DeepSeek V4 Flash 0731 vs GPT-5.2
GPT-5.2 leads DeepSeek V4 Flash 0731 on coding, 55.9 to 51.8 on the Vibe Coding Index, and DeepSeek V4 Flash 0731 costs 50.0x less per output token ($0.28 vs $14.00 per million). Updated October 21, 2026; 38 benchmarks have a result for both.
DeepSeek V4 Flash 0731 or GPT-5.2: which to pick
Pick GPT-5.2 for coding: it scores 55.9 on the Vibe Coding Index against 51.8 for DeepSeek V4 Flash 0731, though on 3 benchmarks against 4.
Pick DeepSeek V4 Flash 0731 for output tokens 50.0x cheaper ($0.28 vs $14.00 per million) and a lower cost per solved task ($0.01 against $1.91).
- Shared results
- 38
- DeepSeek V4 Flash 0731 only
- 17
- GPT-5.2 only
- 76
- Like-for-like categories
- 7 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
GPT-5.2 scores 55.9 against 51.8.
4 and 3 benchmarks behind the scores
Intelligence
AA Intelligence Index
DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 leads, 34.3 against 30.4.
Same board, both measured
Agentic work
AA Agentic Index
No call
Only DeepSeek V4 Flash 0731 has a AA Agentic Index result, 41.0.
Directional only
Reasoning
GPQA Diamond
GPT-5.2
GPT-5.2 leads, 91.7% against 89.9%.
Same board, both measured
Math
AIME
No call
Only GPT-5.2 has a AIME result, 96.9%.
Directional only
Knowledge
MMLU Pro
GPT-5.2
GPT-5.2 leads, 86.2% against 86.2%.
Same board, both measured
Web design
Design Arena: websites
DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 leads, 1238 against 1195.
Same board, both measured
Long documents
Listed context window
DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 takes 1.05M tokens against 400K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 $0.0003, GPT-5.2 $0.0088 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 $0.01, GPT-5.2 $1.91.
Published attempt cost and pass rate
At a glance
- Coding index
- DeepSeek V4 Flash 0731 51.8 from 4 benchmarks, GPT-5.2 55.9 from 3 benchmarks.
- Widest category
- Math, 15.0 points apart on average over 4 shared benchmarks.
- Widest single gap
- ProofBench: DeepSeek V4 Flash 0731 56.0% against GPT-5.2 15.0%, 41.0 points apart on a 0 to 100 scale.
- Output price
- DeepSeek V4 Flash 0731 $0.28 against GPT-5.2 $14.00 per million output tokens, 50.0x apart.
- Context window
- DeepSeek V4 Flash 0731 takes 1.05M tokens, GPT-5.2 400K.
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 | DeepSeek V4 Flash 0731 | Bars | GPT-5.2 | Ties |
|---|---|---|---|---|
| Coding | 4 | 1 | ||
| Intelligence | 2 | 0 | ||
| Reasoning | 6 | 3 | ||
| Math | 1 | 3 | ||
| Knowledge | 0 | 6 | ||
| Web design | 6 | 2 | ||
| General | 0 | 3 | 1 tie | |
| All shared boards | 19 | 18 | 1 tie |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index41
| Benchmark | DeepSeek V4 Flash 0731 | GPT-5.2 | Leader |
|---|---|---|---|
| SWE-bench Verified | Not measured | 72.8% | Only GPT-5.2 measured |
| SWE-bench Multilingual | Not measured | 66.7% | Only GPT-5.2 measured |
| LMArena WebDev | Not measured | 1417 | Only GPT-5.2 measured |
| AA Coding Index | 69.1 | Not measured | Only DeepSeek V4 Flash 0731 measured |
| LiveCodeBench (Vals AI) | 87.3% | 85.4% | DeepSeek V4 Flash 0731 |
| SWE-bench Verified (Vals AI) | 88.8% | 75.8% | DeepSeek V4 Flash 0731 |
| Terminal-Bench 4 (Vals AI) | 18.7% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| Vibe Code Bench (Vals AI) | 74.7% | 53.5% | DeepSeek V4 Flash 0731 |
| Code migration (Vals AI) | 38.6% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| ProgramBench (Vals AI) | 0.0% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| IOI (Vals AI) | 32.7% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| Terminal-Bench 2.1 (Vals AI) | 67.0% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| SkillsBench (Vals AI) | 50.7% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| CyberBench (Vals AI) | 72.1% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| LiveCodeBench (Artificial Analysis run) | Not measured | 89.4% | Only GPT-5.2 measured |
