DeepSeek V3.2 vs GPT-5.4
Head to head on 24 shared benchmarks, with price, context window, and release dates.
Benchmark scores
48evals · sort by any column, ⤢ to expand fullscreen
| Model | |||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT-5.4 OpenAI | 65.3% | 81.0% | 1836elo | 78.3 | 47.6% | 74.8% | 11.3% | 48.4% | 78.5% | 82.0% | 77.5% | 79.3% | 70.2% | 82.6% | 94.1% | 88.1% | 82.8% | 56.0% | 47.1% | 74.0% | 37.9% | 80 | 67.4% | 36.0% | - | - | - | 13.1% | - | 1.47 | 34.1% | 9.0% | 1272elo | 87.2% | - | 92063 | - | 41.3% | 77.5% | - | 68.3% | 86 | 52.2% | 43.3% | -0.61 | 40.8 | - | - | 60.3% |
| DeepSeek V3.2 DeepSeek | 47.9% | 16.0% | 1511elo | 66.8 | 22.1% | 75.1% | 11.2% | 49.0% | 68.0% | 65.9% | 64.6% | 50.0% | 48.2% | 70.4% | 85.0% | 77.2% | 57.5% | 8.0% | 38.7% | 68.2% | 32.6% | 65 | 5.1% | 78.9% | 74.7% | 74.2% | 59.0% | - | 0.0% | - | - | - | - | - | 31.3% | - | 59.3% | - | - | 83.7% | - | - | - | - | - | - | 59.0% | 70.0% | 51.7% |
2 / 2 models
| Attribute | DeepSeek V3.2 DeepSeek | GPT-5.4 OpenAI |
|---|---|---|
| Description | DeepSeek V3.2 is an AI model from DeepSeek, released with open weights. | GPT-5.4 is an AI model from OpenAI. |
| Family | Deepseek | GPT |
| Params | - | - |
| Open weights | open | closed |
| License | mit | proprietary |
| Released | 1 Dec 2025 | 5 Mar 2026 |
| Context | 163,840 | 1,050,000 |
| Input $/1M | $0.28 | $2.50 |
| Output $/1M | $0.42 | $15.00 |
| Cache read $/1M | $0.130 | $0.250 |
| Cache write $/1M | - | - |
| Speed | 0 tok/s | 0 tok/s |
| Latency | 0.00 s | 0.00 s |
| Modalities | text | text, image, file |
| Providers | - | - |
| Publisher | DeepSeek | OpenAI |
| Reported scores | Aider Polyglot Benchmark74.2% (Aider) MathArena57.47% GPQA Diamond75.1% (AA) Humanity's Last Exam (HLE)11.2% (AA) | GPQA Diamond74.8% (AA) SciCode47.1% (AA) τ²-bench (Tau²-bench)36.0% (AA) Terminal-Bench (Hard)37.9% (AA) Vibe Code Bench v1.167.4% |
FAQ
- Is DeepSeek V3.2 or GPT-5.4 better?
- Across the 24 benchmarks both models report on Sophon, GPT-5.4 leads on 21 of them. Which one is "better" depends on the benchmark - the table above shows every shared score.
- How do DeepSeek V3.2 and GPT-5.4 compare on GPQA Diamond?
- DeepSeek V3.2 scores 75.1% and GPT-5.4 scores 74.8% on GPQA Diamond.
- Which is cheaper, DeepSeek V3.2 or GPT-5.4?
- DeepSeek V3.2 is cheaper at $0.42 per million output tokens against $15.00 for GPT-5.4 - about 35.7x.
- When were DeepSeek V3.2 and GPT-5.4 released?
- DeepSeek V3.2 was released 1 Dec 2025 by DeepSeek. GPT-5.4 was released 5 Mar 2026 by OpenAI.
Comparing something else? Build your own side-by-side.