BenchLeader
Menu
Best LLMsChoose a modelCoding plansLeaderboardsModelsBenchmarksProvidersReleases4

GPT-5.3-Codex vs GPT-5.4

Full profiles: GPT-5.3-Codex (OpenAI) and GPT-5.4 (OpenAI): every benchmark, price and rank.

Updated · Data from 12 sources

As of 2 Nov 2026, GPT-5.3-Codex and GPT-5.4 are neck and neck on quality (62.9 vs 63.3 on the BenchLeader Index, inside the ±7.4 margin), and GPT-5.3-Codex is 1.2× cheaper at $4.81 vs $5.63 per 1M tokens blended. Choose GPT-5.3-Codex for agents & tools (80 vs 64) and instruction following (74 vs 72) and fast first answers (64.7 s vs 198.2 s); choose GPT-5.4 for coding (65 vs 61).

Key facts (as of 2 Nov 2026)

  • GPT-5.3-Codex scores 62.9 on the BenchLeader Index and GPT-5.4 scores 63.3 (BenchLeader, 2 Nov 2026).
  • GPT-5.3-Codex costs $4.81 per 1M tokens blended (list price, 2 Nov 2026).
  • GPT-5.4 costs $5.63 per 1M tokens blended (list price, 2 Nov 2026).
  • Of the 13 benchmarks both models have, GPT-5.4 leads 8 (BenchLeader, 2 Nov 2026).
GPT-5.3-Codex benchmarks xhigh OpenAI GPT-5.4 benchmarks xhigh OpenAIAdd a third model

Which to pick for…

The pick's numbers come first. Category scores on the BenchLeader scale (50 is the average model, 15 points one standard deviation). High confidence means a lead of half a standard deviation or more; under one point is called either.

Use casePickWhyConfidence
CodingGPT-5.465 vs 61 category score; LiveCodeBench 84.1% vs 87.3%Medium
Agents & tool useGPT-5.3-Codex80 vs 64 category score; Terminal-Bench Hard 53.0% vs 57.6%High
Long documentsEither67 vs 66 category score; AA-LCR v1.1 83.3% vs 82.0%Either
Images and other mediaEither61 vs 61 category score; MMMU-Pro 78.5% vs 78.4%Either
Following instructionsGPT-5.3-Codex74 vs 72 category score; IFBench 75.4% vs 74.0%Medium
Lowest costGPT-5.3-Codex$4.81 vs $5.63 per 1M blended; agentic coding session $0.161 vs $0.210–
Fastest first answerGPT-5.3-Codex64.7 s (GPT-5.3-Codex (xhigh)) vs 198.2 s (GPT-5.4 (xhigh)) to the first answer token–
High volume on a budgetNeither flagshipA smaller model is usually the better fit: Cheapest good LLMs, Fastest LLMs–

Key differences

  • GPT-5.3-Codex (xhigh) and GPT-5.4 (xhigh) are neck and neck on quality (62.9 vs 63.3, within the ±7.4 margin).
  • GPT-5.3-Codex (xhigh) is stronger in agents & tools, instruction following, long context, multimodal.
  • GPT-5.4 (xhigh) is stronger in coding, knowledge, maths, reasoning.
  • GPT-5.3-Codex (xhigh) is 1.2× cheaper ($4.81 vs $5.63 per 1M blended).

At a glance

SpecGPT-5.3-CodexGPT-5.4
DeveloperOpenAIOpenAI
Released5 Feb 20265 Mar 2026
Context window (tokens)400,0001,050,000
Compared atGPT-5.3-Codex (xhigh)GPT-5.4 (xhigh)
Open weightsNoNo
Price in / out / cached input per 1M$1.75 / $14.00 / $0.175$2.50 / $15.00 / $0.250

Category by category

Category scores on the BenchLeader scale: 50 is the average evaluated model, 15 points is one standard deviation. A gap under one point is called level.

CategoryGPT-5.3-Codex (xhigh)GPT-5.4 (xhigh)Leader
Reasoning–66.1not comparable
Coding60.964.5GPT-5.4
Agents & tools80.063.6GPT-5.3-Codex
Maths–65.1not comparable
Knowledge–58.6not comparable
Instruction following73.572.3GPT-5.3-Codex
Multimodal61.161.0level
Long context66.665.9level

On one chart

Each domain on the index's own scale, where 50 is an average model. Click a domain to see the benchmarks behind it. Pick up to 5 models from the list.

