GPT-5.2-Codex vs GPT-5.4 Mini
GPT-5.2-Codex leads GPT-5.4 Mini on coding, 45.4 to 35.8 on the Vibe Coding Index, and GPT-5.4 Mini costs 3.1x less per output token ($4.50 vs $14.00 per million). Updated October 22, 2026; 14 benchmarks have a result for both.
GPT-5.2-Codex or GPT-5.4 Mini: which to pick
Pick GPT-5.2-Codex for coding: it scores 45.4 on the Vibe Coding Index against 35.8 for GPT-5.4 Mini, though on 3 benchmarks against 5.
Pick GPT-5.4 Mini for output tokens 3.1x cheaper ($4.50 vs $14.00 per million) and work measured by LMArena WebDev or SWE-bench Verified (Vals AI), where it leads.
- Shared results
- 14
- GPT-5.2-Codex only
- 8
- GPT-5.4 Mini only
- 75
- Like-for-like categories
- 5 of 9
Only benchmarks that tested both models count toward a pick; a result only one model has is shown but not compared.
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, six public coding benchmarks
GPT-5.2-Codex
GPT-5.2-Codex scores 45.4 against 35.8.
3 and 5 benchmarks behind the scores
Intelligence
AA Intelligence Index
GPT-5.2-Codex
GPT-5.2-Codex leads, 28.5 against 24.1.
Same board, both measured
Agentic work
AA Agentic Index
No call
Only GPT-5.4 Mini has a AA Agentic Index result, 17.9.
Directional only
Reasoning
GPQA Diamond
No call
Only GPT-5.4 Mini has a GPQA Diamond result, 83.1%.
Directional only
Math
AIME
No call
Only GPT-5.4 Mini has a AIME result, 95.6%.
Directional only
Knowledge
MMLU Pro
No call
Only GPT-5.4 Mini has a MMLU Pro result, 84.5%.
Directional only
Long documents
Listed context window
No call
Both take 400K tokens.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
GPT-5.4 Mini
GPT-5.2-Codex $0.0088, GPT-5.4 Mini $0.0030 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
GPT-5.2-Codex
GPT-5.2-Codex $0.43, GPT-5.4 Mini $0.70.
Published attempt cost and pass rate
At a glance
- Coding index
- GPT-5.2-Codex 45.4 from 3 benchmarks, GPT-5.4 Mini 35.8 from 5 benchmarks.
- Widest category
- Agentic work, 8.8 points apart on average over 1 shared benchmark.
- Widest single gap
- Terminal-Bench Hard (Artificial Analysis run): GPT-5.2-Codex 37.1% against GPT-5.4 Mini 52.3%, 15.2 points apart on a 0 to 100 scale.
- Output price
- GPT-5.2-Codex $14.00 against GPT-5.4 Mini $4.50 per million output tokens, 3.1x apart.
- Context window
- Both take 400K tokens.
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-5.2-Codex | Bars | GPT-5.4 Mini | Ties |
|---|---|---|---|---|
| Coding | 2 | 6 | ||
| Intelligence | 1 | 0 | ||
| Agentic work | 1 | 0 | ||
| Reasoning | 2 | 0 | ||
| General | 2 | 0 | ||
| All shared boards | 8 | 6 |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index26
| Benchmark | GPT-5.2-Codex | GPT-5.4 Mini | Leader |
|---|---|---|---|
| SWE-bench Verified | 72.8% | Not measured | Only GPT-5.2-Codex measured |
| SWE-bench Multilingual | 66.3% | Not measured | Only GPT-5.2-Codex measured |
| LMArena WebDev | 1339 | 1397 | GPT-5.4 Mini |
| AA Coding Index | Not measured | 56.1 | Only GPT-5.4 Mini measured |
| LiveCodeBench (Vals AI) | 88.0% | 81.5% | GPT-5.2-Codex |
| SWE-bench Verified (Vals AI) | 72.4% | 73.0% | GPT-5.4 Mini |
| Terminal-Bench 4 (Vals AI) | Not measured | 2.5% | Only GPT-5.4 Mini measured |
| Vibe Code Bench (Vals AI) | 37.9% | 48.0% | GPT-5.4 Mini |
| Code migration (Vals AI) | Not measured | 12.9% | Only GPT-5.4 Mini measured |
