DeepSeek V3.2 vs DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 leads DeepSeek V3.2 on coding, 51.8 to 30.0 on the Vibe Coding Index, and DeepSeek V4 Flash 0731 costs 3.1x less per output token ($0.28 vs $0.88 per million). Updated October 21, 2026; 26 benchmarks have a result for both.
DeepSeek V3.2 or DeepSeek V4 Flash 0731: which to pick
Pick DeepSeek V4 Flash 0731 for coding: it scores 51.8 on the Vibe Coding Index against 30.0 for DeepSeek V3.2.
DeepSeek V3.2 is neither cheaper nor larger in context, and leads no benchmark both models were measured on.
- Shared results
- 26
- DeepSeek V3.2 only
- 49
- DeepSeek V4 Flash 0731 only
- 29
- Like-for-like categories
- 6 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
DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 scores 51.8 against 30.0.
3 and 4 benchmarks behind the scores
Intelligence
AA Intelligence Index
DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 leads, 34.3 against 21.5.
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
DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 leads, 89.9% against 80.3%.
Same board, both measured
Math
AIME
No call
Only DeepSeek V3.2 has a AIME result, 84.6%.
Directional only
Knowledge
MMLU Pro
DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 leads, 86.2% against 84.9%.
Same board, both measured
Web design
Design Arena: websites
DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 leads, 1238 against 1175.
Same board, both measured
Long documents
Listed context window
DeepSeek V4 Flash 0731
DeepSeek V4 Flash 0731 takes 1.05M tokens against 164K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
DeepSeek V4 Flash 0731
DeepSeek V3.2 $0.0007, DeepSeek V4 Flash 0731 $0.0003 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
DeepSeek V4 Flash 0731
DeepSeek V3.2 $0.99, DeepSeek V4 Flash 0731 $0.01.
Published attempt cost and pass rate
At a glance
- Coding index
- DeepSeek V3.2 30.0 from 3 benchmarks, DeepSeek V4 Flash 0731 51.8 from 4 benchmarks.
- Widest category
- Coding, 30.6 points apart on average over 4 shared benchmarks.
- Widest single gap
- Vibe Code Bench (Vals AI): DeepSeek V3.2 5.1% against DeepSeek V4 Flash 0731 74.7%, 69.6 points apart on a 0 to 100 scale.
- Output price
- DeepSeek V3.2 $0.88 against DeepSeek V4 Flash 0731 $0.28 per million output tokens, 3.1x apart.
- Context window
- DeepSeek V3.2 takes 164K tokens, DeepSeek V4 Flash 0731 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 | DeepSeek V3.2 | Bars | DeepSeek V4 Flash 0731 | Ties |
|---|---|---|---|---|
| Coding | 0 | 4 | ||
| Intelligence | 0 | 2 | ||
| Reasoning | 0 | 5 | ||
| Knowledge | 0 | 4 | ||
| Web design | 0 | 8 | ||
| General | 0 | 3 | ||
| All shared boards | 0 | 26 |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index04
| Benchmark | DeepSeek V3.2 | DeepSeek V4 Flash 0731 | Leader |
|---|---|---|---|
| SWE-bench Verified | 70.0% | Not measured | Only DeepSeek V3.2 measured |
| SWE-bench Multilingual | 59.0% | Not measured | Only DeepSeek V3.2 measured |
| LMArena WebDev | 1362 | Not measured | Only DeepSeek V3.2 measured |
| AA Coding Index | 44.2 | 69.1 | DeepSeek V4 Flash 0731 |
| LiveCodeBench (Vals AI) | 80.7% | 87.3% | DeepSeek V4 Flash 0731 |
| SWE-bench Verified (Vals AI) | 67.6% | 88.8% | DeepSeek V4 Flash 0731 |
| Terminal-Bench 4 (Vals AI) | Not measured | 18.7% | Only DeepSeek V4 Flash 0731 measured |
| Vibe Code Bench (Vals AI) | 5.1% | 74.7% | DeepSeek V4 Flash 0731 |
