What are GPT-5.4's benchmark scores?

As of 24 Oct 2026, GPT-5.4 (xhigh effort) scores 64.7 ±2.6 on the BenchLeader Index, top 50, strongest in composite (73) and weakest in knowledge (60). That rests on 32 independent benchmark results, and every score on this page links to its publisher.

How does GPT-5.4 rank?

GPT-5.4 (xhigh effort) is #45 of 785 ranked configurations on 24 Oct 2026 (top 50). 115 other configurations score within its margin, so the exact position matters less than the band. The leader is Claude Opus 5.5 (max effort) at 72.6.

How much does GPT-5.4 cost per 1M tokens?

As of 24 Oct 2026, GPT-5.4 costs $2.50 per 1M input tokens and $15.00 per 1M output tokens, $5.63 blended (three input tokens per output token).

Is GPT-5.4 better than Kimi K3?

As of 24 Oct 2026, GPT-5.4 (xhigh) and Kimi K3 (max) are neck and neck on quality (64.7 vs 64.7, inside the ±4.2 margin of error) and cost about the same, so speed decides it: GPT-5.4 (xhigh) streams 2.5× faster. The head-to-head page has the benchmark-by-benchmark detail.

When was GPT-5.4 released?

OpenAI released GPT-5.4 on 5 Mar 2026. BenchLeader has results for it at 7 reasoning-effort settings.

OpenAIReasoning modelCompare with another model

GPT-5.4

Reasoning effort

GPT-5.4 is an OpenAI proprietary reasoning model, released 5 Mar 2026. Its best configuration (xhigh reasoning effort) ranks #45 of 785 on the BenchLeader Index at 64.7 ±2.6, in the upper half. The ± is the point: 115 other configurations score within that range, so they and this one cannot be told apart on quality alone — price and speed are what separate them. It scores highest in composite (73) and lowest in knowledge (60). At $5.63 per million tokens blended it is among the most expensive ranked models. Output speed of 96 tokens per second puts it faster than most, with a first answer in 110.0 s. It has been measured at 7 reasoning-effort settings; this summary describes the best-scoring one, and the tabs above switch between them. Last measured 13 Oct 2026.

Blended price: $5.63/M; $2.50 in · $15.00 out
Output speed: 96 tok/s; measured by Artificial Analysis
First answer: 110 s; first token 1.53 s
Context: 1.1M
Full answer: 115 s; median, reasoning included
Cost per run: –
Released: 5 Mar 2026

How it scores by categoryDashed line = average model (50). One step of 15 = one standard deviation.

Overall index65
Reasoning69
Coding65
Agents & tools64
Maths65
Knowledge60
Instruction following72
Multimodal61
Long context66
Composite73

Versions

OpenAI has shipped 13 models under this name. Each is ranked on its own results; a newer version often has fewer results so far, which holds its index nearer the average until more arrive.

Model	Released	Index	Rank
GPT-5.6	9 Jul 2026	–	–
GPT-5.5xhigh	23 Apr 2026	66.1	#35
GPT-5.4xhighthis page	5 Mar 2026	64.7	#45
GPT-5.2xhigh	11 Dec 2025	60.7	#104
GPT-5.1high	12 Nov 2025	57.7	#170
GPT-5high	7 Aug 2025	57.9	#165
GPT-4.1	14 Apr 2025	48.3	#429
GPT-4.5	27 Feb 2025	49.7	#388
GPT 4 0125	25 Jan 2024	–	–
GPT-4 1106	6 Nov 2023	–	–
GPT 4 0613	13 Jun 2023	39.7	#677
GPT-4	14 Mar 2023	42.4	#596
GPT-3.5	–	–	–

Reasoning-effort configurations

The same model behaves differently depending on how much it is allowed to think. Each row is one setting, scored only on the benchmarks that were run at that setting. “Not stated” collects results from publishers that did not say which setting they used; for a reasoning model that is usually its thinking mode, but we do not assume it. Pick a setting here or at the top of the page to see its price, speed and category scores.

