DeepSeek V3.2 vs Step 5 Preview
Step 5 Preview leads DeepSeek V3.2 on coding, 61.1 to 30.0 on the Vibe Coding Index, and DeepSeek V3.2 costs 3.1x less per output token ($0.88 vs $2.70 per million). Updated October 22, 2026; 32 benchmarks have a result for both.
DeepSeek V3.2 or Step 5 Preview: which to pick
Pick Step 5 Preview for coding: it scores 61.1 on the Vibe Coding Index against 30.0 for DeepSeek V3.2.
Pick DeepSeek V3.2 for output tokens 3.1x cheaper ($0.88 vs $2.70 per million) and work measured by LMArena Text: medicine and healthcare, where it leads.
- Shared results
- 32
- DeepSeek V3.2 only
- 43
- Step 5 Preview 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
Step 5 Preview
Step 5 Preview scores 61.1 against 30.0.
3 and 4 benchmarks behind the scores
Intelligence
AA Intelligence Index
Step 5 Preview
Step 5 Preview leads, 43.7 against 21.5.
Same board, both measured
Reasoning
GPQA Diamond
No call
Only DeepSeek V3.2 has a GPQA Diamond result, 80.3%.
Directional only
Math
AIME
No call
Only DeepSeek V3.2 has a AIME result, 84.6%.
Directional only
Knowledge
MMLU Pro
No call
Only DeepSeek V3.2 has a MMLU Pro result, 84.9%.
Directional only
Web design
Design Arena: websites
No call
Only DeepSeek V3.2 has a Design Arena: websites result, 1175.
Directional only
Long documents
Listed context window
Step 5 Preview
Step 5 Preview takes 1.05M tokens against 164K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
DeepSeek V3.2
DeepSeek V3.2 $0.0007, Step 5 Preview $0.0024 per turn.
List prices
Cost per solved task
One SWE-bench Verified task, solved
No call
Only DeepSeek V3.2 has the number, $0.99.
Directional only
At a glance
- Coding index
- DeepSeek V3.2 30.0 from 3 benchmarks, Step 5 Preview 61.1 from 4 benchmarks.
- Widest category
- Coding, 28.9 points apart on average over 6 shared benchmarks.
- Widest single gap
- Vibe Code Bench (Vals AI): DeepSeek V3.2 5.1% against Step 5 Preview 70.4%, 65.3 points apart on a 0 to 100 scale.
- Output price
- DeepSeek V3.2 $0.88 against Step 5 Preview $2.70 per million output tokens, 3.1x apart.
- Context window
- DeepSeek V3.2 takes 164K tokens, Step 5 Preview 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 | Step 5 Preview | Ties |
|---|---|---|---|---|
| Coding | 0 | 6 | ||
| Intelligence | 0 | 1 | ||
| Reasoning | 0 | 4 | ||
| Math | 0 | 2 | ||
| Knowledge | 1 | 3 | ||
| General | 0 | 15 | ||
| All shared boards | 1 | 31 |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index06
| Benchmark | DeepSeek V3.2 | Step 5 Preview | 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 |
| LiveBench coding | Not measured | 85.6% | Only Step 5 Preview measured |
| LMArena WebDev | 1362 | 1567 | Step 5 Preview |
| AA Coding Index | 44.2 | Not measured | Only DeepSeek V3.2 measured |
| LiveCodeBench (Vals AI) | 80.7% | Not measured | Only DeepSeek V3.2 measured |
| SWE-bench Verified (Vals AI) | 67.6% | Not measured | Only DeepSeek V3.2 measured |
| Terminal-Bench 4 (Vals AI) | Not measured | 31.8% | Only Step 5 Preview measured |
| Vibe Code Bench (Vals AI) | 5.1% | 70.4% | Step 5 Preview |
| Code migration (Vals AI) | Not measured | 39.0% | Only Step 5 Preview measured |
| IOI (Vals AI) | Not measured | 38.1% | Only Step 5 Preview measured |
| SWE-bench Multimodal (Vals AI) | Not measured | 40.8% | Only Step 5 Preview measured |
| LiveCodeBench (Artificial Analysis run) | 86.2% | Not measured | Only DeepSeek V3.2 measured |
| SciCode | Not measured | 58.9% | Only Step 5 Preview 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 |
| LMArena Text: coding | 1453 | 1485 | Step 5 Preview |
| LMArena Text: software and IT | 1449 | 1473 | Step 5 Preview |
| LMArena WebDev: React | 1350 | 1583 | Step 5 Preview |
| LMArena WebDev: HTML | 1371 | 1556 | Step 5 Preview |
| LMArena WebDev: from an image | Not measured | 1550 | Only Step 5 Preview measured |
| Neither model has a result on the other 18 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index01
| Benchmark | DeepSeek V3.2 | Step 5 Preview | Leader |
|---|---|---|---|
| AA Intelligence Index | 21.5 | 43.7 | Step 5 Preview |