| Code migration (Vals AI) | Not measured | 38.6% | Only DeepSeek V4 Flash 0731 measured |
| ProgramBench (Vals AI) | Not measured | 0.0% | Only DeepSeek V4 Flash 0731 measured |
| IOI (Vals AI) | Not measured | 32.7% | Only DeepSeek V4 Flash 0731 measured |
| Terminal-Bench 2.1 (Vals AI) | Not measured | 67.0% | Only DeepSeek V4 Flash 0731 measured |
| SkillsBench (Vals AI) | Not measured | 50.7% | Only DeepSeek V4 Flash 0731 measured |
| CyberBench (Vals AI) | Not measured | 72.1% | Only DeepSeek V4 Flash 0731 measured |
| LiveCodeBench (Artificial Analysis run) | 86.2% | Not measured | Only DeepSeek V3.2 measured |
| SciCode | Not measured | 50.3% | Only DeepSeek V4 Flash 0731 measured |
| Terminal-Bench Hard (Artificial Analysis run) | 35.6% | Not measured | Only DeepSeek V3.2 measured |
| Terminal-Bench | 39.6% | Not measured | Only DeepSeek V3.2 measured |
| FrontierCode | Not measured | 18.8% | Only DeepSeek V4 Flash 0731 measured |
| WeirdML | Not measured | 63.0% | Only DeepSeek V4 Flash 0731 measured |
| ALE-Bench | Not measured | 1306.1 | Only DeepSeek V4 Flash 0731 measured |
| LMArena Text: coding | 1453 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: software and IT | 1449 | Not measured | Only DeepSeek V3.2 measured |
| LMArena WebDev: React | 1350 | Not measured | Only DeepSeek V3.2 measured |
| LMArena WebDev: HTML | 1371 | Not measured | Only DeepSeek V3.2 measured |
| Neither model has a result on the other 14 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | DeepSeek V3.2 | DeepSeek V4 Flash 0731 | Leader |
|---|---|---|---|
| AA Intelligence Index | 21.5 | 34.3 | DeepSeek V4 Flash 0731 |
| Epoch Capabilities Index | 146.1 | 154.8 | DeepSeek V4 Flash 0731 |
Agentic workDirectional onlyPublic boards, shown as published, kept out of the index10 boards, none shared
| Benchmark | DeepSeek V3.2 | DeepSeek V4 Flash 0731 | Leader |
|---|---|---|---|
| AA Agentic Index | Not measured | 41.0 | Only DeepSeek V4 Flash 0731 measured |
| Finance Agent (Vals AI) | Not measured | 49.5% | Only DeepSeek V4 Flash 0731 measured |
| Tax Agent Bench (Vals AI) | Not measured | 28.1% | Only DeepSeek V4 Flash 0731 measured |
| Legal Research Bench (Vals AI) | Not measured | 30.3% | Only DeepSeek V4 Flash 0731 measured |
| Excel Modeling Benchmark (Vals AI) | Not measured | 57.0% | Only DeepSeek V4 Flash 0731 measured |
| BioMysteryBench (Vals AI) | Not measured | 64.4% | Only DeepSeek V4 Flash 0731 measured |
| Terminal-Bench-Science (Vals AI) | Not measured | 4.3% | Only DeepSeek V4 Flash 0731 measured |
| τ²-Bench Telecom | 90.6% | Not measured | Only DeepSeek V3.2 measured |
| APEX-Agents | 21.3% | Not measured | Only DeepSeek V3.2 measured |
| Vending-Bench 2 | $1,034 | Not measured | Only DeepSeek V3.2 measured |
| Neither model has a result on the other 31 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index05
| Benchmark | DeepSeek V3.2 | DeepSeek V4 Flash 0731 | Leader |
|---|---|---|---|
| GPQA Diamond | 80.3% | 89.9% | DeepSeek V4 Flash 0731 |
| GPQA Diamond (Artificial Analysis run) | 84.0% | 90.8% | DeepSeek V4 Flash 0731 |
| Humanity's Last Exam (Artificial Analysis run) | 24.6% | 38.6% | DeepSeek V4 Flash 0731 |
| GPQA Diamond (Epoch AI run) | Not measured | 91.0% | Only DeepSeek V4 Flash 0731 measured |
| ARC-AGI | 57.0% | 89.0% | DeepSeek V4 Flash 0731 |
| ARC-AGI-2 | 4.0% | 61.4% | DeepSeek V4 Flash 0731 |
| SimpleBench | Not measured | 61.1% | Only DeepSeek V4 Flash 0731 measured |
