DeepSeek V4 vs GPT 5.4 Mini: Benchmark Comparison
DeepSeek V4 has the higher score on 13 of 15 shared benchmarks; GPT 5.4 Mini leads on 1.
The largest observed score gap is 21.91 pts on Tax Agent Bench , where DeepSeek V4 leads.
Reported ±1 standard-error ranges overlap on 4 of 15 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 | GPT 5.4 Mini | Gap | Reported uncertainty |
|---|---|---|---|---|
| Vals Index | 38.63% ±1.16 | 33.17% ±1.28 | 5.46 pts | Reported ±1 SE ranges do not overlap |
| Harvey's Legal Agent Benchmark | 3.75% ±1.43 | 0.00% ±0.00 | 3.75 pts | Reported ±1 SE ranges do not overlap |
| Legal Research Bench | 23.08% ±2.93 | 12.50% ±2.30 | 10.58 pts | Reported ±1 SE ranges do not overlap |
| EMB | 51.62% ±3.05 | 45.43% ±3.65 | 6.20 pts | Reported ±1 SE ranges overlap |
| Finance Agent (v2) | 44.08% ±0.65 | 45.36% ±0.45 | 1.28 pts | Reported ±1 SE ranges do not overlap |
| Tax Agent Bench | 58.50% ±3.13 | 36.59% ±3.10 | 21.91 pts | Reported ±1 SE ranges do not overlap |
| TaxEval v2 | 72.08% ±0.88 | 71.22% ±0.90 | 0.86 pts | Reported ±1 SE ranges overlap |
| GPQA Diamond | 89.39% ±1.65 | 83.08% ±2.46 | 6.31 pts | Reported ±1 SE ranges do not overlap |
| MMLU Pro | 87.25% ±0.34 | 84.55% ±0.36 | 2.69 pts | Reported ±1 SE ranges do not overlap |
| Code Migration | 26.20% ±4.04 | 12.94% ±3.64 | 13.27 pts | Reported ±1 SE ranges do not overlap |
| LiveCodeBench | 87.48% ±0.95 | 81.47% ±1.09 | 6.02 pts | Reported ±1 SE ranges do not overlap |
| ProgramBench | 0.00% ±0.00 | 0.00% ±0.00 | 0.00 pts | Reported ±1 SE ranges overlap |
| SWE-bench | 77.40% ±1.87 | 73.00% ±1.99 | 4.40 pts | Reported ±1 SE ranges do not overlap |
| Terminal-Bench 4.0 | 11.11% ±1.34 | 2.52% ±1.01 | 8.59 pts | Reported ±1 SE ranges do not overlap |
| Vibe Code Bench v1.1 | 49.93% ±4.77 | 47.97% ±5.61 | 1.96 pts | Reported ±1 SE ranges overlap |
Performance by category
| Category | DeepSeek V4 average | GPT 5.4 Mini average |
|---|---|---|
| Legal | 13.41% | 6.25% |
| Finance | 56.57% | 49.65% |
| Academic | 88.32% | 83.82% |
| Coding | 42.02% | 36.32% |
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 cost | GPT 5.4 Mini cost | DeepSeek V4 latency | GPT 5.4 Mini latency |
|---|---|---|---|---|
| Vals Index | $1.31 | $1.45 | 32m09s | 29m49s |
| Harvey's Legal Agent Benchmark | $0.74 | $1.00 | 6m42s | 11m21s |
| Legal Research Bench | $0.68 | $1.86 | 17m05s | 37m36s |
| EMB | $0.80 | $2.15 | 13m42s | 43m21s |
| Finance Agent (v2) | $0.88 | $1.20 | 25m26s | 35m54s |
| Tax Agent Bench | $0.72 | $0.81 | 20m45s | 16m46s |
| TaxEval v2 | N/A | N/A | 3m10s | 44.58s |
| GPQA Diamond | N/A | N/A | 15m00s | 63.11s |
| MMLU Pro | N/A | N/A | 9m38s | 20.00s |
| Code Migration | $1.07 | $1.95 | 28m58s | 28m53s |
| LiveCodeBench | N/A | N/A | 11m56s | 3m09s |
| ProgramBench | N/A | N/A | 1h03m | 53m45s |
| SWE-bench | $0.44 | $0.51 | 10m35s | 5m26s |
| Terminal-Bench 4.0 | $3.38 | $1.43 | 1h28m | 29m41s |
| Vibe Code Bench v1.1 | $2.21 | $1.19 | 56m59s | 34m15s |
Results available only for DeepSeek V4
- LegalBench
- MedCode
- MedScribe
- ProofBench v1.1
- SkillsBench
- Public Benefits Bench v1.1
Results available only for GPT 5.4 Mini
- MortgageTax
- MMMU Pro
- SAGE