| Epoch Capabilities Index | 146.1 | Not measured | Only DeepSeek V3.2 measured |
Agentic workDirectional onlyPublic boards, shown as published, kept out of the index14 boards, none shared
| Benchmark | DeepSeek V3.2 | Step 5 Preview | Leader |
|---|---|---|---|
| LiveBench research | Not measured | 72.8% | Only Step 5 Preview measured |
| LiveBench browser | Not measured | 24.0% | Only Step 5 Preview measured |
| LiveBench long agentic | Not measured | 27.5% | Only Step 5 Preview measured |
| SREBench (Vals AI) | Not measured | 0.4% | Only Step 5 Preview measured |
| Finance Agent (Vals AI) | Not measured | 50.7% | Only Step 5 Preview measured |
| Tax Agent Bench (Vals AI) | Not measured | 27.3% | Only Step 5 Preview measured |
| Legal Research Bench (Vals AI) | Not measured | 40.4% | Only Step 5 Preview measured |
| Excel Modeling Benchmark (Vals AI) | Not measured | 57.1% | Only Step 5 Preview measured |
| BioMysteryBench (Vals AI) | Not measured | 67.4% | Only Step 5 Preview measured |
| Terminal-Bench-Science (Vals AI) | Not measured | 2.9% | Only Step 5 Preview measured |
| Vals Science Index | Not measured | 29.6% | Only Step 5 Preview 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 27 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index04
| Benchmark | DeepSeek V3.2 | Step 5 Preview | Leader |
|---|---|---|---|
| GPQA Diamond | 80.3% | Not measured | Only DeepSeek V3.2 measured |
| LiveBench reasoning | Not measured | 83.8% | Only Step 5 Preview measured |
| MysteryMechanism (Vals AI) | Not measured | 18.5% | Only Step 5 Preview measured |
| GPQA Diamond (Artificial Analysis run) | 84.0% | Not measured | Only DeepSeek V3.2 measured |
| Humanity's Last Exam (Artificial Analysis run) | 24.6% | 46.5% | Step 5 Preview |
| ARC-AGI | 57.0% | Not measured | Only DeepSeek V3.2 measured |
| ARC-AGI-2 | 4.0% | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: hard prompts | 1434 | 1461 | Step 5 Preview |
| LMArena Text: expert | 1436 | 1485 | Step 5 Preview |
| LMArena Text: hard prompts in English | 1446 | 1454 | Step 5 Preview |
| LMArena Vision: diagrams | Not measured | 1308 | Only Step 5 Preview measured |
| LMArena Vision: homework | Not measured | 1265 | Only Step 5 Preview measured |
| Neither model has a result on the other 10 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | DeepSeek V3.2 | Step 5 Preview | 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 |
| ProofBench | Not measured | 42.0% | Only Step 5 Preview measured |
| LMArena Text: math | 1435 | 1458 | Step 5 Preview |
| LMArena Text: mathematical work | 1440 | 1460 | Step 5 Preview |
| Neither model has a result on the other 6 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index13
| Benchmark | DeepSeek V3.2 | Step 5 Preview | Leader |
|---|---|---|---|
| MMLU Pro | 84.9% | Not measured | Only DeepSeek V3.2 measured |
| LegalBench (Vals AI) | 76.1% | Not measured | Only DeepSeek V3.2 measured |
| CaseLaw (Vals AI) | 55.4% | Not measured | Only DeepSeek V3.2 measured |
| CorpFin (Vals AI) | 51.0% | Not measured | Only DeepSeek V3.2 measured |
| TaxEval (Vals AI) | 68.2% | Not measured | Only DeepSeek V3.2 measured |
| MedQA (Vals AI) | 93.9% | Not measured | Only DeepSeek V3.2 measured |
| MMLU Pro (Artificial Analysis run) | 86.2% | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: science | 1446 | 1454 | Step 5 Preview |
| LMArena Text: medicine and healthcare | 1440 | 1439 | DeepSeek V3.2 |
| LMArena Text: legal and government | 1425 | 1437 | Step 5 Preview |
| LMArena Text: business and finance | 1410 | 1435 | Step 5 Preview |
| Neither model has a result on the other 7 boards here | |||
Web designDirectional onlyPublic boards, shown as published, kept out of the index9 boards, none shared
| Benchmark | DeepSeek V3.2 | Step 5 Preview | Leader |
|---|---|---|---|
| Design Arena: websites | 1175 | Not measured | Only DeepSeek V3.2 measured |
| Design Arena: UI components | 1146 | Not measured | Only DeepSeek V3.2 measured |
| Design Arena: game dev | 1131 | Not measured | Only DeepSeek V3.2 measured |
| Design Arena: data viz | 1158 | Not measured | Only DeepSeek V3.2 measured |