Effort	Index	Rank	Speed	First answer	Chat reply cost	Categories
no reasoning	52.6	#310	71 tok/s	0.88 s	$0.0055	Agents & tools 67 · Coding 55 · Composite 50 · Instruction following 50 · Knowledge 54 · Long context 54 · Maths 49 · Multimodal 53 · Reasoning 48
low	59.7	#125	72 tok/s	2.14 s	$0.0055	Agents & tools 72 · Composite 60 · Instruction following 65 · Knowledge 64 · Long context 63 · Maths 60 · Multimodal 61 · Reasoning 53
medium	55.0	#237	56 tok/s	1.53 s	$0.0055	Coding 48 · Maths 65 · Reasoning 61
high	57.9	#167	56 tok/s	1.53 s	$0.0055	Agents & tools 52 · Coding 55 · Human preference 67 · Knowledge 50 · Maths 67 · Multimodal 65 · Reasoning 61
xhighbest	64.7	#45	96 tok/s	110 s	$0.0055	Agents & tools 64 · Coding 65 · Composite 73 · Instruction following 72 · Knowledge 60 · Long context 66 · Maths 65 · Multimodal 61 · Reasoning 69
thinking	–	–	56 tok/s	1.53 s	$0.0055	Knowledge 70 · Multimodal 65
not stated	58.2	#155	56 tok/s	1.53 s	$0.0055	Agents & tools 59 · Coding 51 · Human preference 64 · Maths 64 · Multimodal 64 · Reasoning 62

How its index has moved

8 Sept 2026 to 23 Oct 2026

xhigh
low
not stated
high

The index is recomputed from scratch every day, so a line moves when a new benchmark result lands, when a publisher revises a score, or when the models it is normalised against change. Early movement usually means the score is still settling.

Benchmark results

See all LLM benchmarks. One column per reasoning effort. Rank is among every configuration of every model on that benchmark. Hover a score for the run it came from.

Reasoning

Benchmark	no reasoning	low	medium	high	xhigh	not stated	Source
GPQA Diamond	74.8%#152	84.8%#93	88.9%#55	89.9%#47	93.3%#19	–	Epoch AI Benchmarking Hub
Humanity's Last Exam	–	–	–	–	36.2%#9	–	Scale AI / CAIS
Arena Hard Prompts	–	–	–	1497#39	–	1487#55	Arena (LMArena)
LiveBench 1.0 Reasoningnot in index	–	–	–	–	88.1%#24	–	LiveBench
GPQA Diamond (AA)not in index	74.8%#254	87.1%#103	–	–	92.0%#41	–	Artificial Analysis
Humanity's Last Exam (AA)not in index	11.3%#296	30.8%#143	–	–	43.7%#56	–	Artificial Analysis
GPQA Diamond (Vals)not in index	–	–	–	–	91.7%#18	–	Vals AI
Kagi LLM Benchmarkeffort not stated	–	–	–	–	–	63.8%#40	Kagi LLM Benchmark
ARC-AGI-1	–	68.2%#147	86.2%#105	92.7%#53	93.7%#48	–	ARC Prize
ARC-AGI-2	–	29.2%#134	55.4%#100	67.5%#65	74.0%#51	–	ARC Prize
ARC-AGI-3	–	–	–	0.2%#58	–	–	ARC Prize
CritPt	0.6%#254	7.4%#142	–	–	23.4%#51	–	Artificial Analysis
Chess Puzzlesnot in index	3.0%#135	19.0%#95	37.0%#44	42.0%#36	48.0%#19	–	Epoch AI Benchmarking Hub
Mystery Game Puzzlesnot in index	16.0%#84	17.0%#75	28.0%#39	–	37.0%#19	–	Epoch AI Benchmarking Hub
CL-benchnot in index	–	–	–	–	27.9%#1	–	Epoch AI Benchmarking Hub
CL-bench Lifenot in index	–	–	–	19.3%#3	21.7%#2	13.8%#7	Epoch AI Benchmarking Hub
LMCAnot in index	–	–	–	–	52.0%#57	–	Epoch AI Benchmarking Hub
DTBenchnot in index	–	–	–	–	94.4%#56	–	Epoch AI Benchmarking Hub
ForecastBenchnot in indexeffort not stated	–	–	–	–	–	63.1%#48	ForecastBench

