# DeepSeek V4 Pro 0813 vs Step 3.7 Flash

> Side-by-side benchmark comparison for DeepSeek V4 Pro 0813 and Step 3.7 Flash across agentic, coding, multimodal, knowledge, reasoning, multilingual, and math tasks.

- Canonical page: https://benchlm.ai/compare/deepseek-v4-pro-0813-vs-step-3-7-flash
- Last updated: September 27, 2026

- Shared sourced benchmarks: 6
- HTML indexing: indexable
- Ranking lane: BenchAlign v5.7

## Quick Verdict

Pick DeepSeek V4 Pro 0813 if you want the stronger benchmark profile. Step 3.7 Flash only makes more sense when its price, context window, or workload-specific category wins matter more than the overall score.

## Summary

- DeepSeek V4 Pro 0813 leads overall 63.48 to 41.41.
- The clearest category separation is in knowledge, where the averages are 63.4 for DeepSeek V4 Pro 0813 and 43.1 for Step 3.7 Flash.
- The biggest single benchmark swing is Terminal-Bench 2.1 in Agentic, with scores of 87.9% and 59.5%.
- Step 3.7 Flash is the cheaper option on output tokens, which matters if you expect large responses or heavy interactive use.
- DeepSeek V4 Pro 0813 also has the larger context window at 1M.

## Model Snapshot

| Property | DeepSeek V4 Pro 0813 | Step 3.7 Flash |
|----------|----------|----------|
| Creator | DeepSeek | StepFun |
| Type | Open Weight | Open Weight |
| Reasoning | Reasoning | Reasoning |
| Context | 1M | 256K |
| Overall Score | 63.48 | 41.41 |
| Benchmarks Covered | 42 | 11 |
| Pricing (input/output) | $1.32 / $3.96 | $0.20 / $1.15 |

## Category Breakdown

### Agentic

- Winner: Directional only
- DeepSeek V4 Pro 0813 public-lane score: 55 (Supported · #29/105)
- Step 3.7 Flash public-lane score: 35.2 (Estimated · #56/105)

| Benchmark | DeepSeek V4 Pro 0813 | Step 3.7 Flash | Winner |
|-----------|-----------|-----------|--------|
| Terminal-Bench 2.0 | 67.9% | Coming soon | Coming soon |
| Terminal-Bench 2.1 | 87.9% | 59.5% | DeepSeek V4 Pro 0813 |
| BrowseComp | 83.4% | 75.8% | DeepSeek V4 Pro 0813 |
| HLE w/ tools | 60.0% | 47.2% | DeepSeek V4 Pro 0813 |
| MCP Atlas | 73.6% | Coming soon | Coming soon |
| Toolathlon | 51.8% | 49.5% | DeepSeek V4 Pro 0813 |
| CyberGym | 83.3% | Coming soon | Coming soon |
| Toolathlon-Verified | 74.1% | Coming soon | Coming soon |
| Agents' Last Exam | 25.7% | Coming soon | Coming soon |
| AutomationBench | 31.8% | Coming soon | Coming soon |
| Terminal-Bench 2.1 (Vals) | 54.7% | Coming soon | Coming soon |
| DeepSearchQA | Coming soon | 92.8% | Coming soon |
| Claw-Eval | Coming soon | 67.1% | Coming soon |
| Gert Labs | Coming soon | 51.57% | Coming soon |

### Coding

- Winner: Directional only
- DeepSeek V4 Pro 0813 public-lane score: 50.3 (Supported · #39/135)
- Step 3.7 Flash public-lane score: 32.4 (Estimated · #81/135)

| Benchmark | DeepSeek V4 Pro 0813 | Step 3.7 Flash | Winner |
|-----------|-----------|-----------|--------|
| LiveCodeBench Pass@1-COT | 93.5% | Coming soon | Coming soon |
| Codeforces | 3206.0 | Coming soon | Coming soon |
| SWE-bench Verified | 80.6% | Coming soon | Coming soon |
| SWE-bench Pro | 55.4% | 56.3% | Step 3.7 Flash |
| SWE Multilingual | 76.2% | Coming soon | Coming soon |
| Terminal-Bench 2.0 | 67.9% | Coming soon | Coming soon |
| Vibe Code Bench | 49.93% | Coming soon | Coming soon |
| Terminal-Bench 2.1 | 87.9% | 59.5% | DeepSeek V4 Pro 0813 |
| NL2Repo | 61.5% | Coming soon | Coming soon |
| DeepSWE | 62.7% | Coming soon | Coming soon |
| DSBench-FullStack | 71.1% | Coming soon | Coming soon |
| DSBench-Hard | 67.2% | Coming soon | Coming soon |
| OpenHarmony Bench | 59.0% | Coming soon | Coming soon |
| LiveCodeBench (Vals) | 87.5% | Coming soon | Coming soon |
| SWE-bench (Vals) | 96.4% | Coming soon | Coming soon |

