GPT-5.3-Codex vs GPT-5.4
Updated October 21, 2026. 33 benchmarks have a result for both. Coding is ranked on the Vibe Coding Index; every other board is shown as published.
Decision reading
Pick GPT-5.4 for coding: it scores 37.7 on the Vibe Coding Index against 29.5 for GPT-5.3-Codex.
Pick GPT-5.3-Codex when output tokens cost 1.1x less, or when a solved task costs $0.59 against $1.02.
- Shared results
- 33
- GPT-5.3-Codex only
- 1
- GPT-5.4 only
- 76
- Like-for-like categories
- 6 of 9
Shared results can support a head-to-head reading. Results only one model has describe coverage, not superiority.
Which one for your work
Each card is decided by one number both models carry. Where only one has it, the card says so and makes no call.
Coding
Vibe Coding Index, four public coding benchmarks
GPT-5.4
GPT-5.4 scores 37.7 against 29.5.
One benchmark behind each score
Intelligence
AA Intelligence Index
GPT-5.4
GPT-5.4 leads, 39.0 against 32.5.
Same board, both measured
Reasoning
GPQA Diamond
No call
Only GPT-5.4 has a GPQA Diamond result, 91.7%.
Directional only
Math
AIME
No call
Only GPT-5.4 has a AIME result, 96.7%.
Directional only
Knowledge
MMLU Pro
No call
Only GPT-5.4 has a MMLU Pro result, 87.5%.
Directional only
Web design
Design Arena: websites
GPT-5.4
GPT-5.4 leads, 1199 against 1163.
Same board, both measured
Long documents
Listed context window
GPT-5.4
GPT-5.4 takes 1.05M tokens against 400K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
GPT-5.3-Codex
GPT-5.3-Codex $0.0088, GPT-5.4 $0.010 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
GPT-5.3-Codex
GPT-5.3-Codex $0.59, GPT-5.4 $1.02.
Published attempt cost and pass rate
At a glance
- Coding index
- GPT-5.3-Codex 29.5 from 1 benchmark, GPT-5.4 37.7 from 1 benchmark.
- Widest category
- Web design, 5.0 points apart on average over 8 shared benchmarks.
- Widest single gap
- LMArena WebDev: React: GPT-5.3-Codex 1395 against GPT-5.4 1454, 9.8 points apart on a 0 to 100 scale.
- Output price
- GPT-5.3-Codex $14.00 against GPT-5.4 $15.00 per million output tokens, 1.1x apart.
- Context window
- GPT-5.3-Codex takes 400K tokens, GPT-5.4 1.05M.
Benchmark by benchmark
First the score: boards each model leads, per category, where both have a result. Then every board, a category to a fold; bars are each result on a 0 to 100 scale, so compare the two in a row, not one row with another.
| Category | GPT-5.3-Codex | Bars | GPT-5.4 | Ties |
|---|---|---|---|---|
| Coding | 4 | 7 | ||
| Intelligence | 0 | 2 | ||
| Agentic work | 2 | 6 | ||
| Reasoning | 0 | 2 | ||
| Web design | 0 | 8 | ||
| General | 2 | 0 | ||
| All shared boards | 8 | 25 |
CodingLike-for-likeFour of these boards feed the Vibe Coding Index47
| Benchmark | GPT-5.3-Codex | GPT-5.4 | Leader |
|---|---|---|---|
| LMArena WebDev | 1407 | 1464 | GPT-5.4 |
| AA Coding Index | Not measured | 71.1 | Only GPT-5.4 measured |
| LiveCodeBench (Vals AI) | 87.3% | 84.1% | GPT-5.3-Codex |
| SWE-bench Verified (Vals AI) | 78.0% | 78.2% | GPT-5.4 |
| Vibe Code Bench (Vals AI) | 61.8% | 67.4% | GPT-5.4 |
| Code migration (Vals AI) | Not measured | 35.0% | Only GPT-5.4 measured |
| ProgramBench (Vals AI) | Not measured | 0.5% | Only GPT-5.4 measured |
| IOI (Vals AI) | 53.8% | Not measured | Only GPT-5.3-Codex measured |
| SkillsBench (Vals AI) | Not measured | 51.7% | Only GPT-5.4 measured |
| Terminal-Bench Hard (Artificial Analysis run) | 53.0% | 57.6% | GPT-5.4 |