Coding

Benchmark	no reasoning	medium	high	xhigh	not stated	Source
SWE-bench Verified (Epoch)	–	–	76.9%#8	–	–	Epoch AI Benchmarking Hub
SciCode	–	–	–	56.6%#35	–	SciCode
WeirdML	57.4%#53	–	–	77.7%#24	–	WeirdML
GSO-Bench	–	–	25.5%#15	31.4%#12	–	GSO-Bench
Arena Coding	–	–	1521#34	–	1511#55	Arena (LMArena)
Arena WebDev	–	1444#74	1463#69	–	1400#89	Arena (LMArena)
LiveBench 1.0 Codingnot in index	–	–	–	77.5%#37	–	LiveBench
LiveCodeBench	–	–	–	84.1%#39	–	Vals AI
SWE-bench (Vals)not in index	–	–	–	78.2%#30	–	Vals AI
SWE-Bench Pro	–	–	–	59.1%#2	–	Scale AI SEAL
Code Migrationnot in index	–	–	–	35.0%#36	–	Vals AI
ProgramBenchnot in index	–	–	0.5%#18	0.0%#32	–	Vals AI
Vibe Code Bench v1.1not in index	–	–	–	67.4%#45	48.5%#62	Vals AI
SWE-Bench Pro (private)not in index	–	–	–	59.1%#2	–	Scale AI SEAL
SWE Atlas: Codebase QnAnot in index	–	–	–	40.8%#10	–	Scale AI SEAL
SWE Atlas: Refactoringnot in index	–	–	–	44.3%#8	–	Scale AI SEAL
SWE Atlas: Test Writingnot in index	–	–	–	44.4%#8	–	Scale AI SEAL
MirrorCodenot in index	–	–	15.6%#8	–	–	Epoch AI Benchmarking Hub
DeepSWE v1.1not in index	–	–	–	51.8%#46	–	Epoch AI Benchmarking Hub
ALE-Benchnot in index	1086.0#53	1520.7#23	1607#17	–	–	Epoch AI Benchmarking Hub
AlgoTunenot in index	–	–	1.9#3	–	–	AlgoTune
Terminal-Bench 2.1 (AA)not in index	–	–	–	78.3%#52	–	Artificial Analysis

Agents & tools

Benchmark	no reasoning	low	high	xhigh	not stated	Source
Terminal-Bencheffort not stated	–	–	–	–	81.8%#2	Terminal-Bench
APEX-Agentseffort not stated	–	–	–	–	52.4%#26	Mercor
Arena Agent	–	–	1.4#25	–	–	Arena (LMArena)
LiveBench Agentic Codingnot in index	–	–	–	53.8%#36	–	LiveBench
Terminal-Bench Hard	37.9%#62	43.2%#41	–	57.6%#11	–	Artificial Analysis
τ²-Bench Telecom (AA)not in index	36.0%#235	74.6%#140	–	87.1%#80	–	Artificial Analysis
MCP Atlas	–	–	–	70.6%#19	–	Scale AI SEAL
HiL-Bencheffort not stated	–	–	–	–	9.7%#28	Scale AI SEAL
τ³-bench Bankingnot in index	–	–	–	39.4%#10	–	τ²-bench
τ³-Banking (AA)	–	–	–	39.6%#31	–	Artificial Analysis
APEX-Agents (AA)not in index	–	–	–	33.3%#9	–	Artificial Analysis
Harvey's Legal Agent Benchmark v1not in index	–	–	–	0.0%#61	–	Vals AI
SkillsBenchnot in index	–	–	–	51.7%#19	–	Vals AI
Terminal-Bench 2.0 (Vals)not in index	–	–	58.4%#10	–	–	Vals AI
DeepResearch Benchnot in index	–	35.1%#39	–	–	–	DeepResearch Bench
PostTrainBenchnot in index	–	–	19.0%#12	–	–	Epoch AI Benchmarking Hub
METR Time Horizonsnot in index	–	–	–	74.3%#7	–	METR
GBAEvalnot in indexeffort not stated	–	–	–	–	45.1%#10	Epoch AI Benchmarking Hub
Vending-Bench 2not in indexeffort not stated	–	–	–	–	6144.2#21	Andon Labs
DeepSearchQAnot in indexeffort not stated	–	–	–	–	63.7%#3	Google