2550 avg75CodingAgents & toolsInstructionfollowingLong contextMultimodalGPT-5.3-Codex, Coding: 60.9GPT-5.3-Codex, Agents & tools: 80.0GPT-5.3-Codex, Instruction following: 73.5GPT-5.3-Codex, Long context: 66.6GPT-5.3-Codex, Multimodal: 61.1GPT-5.4, Coding: 64.5GPT-5.4, Agents & tools: 63.6GPT-5.4, Instruction following: 72.3GPT-5.4, Long context: 65.9GPT-5.4, Multimodal: 61.0

Not drawn, measured for fewer than two of these models: Reasoning, Maths, Knowledge, Human preference.

ModelIndex

Where sources disagree

Across the 13 benchmarks both have, GPT-5.3-Codex leads 5, GPT-5.4 leads 8.

Overall indexGPT-5.3-Codex (xhigh)GPT-5.4 (xhigh)Leader
BenchLeader Index62.963.3neck and neck

Each publisher runs its own tasks, harnesses and effort settings, so their overall verdicts can differ; BenchLeader reconciles them on one scale. † measured at another effort setting of the same model (the publisher did not test the one compared here).

Price, speed and latency

List prices blended at three input tokens per output token. Per-task prices use the token counts on the methodology page. Cost, tokens and time per benchmark task are measured: Artificial Analysis’ averages over its whole Intelligence Index run.

MeasureGPT-5.3-Codex (xhigh)GPT-5.4 (xhigh)
Blended price per 1M tokens$4.81$5.63
Input / output per 1M$1.75 / $14.00$2.50 / $15.00
Cached input per 1M$0.175$0.250
Price per task: chat reply$0.0049$0.0055
Price per task: summarize a 30-page report$0.029$0.039
Price per task: code edit$0.032$0.037
Price per task: agentic coding session$0.161$0.210
Price per task: structured extraction$0.0063$0.0080
Agentic coding session, input mostly cached$0.090$0.109
Output speed114 tok/s108 tok/s
Time to first answer64.7 s198.2 s
Time to a complete answer69.1 s202.8 s

Each model is compared at its best-ranked reasoning-effort setting. GPT-5.3-Codex (xhigh) scores 62.9 and takes 65 s to its first answer token. GPT-5.4 (xhigh) scores 63.3 and takes 198 s to its first answer token; GPT-5.4 (low) scores 58.9 in 2 s, GPT-5.4 (high) scores 58.8 in 2 s. Higher effort means more thinking before the answer: better scores, longer waits and more tokens billed. How reasoning effort changes scores and cost.

Every benchmark

113 benchmarks, grouped by category; each name links to the board and its leaders. † measured at another effort setting of the same model, shown for reference and not counted as a win.

