DeepSeek V4.1 Flash vs Step 5 Preview: Benchmark Comparison
DeepSeek V4.1 Flash has the higher score on 13 of 17 shared benchmarks; Step 5 Preview leads on 4.
The largest observed score gap is 14.34 pts on Vibe Code Bench v1.1 , where DeepSeek V4.1 Flash leads.
Reported ±1 standard-error ranges overlap on 9 of 17 shared benchmarks with comparable uncertainty data. This is not a pairwise statistical significance test.
Shared benchmark results
Scores and reported standard errors come from the latest visible Vals benchmark version available for both models.
| Benchmark | DeepSeek V4.1 Flash | Step 5 Preview | Gap | Reported uncertainty |
|---|---|---|---|---|
| Vals Finance Index | 57.08% ±1.45 | 55.43% ±1.44 | 1.65 pts | Reported ±1 SE ranges overlap |
| Vals Science Index | 29.65% ±0.99 | 29.58% ±1.10 | 0.07 pts | Reported ±1 SE ranges overlap |
| Legal Research Bench | 41.35% ±3.42 | 40.38% ±3.41 | 0.96 pts | Reported ±1 SE ranges overlap |
| EMB v1.1 | 67.76% ±2.56 | 69.55% ±2.49 | 1.79 pts | Reported ±1 SE ranges overlap |
| Finance Agent v2.1 | 65.43% ±2.02 | 58.90% ±2.05 | 6.53 pts | Reported ±1 SE ranges do not overlap |
| Tax Agent Bench | 26.63% ±3.14 | 27.27% ±3.11 | 0.64 pts | Reported ±1 SE ranges overlap |
| ProofBench v1.1 | 54.00% ±5.01 | 42.00% ±4.96 | 12.00 pts | Reported ±1 SE ranges do not overlap |
| BioMysteryBench | 67.78% ±1.11 | 67.41% ±0.37 | 0.37 pts | Reported ±1 SE ranges overlap |
| MysteryMechanism | 21.17% ±2.75 | 18.47% ±2.61 | 2.70 pts | Reported ±1 SE ranges overlap |
| Terminal-Bench-Science | 0.00% ±0.00 | 2.86% ±2.01 | 2.86 pts | Reported ±1 SE ranges do not overlap |
| Code Migration v1.1 | 58.94% ±2.66 | 47.00% ±3.43 | 11.93 pts | Reported ±1 SE ranges do not overlap |
| IOI | 40.28% ±2.56 | 38.06% ±6.69 | 2.22 pts | Reported ±1 SE ranges overlap |
| SWE-bench Multimodal | 46.46% ±2.28 | 40.83% ±2.25 | 5.63 pts | Reported ±1 SE ranges do not overlap |
| Terminal-Bench 4.0 | 19.70% ±1.75 | 31.82% ±2.62 | 12.12 pts | Reported ±1 SE ranges do not overlap |
| Vibe Code Bench v1.1 | 84.74% ±2.89 | 70.40% ±4.27 | 14.34 pts | Reported ±1 SE ranges do not overlap |
| SRE Bench | 0.76% ±0.54 | 0.38% ±0.38 | 0.38 pts | Reported ±1 SE ranges overlap |
| Public Benefits Bench v1.1 | 64.28% ±1.25 | 58.12% ±1.28 | 6.16 pts | Reported ±1 SE ranges do not overlap |
Performance by category
| Category | DeepSeek V4.1 Flash average | Step 5 Preview average |
|---|---|---|
| Legal | 41.35% | 40.38% |
| Finance | 53.27% | 51.91% |
| Math | 54.00% | 42.00% |
| Science | 29.65% | 29.58% |
| Coding | 50.02% | 45.62% |
| Cyber | 0.76% | 0.38% |
| Social Mobility | 64.28% | 58.12% |
Cost and latency
Cost per test appears only where the benchmark reports it for both models. Latency is the measured completion time for that benchmark.
| Benchmark | DeepSeek V4.1 Flash cost | Step 5 Preview cost | DeepSeek V4.1 Flash latency | Step 5 Preview latency |
|---|---|---|---|---|
| Vals Finance Index | $0.25 | $0.82 | 11m37s | 24m42s |
| Vals Science Index | $0.29 | $1.78 | 1h07m | 1h07m |
| Legal Research Bench | $0.25 | $0.84 | 11m42s | 36m08s |
| EMB v1.1 | $0.23 | $1.33 | 13m00s | 43m35s |
| Finance Agent v2.1 | $0.37 | $0.58 | 16m00s | 9m03s |
| Tax Agent Bench | $0.13 | $0.55 | 5m49s | 21m29s |
| ProofBench v1.1 | $0.13 | $0.71 | 9m07s | 38m34s |
| BioMysteryBench | $0.23 | $0.49 | 14m50s | 21m35s |
| MysteryMechanism | $0.16 | $0.60 | 8m04s | 29m20s |
| Terminal-Bench-Science | $0.47 | $4.25 | 2h57m | 2h30m |
| Code Migration v1.1 | $1.09 | $8.61 | 1h54m | 3h17m |
| IOI | $0.24 | $2.03 | 18m30s | 1h20m |
| SWE-bench Multimodal | $0.23 | $0.61 | 19m24s | 26m06s |
| Terminal-Bench 4.0 | $0.50 | $3.11 | 48m22s | 1h58m |
| Vibe Code Bench v1.1 | $0.41 | $5.86 | 15m28s | 1h31m |
| SRE Bench | $0.55 | $3.07 | 1h01m | 1h34m |
| Public Benefits Bench v1.1 | $0.07 | $0.50 | 9m59s | 27m19s |
Results available only for DeepSeek V4.1 Flash
- Vals Inference Provider Index
- LegalBench
- MedCode
- MedScribe
- SAGE
- ProgramBench
- SkillsBench
- Vibe Code Bench 1-100
- CyberBench v1.1
Results available only for Step 5 Preview
None.