DeepSeek V3 vs DeepSeek R1
Side-by-side comparison of DeepSeek V3 (DeepSeek) and DeepSeek R1 (DeepSeek). Exact API pricing per million tokens, context windows, output speed, and total cost on real-world prompts.
Specifications
| Spec | DeepSeek V3 | DeepSeek R1 |
|---|---|---|
| Provider | DeepSeek | DeepSeek |
| Model id | deepseek-v3 | deepseek-r1 |
| Input price (per 1M tokens) | $0.28 | $0.28 |
| Output price (per 1M tokens) | $0.42 | $0.42 |
| Context window | 128,000 | 128,000 |
| Output speed (tokens/sec) | ~60 | ~40 |
Cost on real prompts
Total cost = (input tokens × input price) + (output tokens × output price). Numbers below use the exact pricing tables published by each provider.
| Scenario | Input | Output | DeepSeek V3 | DeepSeek R1 | Cheaper |
|---|---|---|---|---|---|
| Short question + answer | 50 | 150 | $0.000077 | $0.000077 | Tied |
| Code review on one file | 500 | 1,500 | $0.00077 | $0.00077 | Tied |
| Long document summary | 5,000 | 500 | $0.001610 | $0.001610 | Tied |
| Heavy reasoning task | 2,000 | 8,000 | $0.003920 | $0.003920 | Tied |
| Full codebase analysis | 50,000 | 10,000 | $0.018200 | $0.018200 | Tied |
Want the exact cost for your prompt instead of these examples? Open the cost calculator pre-loaded with both models →
When to pick which
Heuristics derived from the spec table above. Always validate on your own prompts before committing — these are starting points, not verdicts.
Pick DeepSeek V3 for
- •latency-sensitive UX (chat, autocompletion) — deepseek-v3 streams faster (~60 vs ~40 tok/s)
Pick DeepSeek R1 for
- •complex multi-step reasoning — deepseek-r1 uses chain-of-thought internally
Switching between them
For most use cases, switching providers means updating the model id and the request shape if the providers differ. Within the same provider, it's usually a single-line change.
From DeepSeek V3 to DeepSeek R1
# Before
model = "deepseek-v3"
# After
model = "deepseek-r1"If the providers differ (DeepSeek vs DeepSeek), you'll also need to swap the SDK / endpoint URL. Cross-provider migrations usually take 30 minutes to a few hours depending on how many features (streaming, function calling, tool use) you depend on.
Calculate cost on your own prompt
The examples above use generic input/output ratios. For an exact comparison, paste your real prompt into the calculator — it counts tokens with the right tokenizer for each model and shows side-by-side cost.
Open the calculator with both models →