Maths

Benchmark	no reasoning	low	medium	high	xhigh	not stated	Source
FrontierMath Tiers 1–3	–	–	–	–	78.6%#19	–	Epoch AI Benchmarking Hub
FrontierMath Tier 4	–	–	–	–	49.0%#26	–	Epoch AI Benchmarking Hub
OTIS Mock AIME	57.8%#167	84.4%#98	95.6%#44	97.8%#31	95.3%#52	–	Epoch AI Benchmarking Hub
ProofBench	–	–	–	–	56.0%#24	–	Vals AI
LiveBench Mathematicsnot in index	–	–	–	–	94.2%#20	–	LiveBench
AIME (Vals)	–	–	–	–	96.7%#5	–	Vals AI
AIME 2026possible contamination	–	–	–	–	99.2%#3	–	MathArena
HMMT February 2026possible contamination	–	–	–	–	97.7%#2	–	MathArena
USAMO 2026not in index	–	–	–	–	95.2%#2	–	MathArena
MathArena Apexpossible contamination	–	–	–	–	54.2%#6	–	MathArena
Arena Maths	–	–	–	1491#29	–	1460#65	Arena (LMArena)

Knowledge

Benchmark	no reasoning	low	high	xhigh	thinking	Source
SimpleQA Verified	–	–	–	45.1%#39	–	Epoch AI Benchmarking Hub
LiveBench 1.0 Data Analysisnot in index	–	–	–	79.3%#15	–	LiveBench
AA-Omniscience	-15.7#224	4.8#108	–	5.8#103	–	Artificial Analysis
MMLU-Pro	–	–	–	87.5%#27	–	Vals AI
LegalBench	–	–	–	86.0%#17	–	Vals AI
CorpFin	–	–	–	65.3%#32	–	Vals AI
TaxEval	–	–	–	74.0%#42	–	Vals AI
MedQA	–	–	–	96.1%#5	–	Vals AI
PRBench Finance	–	–	45.6%#21	–	–	Scale AI SEAL
PRBench Legal	–	–	44.4%#22	–	–	Scale AI SEAL
MultiNRCpossible contamination	–	–	–	–	58.3%#7	Scale AI SEAL
CaseLaw v2not in index	–	–	–	63.8%#16	–	Vals AI
MedCodenot in index	–	–	–	41.3%#54	–	Vals AI
MedScribenot in index	–	–	–	77.5%#67	–	Vals AI
EBR-benchnot in index	–	–	–	25.4%#15	–	Epoch AI Benchmarking Hub
AA-Omniscience: accuracynot in index	37.4%#143	47.9%#73	–	50.9%#61	–	Artificial Analysis
AA-Omniscience: non-hallucinationnot in index	15.3%#336	17.4%#315	–	8.3%#463	–	Artificial Analysis

Instruction following

Benchmark	no reasoning	low	high	xhigh	not stated	Source
LiveBench Languagenot in index	–	–	–	82.6%#25	–	LiveBench
LiveBench Instruction Followingnot in index	–	–	–	70.2%#29	–	LiveBench
IFBench	48.4%#183	65.9%#101	–	74.0%#37	–	Artificial Analysis
Arena Instruction Followingnot in index	–	–	1471#37	–	1460#52	Arena (LMArena)

Human preference

Benchmark	high	not stated	Source
Arena Text	1475#37	1465#55	Arena (LMArena)
EQ-Bench 4effort not stated	–	1272#7	EQ-Bench
Arena Creative Writingnot in index	1442#55	1435#65	Arena (LMArena)
Arena Multi-turnnot in index	1493#21	1480#40	Arena (LMArena)
Arena Longer Queriesnot in index	1481#41	1477#50	Arena (LMArena)

Multimodal

Benchmark	no reasoning	low	high	xhigh	thinking	not stated	Source
Arena Vision	–	–	1303#18	–	–	1292#27	Arena (LMArena)
MMMU-Pro	70.6%#141	78.0%#78	–	78.4%#75	–	–	Artificial Analysis
VISTA	–	–	–	–	50.9%#8	–	Scale AI SEAL
MMMU-Pro (Vals)not in index	–	–	–	87.5%#14	–	–	Vals AI
MortgageTaxnot in index	–	–	–	68.3%#17	–	–	Vals AI
SAGEnot in index	–	–	–	43.3%#50	–	–	Vals AI
VTBnot in index	–	–	29.2%#2	–	–	–	Scale AI SEAL
Blueprint-Bench 2not in indexeffort not stated	–	–	–	–	–	27.1%#21	Epoch AI Benchmarking Hub

