Model comparison
DeepSeek V4 Pro vs Step 5 Preview
DeepSeek V4 Pro is the stronger model overall, scoring 54.3 to 47.9 on the Noometry Index.
Last verified . 17 shared benchmarks.
DeepSeek
StepFunSummary
- They share 17 benchmarks with published results for both. DeepSeek V4 Pro scores higher in 8 categories and Step 5 Preview in 0 categories; 4 gaps are clear of the uncertainty.
- The widest gap is in math, where DeepSeek V4 Pro leads 64.8 to 46.1.
- The biggest single-benchmark swing is ProofBench: 50% for DeepSeek V4 Pro and 42% for Step 5 Preview.
- DeepSeek V4 Pro is cheaper at $0.66 / $1.98 per million input/output tokens, against $1 / $2.70 for Step 5 Preview.
- Step 5 Preview accepts more context: 1.02M tokens versus 1M.
- DeepSeek V4 Pro has downloadable open weights; the other is API-only.
Side by side
| DeepSeek V4 Pro | Step 5 Preview | |
|---|---|---|
| Provider | DeepSeek | StepFun |
| Noometry Index | 54.3 | 47.9 |
| Released | 2026-04-24 | 2026-09-16 |
| Weights | Open | Proprietary |
| Context window | 1M | 1.02M |
| Max output | 393K | 66K |
| Input $ / M tokens | $0.66 | $1 |
| Output $ / M tokens | $1.98 | $2.70 |
| Results tracked | 48 | 18 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding Too close to call
DeepSeek V4 Pro: 52.4 (#34), Step 5 Preview: 51.8 (#36)
| Benchmark | DeepSeek V4 Pro | Step 5 Preview |
|---|---|---|
| LMArena WebDev | 1582 | 1564 |
| SciCode | 51% | 58.9% |
| LMArena Coding | 1470 | 1480 |
| SWE-bench Verified | 77.6% | — |
| FrontierCode | 28.6% | — |
| WeirdML | 66.2% | — |
| ALE-Bench | 1,403 | — |
Agentic & Tool Use Not comparable
DeepSeek V4 Pro: 32.8 (#58), Step 5 Preview: —
| Benchmark | DeepSeek V4 Pro | Step 5 Preview |
|---|---|---|
| APEX-Agents | 47.3% | — |
| Vending-Bench 2 | 3,285 | — |
Reasoning DeepSeek V4 Pro leads
DeepSeek V4 Pro: 56.5 (#24), Step 5 Preview: 40.0 (#57)
| Benchmark | DeepSeek V4 Pro | Step 5 Preview |
|---|---|---|
| CritPt | 18% | 20.9% |
| LMArena Hard Prompts | 1461 | 1465 |
| ARC-AGI-2 | 61.3% | — |
| Kagi LLM Benchmark | 53.5% | — |
| NYT Connections (extended) | 91.3% | — |
| ARC-AGI-1 | 90.5% | — |
| Chess Puzzles | 47% | — |
| Mystery Game Puzzles | 43% | — |
| DTBench | 93.9% | — |
| LMCA | 45.5% | — |
| Surface Evolver Bench | 40% | — |
| Epoch Capabilities Index | 155.31 | — |
| ForecastBench | 56.1 | — |
Math DeepSeek V4 Pro leads
DeepSeek V4 Pro: 64.8 (#30), Step 5 Preview: 46.1 (#72)
| Benchmark | DeepSeek V4 Pro | Step 5 Preview |
|---|---|---|
| ProofBench | 50% | 42% |
| LMArena Math | 1455 | 1470 |
| FrontierMath (Tiers 1-3) | 64.6% | — |
| FrontierMath Tier 4 | 26.8% | — |
| MathArena Final-Answer Competitions | 76.6% | — |
| OTIS Mock AIME 2024-2025 | 98.6% | — |
Knowledge DeepSeek V4 Pro leads
DeepSeek V4 Pro: 59.5 (#31), Step 5 Preview: 41.2 (#112)
| Benchmark | DeepSeek V4 Pro | Step 5 Preview |
|---|---|---|
| LMArena Expert | 1464 | 1470 |
| GPQA Diamond | 91.7% | — |
| SimpleQA Verified | 52.9% | — |
| Vectara Hallucination Rate | 8.6% | — |
Multimodal Not comparable
DeepSeek V4 Pro: —, Step 5 Preview: 41.0 (#41)
| Benchmark | DeepSeek V4 Pro | Step 5 Preview |
|---|---|---|
| LMArena Vision | — | 1267 |
Multilingual Too close to call
DeepSeek V4 Pro: 54.4 (#45), Step 5 Preview: 53.8 (#54)
| Benchmark | DeepSeek V4 Pro | Step 5 Preview |
|---|---|---|
| LMArena Non-English | 1439 | 1432 |
| LMArena Chinese | 1486 | 1519 |
| LMArena Russian | 1453 | 1436 |
| LMArena Spanish | 1458 | 1439 |
| LMArena French | 1472 | — |
| LMArena German | 1458 | — |
| LMArena Japanese | 1445 | — |
| LMArena Korean | 1447 | — |
Instruction Following Too close to call
DeepSeek V4 Pro: 76.1 (#47), Step 5 Preview: 76.0 (#50)
| Benchmark | DeepSeek V4 Pro | Step 5 Preview |
|---|---|---|
| LMArena Instruction Following | 1448 | 1444 |
Long Context Too close to call
DeepSeek V4 Pro: 45.0 (#51), Step 5 Preview: 44.5 (#64)
| Benchmark | DeepSeek V4 Pro | Step 5 Preview |
|---|---|---|
| LMArena Longer Query | 1458 | 1455 |
| CL-bench Life | 13.5% | — |
Writing & Preference DeepSeek V4 Pro leads
DeepSeek V4 Pro: 65.5 (#46), Step 5 Preview: 62.8 (#71)
| Benchmark | DeepSeek V4 Pro | Step 5 Preview |
|---|---|---|
| LMArena Text | 1451 | 1442 |
| LMArena Creative Writing | 1446 | 1410 |
| LMArena Multi-Turn | 1467 | 1447 |
| EQ-Bench Creative Writing | 1553 | — |
| EQ-Bench 4 | 1166 | — |
Frequently asked questions
Is DeepSeek V4 Pro better than Step 5 Preview?
DeepSeek V4 Pro is the stronger model overall, scoring 54.3 to 47.9 on the Noometry Index.
Which is cheaper, DeepSeek V4 Pro or Step 5 Preview?
DeepSeek V4 Pro is cheaper. It lists at $0.66 per million input tokens and $1.98 per million output tokens; Step 5 Preview lists at $1 and $2.70.
Is DeepSeek V4 Pro or Step 5 Preview better for coding?
They score almost the same on coding (52.4 vs 51.8); test both on your own repository before choosing.
Which has the bigger context window?
Step 5 Preview does, with 1.02M tokens against 1M.
How many benchmarks do DeepSeek V4 Pro and Step 5 Preview share?
17 benchmarks have published results for both models. DeepSeek V4 Pro has 48 scored results on Noometry and Step 5 Preview has 18.