DeepSeek API Price per Million Tokens vs Frontier Western Labs
| Model | Input, cache miss ($/1M tokens) | Output ($/1M tokens) | Notes |
|---|---|---|---|
| DeepSeek V4.1-Flash (deepseek-flash) | $0.30 peak / $0.15 off-peak | $1.20 peak / $0.60 off-peak | Model version V4.1-Flash-1017; 1M context |
| DeepSeek V4-Pro (deepseek-v4-pro) | $1.32 peak / $0.66 off-peak | $3.96 peak / $1.98 off-peak | Model version V4-Pro-0813; 1M context |
| Anthropic Claude Opus 4.7 (reference) | $5 | $25 | Anthropic list price, one rate at all hours |
DeepSeek-V3 Disclosed Training Cost
| Figure | Value | Source / status |
|---|---|---|
| Official V3 training run (GPU-hours) | 2.788M H800 GPU-hours | DeepSeek-V3 technical report (disclosed): 2.664M pre-training, 119K context extension, 5K post-training |
| Official V3 training cost (at $2/GPU-hr) | $5.576M | DeepSeek-disclosed at an assumed H800 rental price of $2 per GPU-hour; excludes prior research and ablation experiments |
| DeepSeek-R1 training cost (at $2/GPU-hr) | $294K (147K H800 GPU-hours) | DeepSeek R1 paper, Table 7: R1-Zero $202K, SFT data creation $10K, R1 $82K |
DeepSeek Model-Family Release Timeline and Parameter Scaling
| Date | Model | Total / active params, context | Significance |
|---|---|---|---|
| Nov 2023 | DeepSeek LLM (v1) | 7B and 67B parameters, trained on 2T tokens | Open-source Base and Chat variants |
| May 2024 | DeepSeek-V2 | 236B total / 21B active, 128K context | Introduced Multi-head Latent Attention (MLA) and DeepSeekMoE; released 6 May 2024 |
| Dec 2024 | DeepSeek-V3 | 671B total / 37B active, trained on 14.8T tokens | Technical report puts the official training run at 2.788M H800 GPU-hours, $5.576M at an assumed $2 per GPU-hour |
| Jan 2025 | DeepSeek-R1 | 671B total / 37B active | Open-weights reasoning model released 20 Jan 2025 under the MIT License; DeepSeek says its performance is on par with OpenAI-o1; 90.8 MMLU in the R1 paper |
| Dec 2025 | DeepSeek-V3.2 | Continued pre-training from V3.1-Terminus (128K context) | Released 1 Dec 2025 with V3.2-Speciale; first DeepSeek model to integrate thinking directly into tool use |
| Apr 24 2026 | DeepSeek-V4 (Pro + Flash) | V4-Pro 1.6T / 49B active; V4-Flash 284B / 13B active; 1M context | Open-sourced preview; V4-Pro pre-trained on 33T tokens and V4-Flash on 32T; DeepSeek says V4-Pro rivals the world's top closed-source models |

DeepSeek in AI Software
KPI data tables
DeepSeek operating metrics behind the KPI charts, newest period first (up to 24 periods per chart). Periods use each company's own fiscal labels where the source does; each table names its source.
| Period | DeepSeek V4-Pro Max | OpenAI GPT-5.4 xHigh | Google Gemini-3.1-Pro High |
|---|---|---|---|
| Codeforces rating | 3,206 | 3,168 | 3,052 |
Source: DeepSeek-V4-Pro model card, DeepSeek-V4-Pro-Max vs Frontier Models table (huggingface.co/deepseek-ai/DeepSeek-V4-Pro); DeepSeek-V4 Technical Report (arxiv.org/abs/2606.19348); DeepSeek API Docs, DeepSeek-V4 Preview Release (24 Apr 2026).
| Period | DeepSeek V4-Pro Max | Claude Opus-4.6 Max | Google Gemini-3.1-Pro High |
|---|---|---|---|
| LiveCodeBench | 93.5 | 88.8 | 91.7 |
| SWE-bench Verified (%) | 80.6 | 80.8 | 80.6 |
Source: DeepSeek-V4-Pro model card, DeepSeek-V4-Pro-Max vs Frontier Models table (huggingface.co/deepseek-ai/DeepSeek-V4-Pro); DeepSeek-V4 Technical Report (arxiv.org/abs/2606.19348); DeepSeek API Docs, DeepSeek-V4 Preview Release (24 Apr 2026).