| Terminal-Bench | 78.4% | 81.8% | GPT-5.4 |
| GSO-Bench | Not measured | 31.4% | Only GPT-5.4 measured |
| WeirdML | 79.3% | 77.7% | GPT-5.3-Codex |
| MirrorCode | Not measured | 15.6% | Only GPT-5.4 measured |
| ALE-Bench | 1655.2 | 1607.0 | GPT-5.3-Codex |
| LMArena Text: coding | Not measured | 1496 | Only GPT-5.4 measured |
| LMArena Text: software and IT | Not measured | 1491 | Only GPT-5.4 measured |
| LMArena WebDev: React | 1395 | 1454 | GPT-5.4 |
| LMArena WebDev: HTML | 1422 | 1478 | GPT-5.4 |
| LMArena WebDev: from an image | 1443 | 1436 | GPT-5.3-Codex |
| Neither model has a result on the other 18 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | GPT-5.3-Codex | GPT-5.4 | Leader |
|---|---|---|---|
| AA Intelligence Index | 32.5 | 39.0 | GPT-5.4 |
| Epoch Capabilities Index | 157.1 | 157.3 | GPT-5.4 |
Agentic workLike-for-likePublic boards, shown as published, kept out of the index26
| Benchmark | GPT-5.3-Codex | GPT-5.4 | Leader |
|---|---|---|---|
| Design Arena: agents building web apps | 1033 | 1059 | GPT-5.4 |
| Design Arena: agents building full-stack apps | 983 | 1017 | GPT-5.4 |
| Design Arena: agents building mobile apps | 1054 | 1087 | GPT-5.4 |
| Design Arena: agents building Android apps | 1084 | 1037 | GPT-5.3-Codex |
| Design Arena: agents building Godot games | 1124 | 1135 | GPT-5.4 |
| τ²-Bench Telecom | 86.0% | 87.1% | GPT-5.4 |
| APEX-Agents | Not measured | 52.4% | Only GPT-5.4 measured |
| METR time horizons | 74.5% | 74.3% | GPT-5.3-Codex |
| DeepResearch Bench | Not measured | 35.1% | Only GPT-5.4 measured |
| PostTrainBench | Not measured | 19.0% | Only GPT-5.4 measured |
| Vending-Bench 2 | $5,940 | $6,144 | GPT-5.4 |
| Neither model has a result on the other 28 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | GPT-5.3-Codex | GPT-5.4 | Leader |
|---|---|---|---|
| GPQA Diamond | Not measured | 91.7% | Only GPT-5.4 measured |
| MMMU Pro (Vals AI) | Not measured | 87.5% | Only GPT-5.4 measured |
| GPQA Diamond (Artificial Analysis run) | 91.5% | 92.0% | GPT-5.4 |
| Humanity's Last Exam (Artificial Analysis run) | 42.5% | 43.7% | GPT-5.4 |
| GPQA Diamond (Epoch AI run) | Not measured | 93.3% | Only GPT-5.4 measured |
| ARC-AGI | Not measured | 93.7% | Only GPT-5.4 measured |
| ARC-AGI-2 | Not measured | 74.0% | Only GPT-5.4 measured |
| Chess puzzles | Not measured | 48.0% | Only GPT-5.4 measured |
| Mystery game puzzles | Not measured | 37.0% | Only GPT-5.4 measured |
| EBR-bench | Not measured | 25.4% | Only GPT-5.4 measured |
| LMArena Text: hard prompts | Not measured | 1485 | Only GPT-5.4 measured |
| LMArena Text: expert | Not measured | 1507 | Only GPT-5.4 measured |
| LMArena Text: hard prompts in English | Not measured | 1479 | Only GPT-5.4 measured |
| LMArena Vision: diagrams | Not measured | 1327 | Only GPT-5.4 measured |
| LMArena Vision: homework | Not measured | 1336 | Only GPT-5.4 measured |
| Neither model has a result on the other 6 boards here | |||
MathDirectional onlyPublic boards, shown as published, kept out of the index7 boards, none shared
| Benchmark | GPT-5.3-Codex | GPT-5.4 | Leader |
|---|---|---|---|
| AIME | Not measured | 96.7% | Only GPT-5.4 measured |
| OTIS Mock AIME | Not measured | 97.8% | Only GPT-5.4 measured |
| FrontierMath, tiers 1 to 3 | Not measured | 78.6% | Only GPT-5.4 measured |
| FrontierMath, tier 4 | Not measured | 49.0% | Only GPT-5.4 measured |
