█▇▄▃▁▃·▃ beats 24/24	very hard · avg 9.1%	▇▇▇▇ 9.2%	at its level

web search▂▁▂▆▄█▃▃▁▁·▁ #15 of 80 * · 12% of its score · 2 tests

Beats Claude Fable 5.1, Claude Sonnet 5.5, Claude Opus 5 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	89.3% ✎	▄··█▄··▄·▄█▄ beats 5/8	hard · avg 32%	▇▇▇ 6.6%	typical
Humanity's Last Exam with tools lab	58.7% ✎	▃·▃·▅····▅·█ beats 8/8	very hard · avg 12%	▇▇ 4.9%	at its level

other tests7 not in a task · 26% of its score

test	result	vs the others in its run	difficulty	counts	why
Humanity's Last Exam lab	42.7% ✎	▃··▃·▃▃···▃█ beats 6/7	very hard · avg 4.4%	▇▇▇▇ 9.4%	at its level
Chess Puzzles (Epoch)	58.6% ✎	█▁▁▁▁▂▁▄▂▂▃▁ beats 131/135	very hard · avg 5.9%	▇▇▇ 7.9%	sharp · at its level · noisy
GPQA Diamond	94.6% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 203/208	hard · avg 40%	▇ 3.1%	sharp · maxed out
SimpleQA Verified	46.3% ✎	▁▁▃▂▃▃▆▆█▃▂▅ beats 44/78	very hard · avg 8.0%	▇ 3.0%	at its level
FrontierScience-Research lab	36.7% ✎	██·······█·█ beats 3/3	very hard · avg 7.4%	▇ 1.5%	at its level · small run
ARC-AGI-1 public evaluation set (ARC Prize model_baseline, pass@2)	98.3% ✎	▁·▇▃▁·█▇▁·▇▃ beats 27/27	very hard · avg 4.1%	▇ 1.4%	sharp · maxed out · noisy
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	91.7% ✎	▁▂▁▂▅█▃▃▄▃▁▁ beats 60/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

Through an API

api · $67.50/M · $30 in / $180 out per 1M tokens · 1.1M context · developers.openai.com ↗ · price sourceUSD per 1M tokens, OpenAI API standard processing, model page 'gpt-5.4-pro' whose default snapshot is gpt-5.4-pro-2026-03-05: input $30, output $180; cached-input, Batch and other tiers excluded. Prompts over 272K input tokens are billed at 2x input and 1.5x output (excluded). Currency: USD. OpenAI official model page, observed 2026-09-28.

$ how the score was worked out · Epoch AI — internal evaluatio… 67%, OpenAI vendor-reported benchm… 22%, MathArena ArXivMath / BrokenA… 9% and 2 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	5	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 67%	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	4	▇▇▇▇▇▇▇ 22%	Factual vendor-reported score (not licensed content)
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	1	▇▇▇ 9%	CC BY-SA 4.0
ARC Prize Foundation model_baseline outputs on ARC-AGI-1/2 public evaluation sets	1	▇ 1%	MIT
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 0%
weighted average		= 80.7