| SciCode | 50.3% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| Terminal-Bench Hard (Artificial Analysis run) | Not measured | 47.0% | Only GPT-5.2 measured |
| Terminal-Bench | Not measured | 64.9% | Only GPT-5.2 measured |
| FrontierCode | 18.8% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| GSO-Bench | Not measured | 27.5% | Only GPT-5.2 measured |
| WeirdML | 63.0% | 72.2% | GPT-5.2 |
| ALE-Bench | 1306.1 | 1293.5 | DeepSeek V4 Flash 0731 |
| LMArena Text: coding | Not measured | 1442 | Only GPT-5.2 measured |
| LMArena Text: software and IT | Not measured | 1434 | Only GPT-5.2 measured |
| LMArena WebDev: HTML | Not measured | 1427 | Only GPT-5.2 measured |
| Neither model has a result on the other 14 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index20
| Benchmark | DeepSeek V4 Flash 0731 | GPT-5.2 | Leader |
|---|---|---|---|
| AA Intelligence Index | 34.3 | 30.4 | DeepSeek V4 Flash 0731 |
| Epoch Capabilities Index | 154.8 | 153.7 | DeepSeek V4 Flash 0731 |
Agentic workDirectional onlyPublic boards, shown as published, kept out of the index19 boards, none shared
| Benchmark | DeepSeek V4 Flash 0731 | GPT-5.2 | Leader |
|---|---|---|---|
| AA Agentic Index | 41.0 | Not measured | Only DeepSeek V4 Flash 0731 measured |
| Design Arena: agents building web apps | Not measured | 1086 | Only GPT-5.2 measured |
| Design Arena: agents building full-stack apps | Not measured | 1043 | Only GPT-5.2 measured |
| Design Arena: agents building mobile apps | Not measured | 1093 | Only GPT-5.2 measured |
| Design Arena: agents building Android apps | Not measured | 1058 | Only GPT-5.2 measured |
| Design Arena: agents building Godot games | Not measured | 1142 | Only GPT-5.2 measured |
| Finance Agent (Vals AI) | 49.5% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| Tax Agent Bench (Vals AI) | 28.1% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| Legal Research Bench (Vals AI) | 30.3% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| Excel Modeling Benchmark (Vals AI) | 57.0% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| BioMysteryBench (Vals AI) | 64.4% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| Terminal-Bench-Science (Vals AI) | 4.3% | Not measured | Only DeepSeek V4 Flash 0731 measured |
| τ²-Bench Telecom | Not measured | 84.8% | Only GPT-5.2 measured |
| METR time horizons | Not measured | 75.3% | Only GPT-5.2 measured |
| Remote Labor Index | Not measured | 2.5% | Only GPT-5.2 measured |
| GDPval | Not measured | 49.7% | Only GPT-5.2 measured |
| DeepResearch Bench | Not measured | 41.1% | Only GPT-5.2 measured |
| Vending-Bench 2 | Not measured | $3,591 | Only GPT-5.2 measured |
| Berkeley Function Calling Leaderboard (BFCL) | Not measured | 55.9% | Only GPT-5.2 measured |
| Neither model has a result on the other 22 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index63
| Benchmark | DeepSeek V4 Flash 0731 | GPT-5.2 | Leader |
|---|---|---|---|
| GPQA Diamond | 89.9% | 91.7% | GPT-5.2 |
| MMMU Pro (Vals AI) | Not measured | 86.7% | Only GPT-5.2 measured |
| GPQA Diamond (Artificial Analysis run) | 90.8% | 90.3% | DeepSeek V4 Flash 0731 |
| Humanity's Last Exam (Artificial Analysis run) | 38.6% | 37.7% | DeepSeek V4 Flash 0731 |
| GPQA Diamond (Epoch AI run) | 91.0% | 91.4% | GPT-5.2 |
| ARC-AGI | 89.0% | 86.2% | DeepSeek V4 Flash 0731 |
| ARC-AGI-2 | 61.4% | 52.9% | DeepSeek V4 Flash 0731 |
| SimpleBench | 61.1% | 45.8% | DeepSeek V4 Flash 0731 |
| Chess puzzles | 32.0% | 48.0% | GPT-5.2 |
| Mystery game puzzles | 34.0% | 23.0% | DeepSeek V4 Flash 0731 |
| EBR-bench | Not measured | 23.0% | Only GPT-5.2 measured |
| VPCT | Not measured | 84.0% | Only GPT-5.2 measured |
