DeepSeek R1
DeepSeek R1 — open-weight reasoning model matching o3 quality at 10× lower cost.
| Type | Price / 1M tokens | Notes |
|---|---|---|
| Input | $0.55 | Per 1M input tokens |
| Output | $2.19 | Per 1M output tokens |
| Cache read | $0.14 | Cached prefix read-back |
DeepSeek R1 at $0.55/M input vs OpenAI o3 at $10/M — the cost-per-reasoning-token gap is 18×. For STEM, code, and math tasks, R1 is the most disruptive value in the market.
Best for
- Math and STEM problem solving
- Complex code debugging and generation
- Multi-step logical reasoning
- Research and analysis tasks
Not ideal for
- Fast real-time inference (reasoning models are slow by design)
- Simple Q&A and classification
- Tasks requiring real-time data access
Real cost scenarios
| Scenario | Est. monthly cost | Breakdown |
|---|---|---|
| Research assistant — 10k complex queries/month | $22 | 2,000 token avg input × 10k = 20M tokens at $0.55/M |
| Code review agent — 5k PRs/month | $165 | 6,000 tokens avg per PR × 5k = 30M input at $0.55/M + 30M output at $2.19/M |
Estimates with typical token distributions. Use the Prompt Cost Calculator for your exact workload.
How to cut costs on DeepSeek R1
- Use the DeepSeek R1 Distill Qwen 32B via Together AI at $0.50/M for 80% of the quality at nearly the same price.
- DeepSeek R1 Distill Llama 70B is a good self-hosting option — open weights, strong reasoning, fits in 40GB VRAM at Q4.
Calculate your DeepSeek R1 costs
Other DeepSeek models
Compare DeepSeek R1
Frequently Asked Questions
How does DeepSeek R1 compare to OpenAI o3?
On MATH-500 and competitive programming benchmarks, R1 and o3 score within 2-5% of each other. R1 is $0.55/M input vs o3's $10/M — a 18× cost difference. For most STEM tasks, R1 is the rational choice. o3 retains an edge on the very hardest reasoning problems.
Can I self-host DeepSeek R1?
Yes — R1 weights are open-source under MIT license. The full model is 671B parameters, requiring a multi-GPU cluster. For practical self-hosting, use the distilled variants: R1 Distill Qwen 32B (fits in 24GB VRAM at Q4) or R1 Distill Llama 70B (fits in 40GB VRAM at Q4).
Pricing verified 2026-08-19 from DeepSeek official pricing. All prices in USD per 1M tokens. See methodology and price changelog.