MetricPublisherGPT-5.3-Codex (xhigh)GPT-5.4 (xhigh)
BenchLeader IndexBenchLeader62.963.3
Agents & tools scoreBenchLeader80.063.6
Coding scoreBenchLeader60.964.5
Instruction following scoreBenchLeader73.572.3
Long context scoreBenchLeader66.665.9
Multimodal scoreBenchLeader61.161.0
Knowledge scoreBenchLeader–58.6
Maths scoreBenchLeader–65.1
Reasoning scoreBenchLeader–66.1
Blended price $/MBenchLeader$4.81$5.63
Output speedBenchLeader114 tok/s108 tok/s
Time to first answerBenchLeader64.7 s198.2 s
Context windowBenchLeader400,000 tokens1,050,000 tokens
Reasoning
GPQA DiamondEpoch AI Benchmarking Hub–93.3%
Humanity's Last ExamScale AI / CAIS–36.2%
Arena Hard PromptsArena (LMArena)–1487†
LiveBench 1.0 ReasoningLiveBench–88.1%
GPQA Diamond (AA)Artificial Analysis91.5%92.0%
Humanity's Last Exam (AA)Artificial Analysis42.5%43.7%
GPQA Diamond (Vals)Vals AI–91.7%
Kagi LLM BenchmarkKagi LLM Benchmark–63.8%†
ARC-AGI-1ARC Prize–93.7%
ARC-AGI-2ARC Prize–74.0%
ARC-AGI-3ARC Prize–0.2%†
CritPtArtificial Analysis16.9%23.4%
Chess PuzzlesEpoch AI Benchmarking Hub–48.0%
Mystery Game PuzzlesEpoch AI Benchmarking Hub–37.0%
CL-benchEpoch AI Benchmarking Hub–27.9%
CL-bench LifeEpoch AI Benchmarking Hub–21.7%
LMCAEpoch AI Benchmarking Hub–52.0%
DTBenchEpoch AI Benchmarking Hub–94.4%
ForecastBenchForecastBench–63.1%†
Coding
SWE-bench Verified (Epoch)Epoch AI Benchmarking Hub74.8%†76.9%†
SciCodeSciCode–56.6%
WeirdMLWeirdML77.9%77.7%
GSO-BenchGSO-Bench–31.4%
Arena CodingArena (LMArena)–1511†
Arena WebDevArena (LMArena)1407†1401†
Show all 113 benchmarks
MetricPublisherGPT-5.3-Codex (xhigh)GPT-5.4 (xhigh)
Coding
LiveBench 1.0 CodingLiveBench–77.5%
LiveCodeBenchVals AI87.3%84.1%
IOIVals AI53.8%–
SWE-bench (Vals)Vals AI78.0%78.2%
SWE-Bench ProScale AI SEAL–59.1%
Code MigrationVals AI–35.0%
ProgramBenchVals AI–0.0%
Vibe Code Bench v1.1Vals AI61.8%67.4%
SWE-Bench Pro (private)Scale AI SEAL–59.1%
SWE Atlas: Codebase QnAScale AI SEAL–40.8%
SWE Atlas: RefactoringScale AI SEAL–44.3%
SWE Atlas: Test WritingScale AI SEAL–44.4%
MirrorCodeEpoch AI Benchmarking Hub–15.6%†
DeepSWE v1.1Epoch AI Benchmarking Hub–51.8%
ALE-BenchEpoch AI Benchmarking Hub1655.21607†
AlgoTuneAlgoTune–1.9†
Terminal-Bench 2.1 (AA)Artificial Analysis–78.3%
GameBench 2 OverallDuelLab GameBench 2–55.2†
GameBench 2 by reasoning settingDuelLab GameBench 2–63.5
Agents & tools
Terminal-BenchTerminal-Bench78.4%†81.8%†
APEX-AgentsMercor–52.4%†
LiveBench Agentic CodingLiveBench–53.8%
Terminal-Bench HardArtificial Analysis53.0%57.6%
τ²-Bench Telecom (AA)Artificial Analysis86.0%87.1%
MCP AtlasScale AI SEAL–70.6%
HiL-BenchScale AI SEAL4.3%†9.7%†
τ³-bench Bankingτ²-bench–39.4%
τ³-Banking (AA)Artificial Analysis–39.6%
APEX-Agents (AA)Artificial Analysis–33.3%
Harvey's Legal Agent Benchmark v1Vals AI–0.0%
SkillsBenchVals AI–51.7%
Terminal-Bench 2.0 (Vals)Vals AI64.0%†58.4%†
DeepResearch BenchDeepResearch Bench–35.1%†
PostTrainBenchEpoch AI Benchmarking Hub–19.0%†
METR Time HorizonsMETR74.5%†74.3%
GBAEvalEpoch AI Benchmarking Hub–45.1%†
Vending-Bench 2Andon Labs5940.1†6144.2†
AA-Briefcase v1.1Artificial Analysis868–
DeepSearchQAGoogle–63.7%†
Maths
FrontierMath Tiers 1–3Epoch AI Benchmarking Hub–78.6%
FrontierMath Tier 4Epoch AI Benchmarking Hub–49.0%
OTIS Mock AIMEEpoch AI Benchmarking Hub–95.3%
ProofBenchVals AI–56.0%
LiveBench MathematicsLiveBench–94.2%
AIME (Vals)Vals AI–96.7%
AIME 2026MathArena–99.2%
HMMT February 2026MathArena–97.7%
USAMO 2026MathArena–95.2%
MathArena ApexMathArena–54.2%
Arena MathsArena (LMArena)–1460†
Knowledge
SimpleQA VerifiedEpoch AI Benchmarking Hub–45.1%
LiveBench 1.0 Data AnalysisLiveBench–79.3%
MMLU-ProVals AI–87.5%
LegalBenchVals AI–86.0%
CorpFinVals AI–65.3%
TaxEvalVals AI–74.0%
MedQAVals AI–96.1%
PRBench FinanceScale AI SEAL–45.6%†
PRBench LegalScale AI SEAL–44.4%†
MultiNRCScale AI SEAL–58.3%†
CaseLaw v2Vals AI–63.8%
MedCodeVals AI–41.3%
MedScribeVals AI–77.5%
EBR-benchEpoch AI Benchmarking Hub–25.4%
Instruction following
LiveBench LanguageLiveBench–82.6%
LiveBench Instruction FollowingLiveBench–70.2%
IFBenchArtificial Analysis75.4%74.0%
Arena Instruction FollowingArena (LMArena)–1461†
Human preference
Arena TextArena (LMArena)–1465†
EQ-Bench 4EQ-Bench–1258†
Arena Creative WritingArena (LMArena)–1434†
Arena Multi-turnArena (LMArena)–1480†
Arena Longer QueriesArena (LMArena)–1477†
Multimodal
Arena VisionArena (LMArena)–1292†
MMMU-ProArtificial Analysis78.5%78.4%
VISTAScale AI SEAL–50.9%†
MMMU-Pro (Vals)Vals AI–87.5%
MortgageTaxVals AI–68.3%
VTBScale AI SEAL–29.2%†
Blueprint-Bench 2Epoch AI Benchmarking Hub–27.1%†
Long context
AA-LCR v1.1Artificial Analysis83.3%82.0%
Arena DocumentArena (LMArena)–1470†
Composite
Epoch Capabilities IndexEpoch AI Benchmarking Hub–157.1†
LiveBench 1.0LiveBench–78.0%
AA Intelligence Index v4.3.2Artificial Analysis32.539.0
AA Progress Index v1Artificial Analysis301.3311.6
Vals RSI IndexVals AI–15.5%
Safety & honesty
MASKScale AI SEAL–89.7%†