### Reasoning

- Winner: Not comparable
- DeepSeek V4 Pro 0813 public-lane score: 56.9 (Unranked · 4 rankable rows)
- Step 3.7 Flash public-lane score: 72.7 (Unranked · 2 rankable rows)

| Benchmark | DeepSeek V4 Pro 0813 | Step 3.7 Flash | Winner |
|-----------|-----------|-----------|--------|
| MRCR 1M | 83.5% | Coming soon | Coming soon |
| CorpusQA 1M | 62.0% | Coming soon | Coming soon |
| ARC-AGI-1 | 90.00% | Coming soon | Coming soon |
| ARC-AGI-2 | 61.3% | Coming soon | Coming soon |

### Multimodal & Grounded

- Winner: Not comparable
- DeepSeek V4 Pro 0813 public-lane score: Coming soon
- Step 3.7 Flash public-lane score: 71 (Unranked · 3 rankable rows)

| Benchmark | DeepSeek V4 Pro 0813 | Step 3.7 Flash | Winner |
|-----------|-----------|-----------|--------|
| SimpleVQA | Coming soon | 79.2% | Coming soon |
| V* | Coming soon | 95.3% | Coming soon |

### Knowledge

- Winner: Directional only
- DeepSeek V4 Pro 0813 public-lane score: 63.4 (Estimated · #29/158)
- Step 3.7 Flash public-lane score: 43.1 (Estimated · #78/158)

| Benchmark | DeepSeek V4 Pro 0813 | Step 3.7 Flash | Winner |
|-----------|-----------|-----------|--------|
| MMLU-Pro | 87.5% | Coming soon | Coming soon |
| SimpleQA | 57.9% | Coming soon | Coming soon |
| Chinese-SimpleQA | 84.4% | Coming soon | Coming soon |
| GPQA | 90.1% | Coming soon | Coming soon |
| GPQA-D | 90.1% | Coming soon | Coming soon |
| HLE | 42.7% | Coming soon | Coming soon |
| GPQA Diamond (Vals) | 92.4% | Coming soon | Coming soon |
| MMLU-Pro (Vals) | 87.0% | Coming soon | Coming soon |

### Instruction Following

- Winner: Not comparable
- DeepSeek V4 Pro 0813 public-lane score: Coming soon
- Step 3.7 Flash public-lane score: 80.6 (#53/124)

### Mathematics

- Winner: Not comparable
- DeepSeek V4 Pro 0813 public-lane score: 80.2 (Unranked · 4 rankable rows)
- Step 3.7 Flash public-lane score: Coming soon

| Benchmark | DeepSeek V4 Pro 0813 | Step 3.7 Flash | Winner |
|-----------|-----------|-----------|--------|
| HMMT Feb 2026 | 95.2% | Coming soon | Coming soon |
| IMOAnswerBench | 89.8% | Coming soon | Coming soon |
| Apex | 38.3% | Coming soon | Coming soon |
| Apex Shortlist | 90.2% | Coming soon | Coming soon |

## FAQ

### Which is better overall, DeepSeek V4 Pro 0813 or Step 3.7 Flash?

DeepSeek V4 Pro 0813 is ahead overall on BenchLM right now.

### Where is the biggest gap between DeepSeek V4 Pro 0813 and Step 3.7 Flash?

The widest category gap is in knowledge, where the averages are 63.4 for DeepSeek V4 Pro 0813 and 43.1 for Step 3.7 Flash.

### How many benchmarks does DeepSeek V4 Pro 0813 cover on BenchLM?

DeepSeek V4 Pro 0813 currently has 42 sourced benchmark scores on BenchLM.

### How many benchmarks does Step 3.7 Flash cover on BenchLM?

Step 3.7 Flash currently has 11 sourced benchmark scores on BenchLM.

## Related Comparisons

- [DeepSeek V4 Pro 0813 vs GPT-6 Astra](/compare/deepseek-v4-pro-0813-vs-gpt-6-astra)
- [Step 3.7 Flash vs GPT-6 Astra](/compare/gpt-6-astra-vs-step-3-7-flash)
- [DeepSeek V4 Pro 0813 vs Claude Opus 5.5](/compare/claude-opus-5-5-vs-deepseek-v4-pro-0813)
- [Step 3.7 Flash vs Claude Opus 5.5](/compare/claude-opus-5-5-vs-step-3-7-flash)

## Explore More

- [DeepSeek V4 Pro 0813 profile](/models/deepseek-v4-pro-0813)
- [Step 3.7 Flash profile](/models/step-3-7-flash)
- [Compare Pricing](/llm-pricing)
- [Alternative Finder](/tools/alternative-finder)
- [LLM Selector](/tools/llm-selector)
- [Overall Rankings](/best/overall)