| LMArena Text: hard prompts | Not measured | 1428 | Only GPT-5.2 measured |
| LMArena Text: expert | Not measured | 1445 | Only GPT-5.2 measured |
| LMArena Text: hard prompts in English | Not measured | 1431 | Only GPT-5.2 measured |
| LMArena Vision: diagrams | Not measured | 1265 | Only GPT-5.2 measured |
| LMArena Vision: homework | Not measured | 1276 | Only GPT-5.2 measured |
| Kagi LLM Benchmark | Not measured | 73.3% | Only GPT-5.2 measured |
| Neither model has a result on the other 4 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index13
| Benchmark | DeepSeek V4 Flash 0731 | GPT-5.2 | Leader |
|---|---|---|---|
| AIME | Not measured | 96.9% | Only GPT-5.2 measured |
| MGSM (Vals AI) | Not measured | 94.0% | Only GPT-5.2 measured |
| AA Math Index | Not measured | 99.0 | Only GPT-5.2 measured |
| AIME 2025 (Artificial Analysis run) | Not measured | 99.0% | Only GPT-5.2 measured |
| OTIS Mock AIME | 94.4% | 96.1% | GPT-5.2 |
| FrontierMath, tiers 1 to 3 | 57.5% | 67.4% | GPT-5.2 |
| FrontierMath, tier 4 | 24.4% | 31.7% | GPT-5.2 |
| ProofBench | 56.0% | 15.0% | DeepSeek V4 Flash 0731 |
| LMArena Text: math | Not measured | 1440 | Only GPT-5.2 measured |
| LMArena Text: mathematical work | Not measured | 1442 | Only GPT-5.2 measured |
| Neither model has a result on the other 3 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index06
| Benchmark | DeepSeek V4 Flash 0731 | GPT-5.2 | Leader |
|---|---|---|---|
| MMLU Pro | 86.2% | 86.2% | GPT-5.2 |
| LegalBench (Vals AI) | 77.7% | 82.8% | GPT-5.2 |
| CaseLaw (Vals AI) | Not measured | 66.0% | Only GPT-5.2 measured |
| CorpFin (Vals AI) | 61.9% | 65.9% | GPT-5.2 |
| TaxEval (Vals AI) | 70.7% | 75.8% | GPT-5.2 |
| MortgageTax (Vals AI) | Not measured | 67.1% | Only GPT-5.2 measured |
| MedQA (Vals AI) | Not measured | 94.1% | Only GPT-5.2 measured |
| MedCode (Vals AI) | 41.4% | 49.8% | GPT-5.2 |
| MMLU Pro (Artificial Analysis run) | Not measured | 87.4% | Only GPT-5.2 measured |
| SimpleQA Verified | 33.6% | 37.1% | GPT-5.2 |
| LMArena Text: science | Not measured | 1425 | Only GPT-5.2 measured |
| LMArena Text: medicine and healthcare | Not measured | 1425 | Only GPT-5.2 measured |
| LMArena Text: legal and government | Not measured | 1427 | Only GPT-5.2 measured |
| LMArena Text: business and finance | Not measured | 1414 | Only GPT-5.2 measured |
| FACTS Benchmark Suite | Not measured | 48.5% | Only GPT-5.2 measured |
| Neither model has a result on the other 3 boards here | |||
Web designLike-for-likePublic boards, shown as published, kept out of the index62
| Benchmark | DeepSeek V4 Flash 0731 | GPT-5.2 | Leader |
|---|---|---|---|
| Design Arena: websites | 1238 | 1195 | DeepSeek V4 Flash 0731 |
| Design Arena: UI components | 1235 | 1190 | DeepSeek V4 Flash 0731 |
| Design Arena: game dev | 1208 | 1195 | DeepSeek V4 Flash 0731 |
| Design Arena: data viz | 1180 | 1202 | GPT-5.2 |
| Design Arena: 3D | 1209 | 1079 | DeepSeek V4 Flash 0731 |
| Design Arena: all code categories | 1230 | 1170 | DeepSeek V4 Flash 0731 |
| Design Arena: SVG | 1185 | 1139 | DeepSeek V4 Flash 0731 |
| Design Arena: ASCII art | 1090 | 1152 | GPT-5.2 |
| Neither model has a result on the other 1 board here | |||
GeneralLike-for-likePublic boards, shown as published, kept out of the index03
| Benchmark | DeepSeek V4 Flash 0731 | GPT-5.2 | Leader |
|---|---|---|---|
| SAGE (Vals AI) | Not measured | 49.3% | Only GPT-5.2 measured |
| MedScribe (Vals AI) | 80.4% | 84.4% | GPT-5.2 |
| IFBench | Not measured | 75.4% | Only GPT-5.2 measured |
| Long context reasoning (AA-LCR) | 79.7% | 82.7% | GPT-5.2 |
| DTBench | 90.9% | 90.9% | Tie |
| LMCA | 41.7% | 43.9% | GPT-5.2 |
| CL-bench | Not measured | 18.2% | Only GPT-5.2 measured |
| Furniture assembly | Not measured | 38.3% | Only GPT-5.2 measured |