Data as of 2026-11-02. Best configuration of each model; every score links to its source on the model pages.

DuelLab GameBench 2, release gb2-20261026-cb69502a12d9 (https://benchmarks.duellab.org/releases/gb2-20261026-cb69502a12d9/), CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Scores mapped to BenchLeader model names; not endorsed by DuelLab.

What changed

  • GPT-5.3-Codex (xhigh): 1 new result (AA-Briefcase v1.1).
  • GPT-5.3-Codex (xhigh): 1 new result (AA Progress Index v1).
  • GPT-5.4 (xhigh): 1 new result (AA Progress Index v1).
  • GPT-5.4 (xhigh): 2 new results (Harvey's Legal Agent Benchmark v1, Vals RSI Index).
  • GPT-5.4 (xhigh): 2 new results (USAMO 2026, τ³-bench Banking).
  • GPT-5.4 (xhigh): 1 score revision.
  • GPT-5.4 (xhigh): 1 new result (SWE Atlas: Codebase QnA).
  • GPT-5.4 (xhigh): 1 score revision.
  • GPT-5.4 (xhigh): BenchLeader Index 64.2 → 63.7.

Rankings: Best LLM overall · Best LLM for agents · Best cheap LLMs

GPT-5.3-Codex vs GPT-5.4: questions

Is GPT-5.3-Codex better than GPT-5.4?
GPT-5.3-Codex (xhigh) and GPT-5.4 (xhigh) are neck and neck on quality (62.9 vs 63.3, within the ±7.4 margin). The BenchLeader Index combines every independent quality benchmark; the two are within the margin of error as of 2 Nov 2026, so check the category scores for your use.
GPT-5.3-Codex vs GPT-5.4: which is cheaper, and will that change?
GPT-5.3-Codex is cheaper today: $4.81 against $5.63 per million tokens, blended at three input tokens per output token. No price change has been announced for either.
Is GPT-5.3-Codex or GPT-5.4 better for coding?
GPT-5.4 scores higher in coding (65 vs 61 on the category index, where 50 is average). On LiveCodeBench, the top shared coding benchmark, it is 87.3% for GPT-5.3-Codex against 84.1% for GPT-5.4.
Which is faster, GPT-5.3-Codex or GPT-5.4?
GPT-5.3-Codex streams faster (114 vs 108 output tokens per second, Artificial Analysis 7-day median).
Is GPT-5.3-Codex or GPT-5.4 better for agents and tool use?
GPT-5.3-Codex scores higher in agentic tasks (80 vs 64 on the category index, where 50 is average). On Terminal-Bench Hard, the top shared agents & tools benchmark, it is 53.0% for GPT-5.3-Codex against 57.6% for GPT-5.4.
More questions (4)
Which is better for long documents?
They are level on long-context benchmarks (67 vs 66). GPT-5.4 accepts more context: 1,050,000 against 400,000 tokens for GPT-5.3-Codex.
Which has lower latency?
GPT-5.3-Codex answers first: 64.7 s against 198.2 s to the first answer token, as Artificial Analysis measures it. Reasoning effort drives most of this: a higher setting thinks longer before it answers.
Do independent benchmarks agree on GPT-5.3-Codex vs GPT-5.4?
Across the 13 benchmarks both have, GPT-5.3-Codex leads 5, GPT-5.4 leads 8.
When was GPT-5.4 released?
GPT-5.4 was released on 5 Mar 2026.