| Design Arena: 3D | 1132 | Not measured | Only DeepSeek V3.2 measured |
| Design Arena: all code categories | 1163 | Not measured | Only DeepSeek V3.2 measured |
| Design Arena: SVG | 1032 | Not measured | Only DeepSeek V3.2 measured |
| Design Arena: ASCII art | 1083 | Not measured | Only DeepSeek V3.2 measured |
| LiveBench design | Not measured | 76.9% | Only Step 5 Preview measured |
GeneralLike-for-likePublic boards, shown as published, kept out of the index015
| Benchmark | DeepSeek V3.2 | Step 5 Preview | Leader |
|---|---|---|---|
| LiveBench data analysis | Not measured | 63.5% | Only Step 5 Preview measured |
| Public Benefits Bench (Vals AI) | Not measured | 58.1% | Only Step 5 Preview measured |
| IFBench | 60.7% | Not measured | Only DeepSeek V3.2 measured |
| Long context reasoning (AA-LCR) | 73.3% | 88.3% | Step 5 Preview |
| DTBench | 85.6% | Not measured | Only DeepSeek V3.2 measured |
| LMCA | 28.8% | Not measured | Only DeepSeek V3.2 measured |
| 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 | 1440 | Step 5 Preview |
| LMArena Text: instruction following | 1413 | 1438 | Step 5 Preview |
| LMArena Text: creative writing | 1400 | 1409 | Step 5 Preview |
| LMArena Vision | Not measured | 1265 | Only Step 5 Preview measured |
| LMArena Text: English | 1439 | 1447 | Step 5 Preview |
| LMArena Text: non-English | 1409 | 1428 | Step 5 Preview |
| LMArena Text: Chinese | 1455 | 1500 | Step 5 Preview |
| LMArena Text: German | 1413 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: Japanese | 1374 | 1383 | Step 5 Preview |
| LMArena Text: Korean | 1370 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: French | 1429 | 1438 | Step 5 Preview |
| LMArena Text: Spanish | 1424 | 1441 | Step 5 Preview |
| LMArena Text: Russian | 1424 | 1433 | Step 5 Preview |
| LMArena Text: Polish | 1419 | Not measured | Only DeepSeek V3.2 measured |
| LMArena Text: multi-turn | 1427 | 1437 | Step 5 Preview |
| LMArena Text: longer queries | 1428 | 1449 | Step 5 Preview |
| LMArena Text: writing and language | 1407 | 1425 | Step 5 Preview |
| LMArena Text: entertainment and media | 1395 | 1412 | Step 5 Preview |
| LMArena Vision: English | Not measured | 1290 | Only Step 5 Preview measured |
| LMArena Vision: reading text | Not measured | 1282 | Only Step 5 Preview measured |
| Neither model has a result on the other 18 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | DeepSeek V3.2 | Step 5 Preview | 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 V3.2 costs less
Repository review
50K input + 3K output tokens
DeepSeek V3.2 costs less
Agent loop
200K input + 10K output tokens
DeepSeek V3.2 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 | StepFun Step 5 Preview |
|---|---|---|
| Maker | DeepSeek | StepFun |
| Weights | Open weights | Open weights |
| Released | December 1, 2025 | October 8, 2026 |
| Context window | 164K | 1.05M |
| Input $/Mtok | $0.29 | $1.00 |
| Output $/Mtok | $0.88 | $2.70 |
| Coding index | 30.0 | 61.1 |
| Independent measurements behind it | 3 | 4 |
| Cost per solved task | $0.99 | Not computable |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is DeepSeek V3.2 better than Step 5 Preview for coding?
Pick Step 5 Preview for coding: it scores 61.1 on the Vibe Coding Index against 30.0 for DeepSeek V3.2. Pick DeepSeek V3.2 for output tokens 3.1x cheaper ($0.88 vs $2.70 per million) and work measured by LMArena Text: medicine and healthcare, where it leads.
Which is cheaper, DeepSeek V3.2 or Step 5 Preview?
DeepSeek V3.2 costs $0.29 per million input tokens and $0.88 per million output tokens; Step 5 Preview costs $1.00 and $2.70. Cost per solved task is known for DeepSeek V3.2 only.
Which has the bigger context window, DeepSeek V3.2 or Step 5 Preview?
Step 5 Preview, with 1.05M tokens against 164K.
How many benchmarks is the DeepSeek V3.2 vs Step 5 Preview comparison based on?
32 benchmarks have a result for both models as of October 22, 2026. DeepSeek V3.2's coding index rests on 3 independent benchmarks and Step 5 Preview's on 4. Where only one model has a result, the page says so instead of naming a winner.