| ProofBench | Not measured | 56.0% | Only GPT-5.4 measured |
| LMArena Text: math | Not measured | 1488 | Only GPT-5.4 measured |
| LMArena Text: mathematical work | Not measured | 1497 | Only GPT-5.4 measured |
| Neither model has a result on the other 6 boards here | |||
KnowledgeDirectional onlyPublic boards, shown as published, kept out of the index13 boards, none shared
| Benchmark | GPT-5.3-Codex | GPT-5.4 | Leader |
|---|---|---|---|
| MMLU Pro | Not measured | 87.5% | Only GPT-5.4 measured |
| LegalBench (Vals AI) | Not measured | 86.0% | Only GPT-5.4 measured |
| CaseLaw (Vals AI) | Not measured | 63.8% | Only GPT-5.4 measured |
| CorpFin (Vals AI) | Not measured | 65.3% | Only GPT-5.4 measured |
| TaxEval (Vals AI) | Not measured | 74.0% | Only GPT-5.4 measured |
| MortgageTax (Vals AI) | Not measured | 68.3% | Only GPT-5.4 measured |
| MedQA (Vals AI) | Not measured | 96.1% | Only GPT-5.4 measured |
| MedCode (Vals AI) | Not measured | 41.3% | Only GPT-5.4 measured |
| SimpleQA Verified | Not measured | 45.1% | Only GPT-5.4 measured |
| LMArena Text: science | Not measured | 1479 | Only GPT-5.4 measured |
| LMArena Text: medicine and healthcare | Not measured | 1458 | Only GPT-5.4 measured |
| LMArena Text: legal and government | Not measured | 1486 | Only GPT-5.4 measured |
| LMArena Text: business and finance | Not measured | 1474 | Only GPT-5.4 measured |
| Neither model has a result on the other 4 boards here | |||
Web designLike-for-likePublic boards, shown as published, kept out of the index08
| Benchmark | GPT-5.3-Codex | GPT-5.4 | Leader |
|---|---|---|---|
| Design Arena: websites | 1163 | 1199 | GPT-5.4 |
| Design Arena: UI components | 1137 | 1215 | GPT-5.4 |
| Design Arena: game dev | 1162 | 1195 | GPT-5.4 |
| Design Arena: data viz | 1164 | 1196 | GPT-5.4 |
| Design Arena: 3D | 1008 | 1065 | GPT-5.4 |
| Design Arena: all code categories | 1145 | 1180 | GPT-5.4 |
| Design Arena: SVG | 1129 | 1141 | GPT-5.4 |
| Design Arena: ASCII art | 1151 | 1188 | GPT-5.4 |
| Neither model has a result on the other 1 board here | |||
GeneralLike-for-likePublic boards, shown as published, kept out of the index20
| Benchmark | GPT-5.3-Codex | GPT-5.4 | Leader |
|---|---|---|---|
| SAGE (Vals AI) | Not measured | 43.3% | Only GPT-5.4 measured |
| MedScribe (Vals AI) | Not measured | 77.5% | Only GPT-5.4 measured |
| IFBench | 75.4% | 73.9% | GPT-5.3-Codex |
| Long context reasoning (AA-LCR) | 83.3% | 82.0% | GPT-5.3-Codex |
| DTBench | Not measured | 94.4% | Only GPT-5.4 measured |
| LMCA | Not measured | 52.0% | Only GPT-5.4 measured |
| CL-bench | Not measured | 27.9% | Only GPT-5.4 measured |
| Furniture assembly | Not measured | 37.5% | Only GPT-5.4 measured |
| CL-bench Life | Not measured | 21.7% | Only GPT-5.4 measured |
| LMArena Text | Not measured | 1469 | Only GPT-5.4 measured |
| LMArena Text: instruction following | Not measured | 1469 | Only GPT-5.4 measured |
| LMArena Text: creative writing | Not measured | 1439 | Only GPT-5.4 measured |
| LMArena Vision | Not measured | 1303 | Only GPT-5.4 measured |
| LMArena Text: English | Not measured | 1467 | Only GPT-5.4 measured |
| LMArena Text: non-English | Not measured | 1465 | Only GPT-5.4 measured |
| LMArena Text: Chinese | Not measured | 1520 | Only GPT-5.4 measured |
| LMArena Text: German | Not measured | 1472 | Only GPT-5.4 measured |
| LMArena Text: Japanese | Not measured | 1485 | Only GPT-5.4 measured |