| ProgramBench (Vals AI) | Not measured | 0.0% | Only GPT-5.4 Mini measured |
| Terminal-Bench 2.1 (Vals AI) | Not measured | 54.7% | Only GPT-5.4 Mini measured |
| SciCode | Not measured | 52.1% | Only GPT-5.4 Mini measured |
| Terminal-Bench Hard (Artificial Analysis run) | 37.1% | 52.3% | GPT-5.4 Mini |
| Terminal-Bench | 66.5% | Not measured | Only GPT-5.2-Codex measured |
| FrontierCode | Not measured | 27.0% | Only GPT-5.4 Mini measured |
| WeirdML | Not measured | 60.3% | Only GPT-5.4 Mini measured |
| ALE-Bench | 1299.9 | 1188.6 | GPT-5.2-Codex |
| LMArena Text: coding | Not measured | 1438 | Only GPT-5.4 Mini measured |
| LMArena Text: software and IT | Not measured | 1436 | Only GPT-5.4 Mini measured |
| LMArena WebDev: React | 1329 | 1386 | GPT-5.4 Mini |
| LMArena WebDev: HTML | 1348 | 1419 | GPT-5.4 Mini |
| Neither model has a result on the other 18 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index10
| Benchmark | GPT-5.2-Codex | GPT-5.4 Mini | Leader |
|---|---|---|---|
| AA Intelligence Index | 28.5 | 24.1 | GPT-5.2-Codex |
| Epoch Capabilities Index | Not measured | 148.7 | Only GPT-5.4 Mini measured |
Agentic workLike-for-likePublic boards, shown as published, kept out of the index10
| Benchmark | GPT-5.2-Codex | GPT-5.4 Mini | Leader |
|---|---|---|---|
| AA Agentic Index | Not measured | 17.9 | Only GPT-5.4 Mini measured |
| Design Arena: agents building web apps | 1051 | Not measured | Only GPT-5.2-Codex measured |
| Design Arena: agents building full-stack apps | 993 | Not measured | Only GPT-5.2-Codex measured |
| Design Arena: agents building mobile apps | 1090 | Not measured | Only GPT-5.2-Codex measured |
| Design Arena: agents building Android apps | 1176 | Not measured | Only GPT-5.2-Codex measured |
| Design Arena: agents building Godot games | 1142 | Not measured | Only GPT-5.2-Codex measured |
| Finance Agent (Vals AI) | Not measured | 45.4% | Only GPT-5.4 Mini measured |
| Tax Agent Bench (Vals AI) | Not measured | 8.8% | Only GPT-5.4 Mini measured |
| Legal Research Bench (Vals AI) | Not measured | 12.5% | Only GPT-5.4 Mini measured |
| Excel Modeling Benchmark (Vals AI) | Not measured | 45.4% | Only GPT-5.4 Mini measured |
| τ²-Bench Telecom | 92.1% | 83.3% | GPT-5.2-Codex |
| DeepResearch Bench | Not measured | 36.3% | Only GPT-5.4 Mini measured |
| Neither model has a result on the other 29 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index20
| Benchmark | GPT-5.2-Codex | GPT-5.4 Mini | Leader |
|---|---|---|---|
| GPQA Diamond | Not measured | 83.1% | Only GPT-5.4 Mini measured |
| MMMU Pro (Vals AI) | Not measured | 79.3% | Only GPT-5.4 Mini measured |
| GPQA Diamond (Artificial Analysis run) | 89.9% | 87.5% | GPT-5.2-Codex |
| Humanity's Last Exam (Artificial Analysis run) | 35.7% | 28.1% | GPT-5.2-Codex |
| GPQA Diamond (Epoch AI run) | Not measured | 86.9% | Only GPT-5.4 Mini measured |
| ARC-AGI | Not measured | 63.7% | Only GPT-5.4 Mini measured |
| ARC-AGI-2 | Not measured | 18.9% | Only GPT-5.4 Mini measured |
| Chess puzzles | Not measured | 23.0% | Only GPT-5.4 Mini measured |
| Mystery game puzzles | Not measured | 11.0% | Only GPT-5.4 Mini measured |
| LMArena Text: hard prompts | Not measured | 1424 | Only GPT-5.4 Mini measured |
| LMArena Text: expert | Not measured | 1435 | Only GPT-5.4 Mini measured |