| Chess puzzles | Not measured | 32.0% | Only DeepSeek V4 Flash 0731 measured |
| Mystery game puzzles | Not measured | 34.0% | Only DeepSeek V4 Flash 0731 measured |
| LMArena Text: hard prompts | 1434 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: expert | 1436 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: hard prompts in English | 1446 | Not measured | Only DeepSeek V3.2 measured |
| Neither model has a result on the other 10 boards here | |||
MathDirectional onlyPublic boards, shown as published, kept out of the index10 boards, none shared
| Benchmark | DeepSeek V3.2 | DeepSeek V4 Flash 0731 | Leader |
|---|---|---|---|
| AIME | 84.6% | Not measured | Only DeepSeek V3.2 measured |
| MGSM (Vals AI) | 86.0% | Not measured | Only DeepSeek V3.2 measured |
| AA Math Index | 92.0 | Not measured | Only DeepSeek V3.2 measured |
| AIME 2025 (Artificial Analysis run) | 92.0% | Not measured | Only DeepSeek V3.2 measured |
| OTIS Mock AIME | Not measured | 94.4% | Only DeepSeek V4 Flash 0731 measured |
| FrontierMath, tiers 1 to 3 | Not measured | 57.5% | Only DeepSeek V4 Flash 0731 measured |
| FrontierMath, tier 4 | Not measured | 24.4% | Only DeepSeek V4 Flash 0731 measured |
| ProofBench | Not measured | 56.0% | Only DeepSeek V4 Flash 0731 measured |
| LMArena Text: math | 1435 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: mathematical work | 1440 | Not measured | Only DeepSeek V3.2 measured |
| Neither model has a result on the other 3 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index04
| Benchmark | DeepSeek V3.2 | DeepSeek V4 Flash 0731 | Leader |
|---|---|---|---|
| MMLU Pro | 84.9% | 86.2% | DeepSeek V4 Flash 0731 |
| LegalBench (Vals AI) | 76.1% | 77.7% | DeepSeek V4 Flash 0731 |
| CaseLaw (Vals AI) | 55.4% | Not measured | Only DeepSeek V3.2 measured |
| CorpFin (Vals AI) | 51.0% | 61.9% | DeepSeek V4 Flash 0731 |
| TaxEval (Vals AI) | 68.2% | 70.7% | DeepSeek V4 Flash 0731 |
| MedQA (Vals AI) | 93.9% | Not measured | Only DeepSeek V3.2 measured |
| MedCode (Vals AI) | Not measured | 41.4% | Only DeepSeek V4 Flash 0731 measured |
| MMLU Pro (Artificial Analysis run) | 86.2% | Not measured | Only DeepSeek V3.2 measured |
| SimpleQA Verified | Not measured | 33.6% | Only DeepSeek V4 Flash 0731 measured |
| LMArena Text: science | 1446 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: medicine and healthcare | 1440 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: legal and government | 1425 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: business and finance | 1410 | Not measured | Only DeepSeek V3.2 measured |
| Neither model has a result on the other 5 boards here | |||
Web designLike-for-likePublic boards, shown as published, kept out of the index08
| Benchmark | DeepSeek V3.2 | DeepSeek V4 Flash 0731 | Leader |
|---|---|---|---|
| Design Arena: websites | 1175 | 1238 | DeepSeek V4 Flash 0731 |
| Design Arena: UI components | 1149 | 1235 | DeepSeek V4 Flash 0731 |
| Design Arena: game dev | 1132 | 1208 | DeepSeek V4 Flash 0731 |
| Design Arena: data viz | 1158 | 1180 | DeepSeek V4 Flash 0731 |
| Design Arena: 3D | 1133 | 1209 | DeepSeek V4 Flash 0731 |
| Design Arena: all code categories | 1164 | 1230 | DeepSeek V4 Flash 0731 |
| Design Arena: SVG | 1032 | 1185 | DeepSeek V4 Flash 0731 |
| Design Arena: ASCII art | 1083 | 1090 | DeepSeek V4 Flash 0731 |
| 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 V3.2 | DeepSeek V4 Flash 0731 | Leader |