| LMArena Text: Korean | Not measured | 1449 | Only GPT-5.4 measured |
| LMArena Text: French | Not measured | 1493 | Only GPT-5.4 measured |
| LMArena Text: Spanish | Not measured | 1452 | Only GPT-5.4 measured |
| LMArena Text: Russian | Not measured | 1480 | Only GPT-5.4 measured |
| LMArena Text: Polish | Not measured | 1472 | Only GPT-5.4 measured |
| LMArena Text: multi-turn | Not measured | 1482 | Only GPT-5.4 measured |
| LMArena Text: longer queries | Not measured | 1473 | Only GPT-5.4 measured |
| LMArena Text: writing and language | Not measured | 1458 | Only GPT-5.4 measured |
| LMArena Text: entertainment and media | Not measured | 1438 | Only GPT-5.4 measured |
| LMArena Vision: English | Not measured | 1301 | Only GPT-5.4 measured |
| LMArena Vision: Chinese | Not measured | 1354 | Only GPT-5.4 measured |
| LMArena Vision: reading text | Not measured | 1312 | Only GPT-5.4 measured |
| LMArena Vision: humor | Not measured | 1284 | Only GPT-5.4 measured |
| LMArena Vision: recognizing things | Not measured | 1251 | Only GPT-5.4 measured |
| LMArena Vision: creative writing | Not measured | 1283 | Only GPT-5.4 measured |
| LMArena Document | Not measured | 1471 | Only GPT-5.4 measured |
| Neither model has a result on the other 12 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | GPT-5.3-Codex | GPT-5.4 | Leader |
|---|---|---|---|
| Neither model has a result on 5 boards here | |||
What each workload costs
Three fixed token mixes at each model's listed price, so per-token rates become a bill.
Chat turn
1K input + 500 output tokens
GPT-5.3-Codex costs less
Repository review
50K input + 3K output tokens
GPT-5.3-Codex costs less
Agent loop
200K input + 10K output tokens
GPT-5.3-Codex costs less
Listed OpenRouter rates, without cache or batch discounts. A workload that does not fit a model's context window would need splitting.
Specification differences
The leading value in a row is marked; a row with a missing value names no leader. Our two rows are marked in red.
| Specification | OpenAI GPT-5.3-Codex | OpenAI GPT-5.4 |
|---|---|---|
| Maker | OpenAI | OpenAI |
| Weights | No public weights listed | No public weights listed |
| Released | February 24, 2026 | March 5, 2026 |
| Context window | 400K | 1.05M |
| Input $/Mtok | $1.75 | $2.50 |
| Output $/Mtok | $14.00 | $15.00 |
| Coding index | 29.5 | 37.7 |
| Independent measurements behind it | 1 | 1 |
| Cost per solved task | $0.59 | $1.02 |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is GPT-5.3-Codex better than GPT-5.4 for coding?
Pick GPT-5.4 for coding: it scores 37.7 on the Vibe Coding Index against 29.5 for GPT-5.3-Codex. Pick GPT-5.3-Codex when output tokens cost 1.1x less, or when a solved task costs $0.59 against $1.02.
Which is cheaper, GPT-5.3-Codex or GPT-5.4?
GPT-5.3-Codex costs $1.75 per million input tokens and $14.00 per million output tokens; GPT-5.4 costs $2.50 and $15.00. Per solved SWE-bench Verified task, GPT-5.3-Codex costs $0.59 and GPT-5.4 $1.02.
Which has the bigger context window, GPT-5.3-Codex or GPT-5.4?
GPT-5.4, with 1.05M tokens against 400K.
How many benchmarks is the GPT-5.3-Codex vs GPT-5.4 comparison based on?
33 benchmarks have a result for both models as of October 21, 2026. GPT-5.3-Codex's coding index rests on 1 independent benchmark and GPT-5.4's on 1. Where only one model has a result, the page says so instead of naming a winner.