| LMArena Text: hard prompts in English | Not measured | 1420 | Only GPT-5.4 Mini measured |
| LMArena Vision: diagrams | Not measured | 1266 | Only GPT-5.4 Mini measured |
| LMArena Vision: homework | Not measured | 1285 | Only GPT-5.4 Mini measured |
| Kagi LLM Benchmark | Not measured | 37.9% | Only GPT-5.4 Mini measured |
| Neither model has a result on the other 7 boards here | |||
MathDirectional onlyPublic boards, shown as published, kept out of the index7 boards, none shared
| Benchmark | GPT-5.2-Codex | GPT-5.4 Mini | Leader |
|---|---|---|---|
| AIME | Not measured | 95.6% | Only GPT-5.4 Mini measured |
| OTIS Mock AIME | Not measured | 88.9% | Only GPT-5.4 Mini measured |
| FrontierMath, tiers 1 to 3 | Not measured | 51.2% | Only GPT-5.4 Mini measured |
| FrontierMath, tier 4 | Not measured | 9.8% | Only GPT-5.4 Mini measured |
| ProofBench | Not measured | 21.0% | Only GPT-5.4 Mini measured |
| LMArena Text: math | Not measured | 1418 | Only GPT-5.4 Mini measured |
| LMArena Text: mathematical work | Not measured | 1429 | Only GPT-5.4 Mini measured |
| Neither model has a result on the other 6 boards here | |||
KnowledgeDirectional onlyPublic boards, shown as published, kept out of the index10 boards, none shared
| Benchmark | GPT-5.2-Codex | GPT-5.4 Mini | Leader |
|---|---|---|---|
| MMLU Pro | Not measured | 84.5% | Only GPT-5.4 Mini measured |
| CaseLaw (Vals AI) | Not measured | 51.7% | Only GPT-5.4 Mini measured |
| CorpFin (Vals AI) | Not measured | 60.9% | Only GPT-5.4 Mini measured |
| TaxEval (Vals AI) | Not measured | 71.2% | Only GPT-5.4 Mini measured |
| MortgageTax (Vals AI) | Not measured | 63.5% | Only GPT-5.4 Mini measured |
| SimpleQA Verified | Not measured | 29.4% | Only GPT-5.4 Mini measured |
| LMArena Text: science | Not measured | 1417 | Only GPT-5.4 Mini measured |
| LMArena Text: medicine and healthcare | Not measured | 1395 | Only GPT-5.4 Mini measured |
| LMArena Text: legal and government | Not measured | 1414 | Only GPT-5.4 Mini measured |
| LMArena Text: business and finance | Not measured | 1418 | Only GPT-5.4 Mini measured |
| Neither model has a result on the other 8 boards here | |||
Web designNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | GPT-5.2-Codex | GPT-5.4 Mini | Leader |
|---|---|---|---|
| Neither model has a result on 9 boards here | |||
GeneralLike-for-likePublic boards, shown as published, kept out of the index20
| Benchmark | GPT-5.2-Codex | GPT-5.4 Mini | Leader |
|---|---|---|---|
| SAGE (Vals AI) | Not measured | 50.8% | Only GPT-5.4 Mini measured |
| Vals Multimodal Index | Not measured | 54.2% | Only GPT-5.4 Mini measured |
| IFBench | 77.6% | 73.3% | GPT-5.2-Codex |
| Long context reasoning (AA-LCR) | 82.3% | 77.0% | GPT-5.2-Codex |
| DTBench | Not measured | 80.0% | Only GPT-5.4 Mini measured |
| LMCA | Not measured | 40.8% | Only GPT-5.4 Mini measured |
| LMArena Text | Not measured | 1412 | Only GPT-5.4 Mini measured |
| LMArena Text: instruction following | Not measured | 1405 | Only GPT-5.4 Mini measured |
| LMArena Text: creative writing | Not measured | 1370 | Only GPT-5.4 Mini measured |
| LMArena Vision | Not measured | 1245 | Only GPT-5.4 Mini measured |
| LMArena Text: English | Not measured | 1412 | Only GPT-5.4 Mini measured |
| LMArena Text: non-English | Not measured | 1405 | Only GPT-5.4 Mini measured |