| LMArena Text | Not measured | 1416 | Only GPT-5.2 measured |
| LMArena Text: instruction following | Not measured | 1409 | Only GPT-5.2 measured |
| LMArena Text: creative writing | Not measured | 1376 | Only GPT-5.2 measured |
| LMArena Vision | Not measured | 1246 | Only GPT-5.2 measured |
| LMArena Text: English | Not measured | 1422 | Only GPT-5.2 measured |
| LMArena Text: non-English | Not measured | 1405 | Only GPT-5.2 measured |
| LMArena Text: Chinese | Not measured | 1447 | Only GPT-5.2 measured |
| LMArena Text: German | Not measured | 1444 | Only GPT-5.2 measured |
| LMArena Text: Japanese | Not measured | 1420 | Only GPT-5.2 measured |
| LMArena Text: Korean | Not measured | 1374 | Only GPT-5.2 measured |
| LMArena Text: French | Not measured | 1454 | Only GPT-5.2 measured |
| LMArena Text: Spanish | Not measured | 1413 | Only GPT-5.2 measured |
| LMArena Text: Russian | Not measured | 1415 | Only GPT-5.2 measured |
| LMArena Text: Polish | Not measured | 1426 | Only GPT-5.2 measured |
| LMArena Text: multi-turn | Not measured | 1419 | Only GPT-5.2 measured |
| LMArena Text: longer queries | Not measured | 1413 | Only GPT-5.2 measured |
| LMArena Text: writing and language | Not measured | 1392 | Only GPT-5.2 measured |
| LMArena Text: entertainment and media | Not measured | 1378 | Only GPT-5.2 measured |
| LMArena Vision: English | Not measured | 1243 | Only GPT-5.2 measured |
| LMArena Vision: Chinese | Not measured | 1305 | Only GPT-5.2 measured |
| LMArena Vision: reading text | Not measured | 1257 | Only GPT-5.2 measured |
| LMArena Vision: humor | Not measured | 1227 | Only GPT-5.2 measured |
| LMArena Vision: captioning | Not measured | 1267 | Only GPT-5.2 measured |
| LMArena Vision: recognizing things | Not measured | 1232 | Only GPT-5.2 measured |
| LMArena Vision: creative writing | Not measured | 1225 | Only GPT-5.2 measured |
| LMArena Document | Not measured | 1405 | Only GPT-5.2 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 | DeepSeek V4 Flash 0731 | GPT-5.2 | 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
DeepSeek V4 Flash 0731 costs less
Repository review
50K input + 3K output tokens
DeepSeek V4 Flash 0731 costs less
Agent loop
200K input + 10K output tokens
DeepSeek V4 Flash 0731 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 | DeepSeek V4 Flash 0731 | OpenAI GPT-5.2 |
|---|---|---|
| Maker | DeepSeek | OpenAI |
| Weights | Open weights | No public weights listed |
| Released | July 31, 2026 | December 10, 2025 |
| Context window | 1.05M | 400K |
| Input $/Mtok | $0.14 | $1.75 |
| Output $/Mtok | $0.28 | $14.00 |
| Coding index | 51.8 | 55.9 |
| Independent measurements behind it | 4 | 3 |
| Cost per solved task | $0.01 | $1.91 |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is DeepSeek V4 Flash 0731 better than GPT-5.2 for coding?
Pick GPT-5.2 for coding: it scores 55.9 on the Vibe Coding Index against 51.8 for DeepSeek V4 Flash 0731, though on 3 benchmarks against 4. Pick DeepSeek V4 Flash 0731 for output tokens 50.0x cheaper ($0.28 vs $14.00 per million) and a lower cost per solved task ($0.01 against $1.91).
Which is cheaper, DeepSeek V4 Flash 0731 or GPT-5.2?
DeepSeek V4 Flash 0731 costs $0.14 per million input tokens and $0.28 per million output tokens; GPT-5.2 costs $1.75 and $14.00. Per solved SWE-bench Verified task, DeepSeek V4 Flash 0731 costs $0.01 and GPT-5.2 $1.91.
Which has the bigger context window, DeepSeek V4 Flash 0731 or GPT-5.2?
DeepSeek V4 Flash 0731, with 1.05M tokens against 400K.
How many benchmarks is the DeepSeek V4 Flash 0731 vs GPT-5.2 comparison based on?
38 benchmarks have a result for both models as of October 21, 2026. DeepSeek V4 Flash 0731's coding index rests on 4 independent benchmarks and GPT-5.2's on 3. Where only one model has a result, the page says so instead of naming a winner.