Long context

Benchmark	no reasoning	low	xhigh	not stated	Source	Trend
AA-LCR	58.3%#288	76.7%#140	82.0%#47	–	Artificial Analysis
Arena Documentnot in indexeffort not stated	–	–	–	1470#18	Arena (LMArena)

Composite

Benchmark	no reasoning	low	xhigh	not stated	Source
Epoch Capabilities Indexnot in indexeffort not stated	–	–	–	157.3#22	Epoch AI Benchmarking Hub
LiveBench 1.0not in index	–	–	78.0%#19	–	LiveBench
AA Intelligence Index v4.3.2	18.2#257	27.6#141	39.0#73	–	Artificial Analysis
Vals RSI Indexnot in index	–	–	15.5%#23	–	Vals AI

Safety & honesty

Benchmark	thinking	Source	Trend
MASKnot in indexpossible contamination	89.7%#8	Scale AI SEAL

Where it wins

Benchmarks where this configuration ranks in the top five of every configuration measured.

Where to run it

Every provider serving this model through OpenRouter, with throughput and first-token latency measured on live traffic over the last 30 minutes and each provider’s own price. Purple marks the best in each column.

Provider	Speed	First token	Input $/M	Output $/M	Context	Quantisation
Amazon Bedrock	124 tok/s	0.92 s	$2.75	$16.50	1.1M	–
Azure	76 tok/s	0.58 s	$2.75	$16.50	1.1M	–
Azure	66 tok/s	1.81 s	$2.50	$15.00	1.1M	–
OpenAI	66 tok/s	1.45 s	$2.50	$15.00	1.1M	–
OpenAI	26 tok/s	9.36 s	$1.25	$7.50	1.1M	–

What a task costs

Estimates from list price, output speed and time to first answer for the best configuration. “With caching” assumes three-quarters of the input is served from the prompt cache at $0.250 per 1M. Reasoning tokens are not modelled.

Workload	Tokens in / out	Cost	With caching	Time
Chat reply	400 / 300	$0.0055	$0.0048	1.9 min
Summarize a 30-page report	12,000 / 600	$0.039	$0.019	1.9 min
Code edit	6,000 / 1,500	$0.037	$0.027	2.1 min
Agentic coding session	60,000 / 4,000	$0.210	$0.109	2.5 min
Structured extraction	2,000 / 200	$0.0080	$0.0046	1.9 min

New this month

Compare GPT-5.4 with…

Questions people ask

What are GPT-5.4's benchmark scores?: As of 24 Oct 2026, GPT-5.4 (xhigh effort) scores 64.7 ±2.6 on the BenchLeader Index, top 50, strongest in composite (73) and weakest in knowledge (60). That rests on 32 independent benchmark results, and every score on this page links to its publisher.
How does GPT-5.4 rank?: GPT-5.4 (xhigh effort) is #45 of 785 ranked configurations on 24 Oct 2026 (top 50). 115 other configurations score within its margin, so the exact position matters less than the band. The leader is Claude Opus 5.5 (max effort) at 72.6.
How much does GPT-5.4 cost per 1M tokens?: As of 24 Oct 2026, GPT-5.4 costs $2.50 per 1M input tokens and $15.00 per 1M output tokens, $5.63 blended (three input tokens per output token).
Is GPT-5.4 better than Kimi K3?: As of 24 Oct 2026, GPT-5.4 (xhigh) and Kimi K3 (max) are neck and neck on quality (64.7 vs 64.7, inside the ±4.2 margin of error) and cost about the same, so speed decides it: GPT-5.4 (xhigh) streams 2.5× faster. The head-to-head page has the benchmark-by-benchmark detail.
When was GPT-5.4 released?: OpenAI released GPT-5.4 on 5 Mar 2026. BenchLeader has results for it at 7 reasoning-effort settings.

Data as of 24 Oct 2026. Compare these configurations.

Cite as: BenchLeader, “GPT-5.4: benchmarks, pricing, speed and rank”, data as of 24 Oct 2026, https://www.benchleader.com/models/gpt-5-4.