| LMArena Text: Chinese | Not measured | 1446 | Only GPT-5.4 Mini measured |
| LMArena Text: German | Not measured | 1408 | Only GPT-5.4 Mini measured |
| LMArena Text: Japanese | Not measured | 1375 | Only GPT-5.4 Mini measured |
| LMArena Text: Korean | Not measured | 1370 | Only GPT-5.4 Mini measured |
| LMArena Text: French | Not measured | 1441 | Only GPT-5.4 Mini measured |
| LMArena Text: Spanish | Not measured | 1405 | Only GPT-5.4 Mini measured |
| LMArena Text: Russian | Not measured | 1417 | Only GPT-5.4 Mini measured |
| LMArena Text: Polish | Not measured | 1415 | Only GPT-5.4 Mini measured |
| LMArena Text: multi-turn | Not measured | 1429 | Only GPT-5.4 Mini measured |
| LMArena Text: longer queries | Not measured | 1408 | Only GPT-5.4 Mini measured |
| LMArena Text: writing and language | Not measured | 1393 | Only GPT-5.4 Mini measured |
| LMArena Text: entertainment and media | Not measured | 1373 | Only GPT-5.4 Mini measured |
| LMArena Vision: English | Not measured | 1244 | Only GPT-5.4 Mini measured |
| LMArena Vision: Chinese | Not measured | 1286 | Only GPT-5.4 Mini measured |
| LMArena Vision: reading text | Not measured | 1253 | Only GPT-5.4 Mini measured |
| LMArena Vision: humor | Not measured | 1235 | Only GPT-5.4 Mini measured |
| LMArena Vision: recognizing things | Not measured | 1218 | Only GPT-5.4 Mini measured |
| LMArena Vision: creative writing | Not measured | 1222 | Only GPT-5.4 Mini measured |
| Neither model has a result on the other 16 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | GPT-5.2-Codex | GPT-5.4 Mini | 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-5.4 Mini costs less
Repository review
50K input + 3K output tokens
GPT-5.4 Mini costs less
Agent loop
200K input + 10K output tokens
GPT-5.4 Mini 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-5.2-Codex | OpenAI GPT-5.4 Mini |
|---|---|---|
| Maker | OpenAI | OpenAI |
| Weights | No public weights listed | No public weights listed |
| Released | January 14, 2026 | March 17, 2026 |
| Context window | 400K | 400K |
| Input $/Mtok | $1.75 | $0.75 |
| Output $/Mtok | $14.00 | $4.50 |
| Coding index | 45.4 | 35.8 |
| Independent measurements behind it | 3 | 5 |
| Cost per solved task | $0.43 | $0.70 |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is GPT-5.2-Codex better than GPT-5.4 Mini for coding?
Pick GPT-5.2-Codex for coding: it scores 45.4 on the Vibe Coding Index against 35.8 for GPT-5.4 Mini, though on 3 benchmarks against 5. Pick GPT-5.4 Mini for output tokens 3.1x cheaper ($4.50 vs $14.00 per million) and work measured by LMArena WebDev or SWE-bench Verified (Vals AI), where it leads.
Which is cheaper, GPT-5.2-Codex or GPT-5.4 Mini?
GPT-5.2-Codex costs $1.75 per million input tokens and $14.00 per million output tokens; GPT-5.4 Mini costs $0.75 and $4.50. Per solved SWE-bench Verified task, GPT-5.2-Codex costs $0.43 and GPT-5.4 Mini $0.70.
Which has the bigger context window, GPT-5.2-Codex or GPT-5.4 Mini?
Neither: both take 400K tokens.
How many benchmarks is the GPT-5.2-Codex vs GPT-5.4 Mini comparison based on?
14 benchmarks have a result for both models as of October 22, 2026. GPT-5.2-Codex's coding index rests on 3 independent benchmarks and GPT-5.4 Mini's on 5. Where only one model has a result, the page says so instead of naming a winner.