|---|---|---|---|
| MedScribe (Vals AI) | Not measured | 80.4% | Only DeepSeek V4 Flash 0731 measured |
| IFBench | 60.7% | Not measured | Only DeepSeek V3.2 measured |
| Long context reasoning (AA-LCR) | 73.3% | 79.7% | DeepSeek V4 Flash 0731 |
| DTBench | 85.6% | 90.9% | DeepSeek V4 Flash 0731 |
| LMCA | 28.8% | 41.7% | DeepSeek V4 Flash 0731 |
| CL-bench | 12.4% | Not measured | Only DeepSeek V3.2 measured |
| CL-bench Life | 7.4% | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text | 1425 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: instruction following | 1413 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: creative writing | 1400 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: English | 1439 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: non-English | 1409 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: Chinese | 1455 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: German | 1413 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: Japanese | 1374 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: Korean | 1370 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: French | 1429 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: Spanish | 1424 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: Russian | 1424 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: Polish | 1419 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: multi-turn | 1427 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: longer queries | 1428 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: writing and language | 1407 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: entertainment and media | 1395 | Not measured | Only DeepSeek V3.2 measured |
| Neither model has a result on the other 22 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | DeepSeek V3.2 | DeepSeek V4 Flash 0731 | 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 V3.2 | DeepSeek V4 Flash 0731 |
|---|---|---|
| Maker | DeepSeek | DeepSeek |
| Weights | Open weights | Open weights |
| Released | December 1, 2025 | July 31, 2026 |
| Context window | 164K | 1.05M |
| Input $/Mtok | $0.29 | $0.14 |
| Output $/Mtok | $0.88 | $0.28 |
| Coding index | 30.0 | 51.8 |
| Independent measurements behind it | 3 | 4 |
| Cost per solved task | $0.99 | $0.01 |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is DeepSeek V3.2 better than DeepSeek V4 Flash 0731 for coding?
Pick DeepSeek V4 Flash 0731 for coding: it scores 51.8 on the Vibe Coding Index against 30.0 for DeepSeek V3.2. DeepSeek V3.2 is neither cheaper nor larger in context, and leads no benchmark both models were measured on.
Which is cheaper, DeepSeek V3.2 or DeepSeek V4 Flash 0731?
DeepSeek V3.2 costs $0.29 per million input tokens and $0.88 per million output tokens; DeepSeek V4 Flash 0731 costs $0.14 and $0.28. Per solved SWE-bench Verified task, DeepSeek V3.2 costs $0.99 and DeepSeek V4 Flash 0731 $0.01.
Which has the bigger context window, DeepSeek V3.2 or DeepSeek V4 Flash 0731?
DeepSeek V4 Flash 0731, with 1.05M tokens against 164K.
How many benchmarks is the DeepSeek V3.2 vs DeepSeek V4 Flash 0731 comparison based on?
26 benchmarks have a result for both models as of October 21, 2026. DeepSeek V3.2's coding index rests on 3 independent benchmarks and DeepSeek V4 Flash 0731's on 4. Where only one model has a result, the page says so instead of naming a winner.