DeepSeek API Pricing 2026
R1, V3 Cost vs Claude
Full DeepSeek R1 and V3 pricing breakdown — tokens, cache discounts, monthly cost examples, and a direct comparison with Claude Sonnet and Opus so you can choose the right model for your budget.
DeepSeek R1 Input
DeepSeek R1 Output
DeepSeek V3 Input
R1 Cache Hit
DeepSeek Model Pricing Table
| Model | Type | Input ($/M) | Output ($/M) | Cache Read ($/M) | Context |
|---|---|---|---|---|---|
| DeepSeek R1 | Reasoning | $0.55 | $2.19 | $0.14 | 64K |
| DeepSeek R1 (distilled) | Reasoning | $0.14 | $0.28 | $0.035 | 32K |
| DeepSeek V3 | Chat/Code | $0.27 | $1.10 | $0.07 | 64K |
DeepSeek vs Claude — Full Comparison
| Feature / Model | DeepSeek R1 | DeepSeek V3 | Claude Sonnet 4.6 | Claude Haiku 4.5 |
|---|---|---|---|---|
| Input price/M | $0.55 | $0.27 | $3.00 | $0.80 |
| Output price/M | $2.19 | $1.10 | $15.00 | $4.00 |
| Cache read/M | $0.14 (75% off) | $0.07 (75% off) | $0.30 (90% off) | $0.08 (90% off) |
| Context window | 64K | 64K | 200K | 200K |
| Vision/image input | No | No | Yes | Yes |
| Tool use / function calling | Limited | Yes | Yes (robust) | Yes (robust) |
| Reasoning / CoT | Native (R1) | Good | Good | Basic |
| Data residency | China-origin | China-origin | US (Anthropic) | US (Anthropic) |
| Enterprise SLA | No | No | Yes | Yes |
Monthly Cost Examples
| Workload (per month) | DeepSeek R1 | DeepSeek V3 | Claude Sonnet 4.6 | Claude Haiku 4.5 |
|---|---|---|---|---|
| 10M in / 2M out | $9.88 | $4.90 | $60 | $16 |
| 100M in / 20M out | $98.80 | $49 | $600 | $160 |
| 100M in / 20M out (80% cache hit) |
$57.60 | $28.40 | $114 | $46.40 |
| 1B in / 100M out | $769 | $380 | $4,500 | $1,200 |
When to Use DeepSeek vs Claude
| Use case | Best choice | Why |
|---|---|---|
| Math / logic reasoning | DeepSeek R1 | Native CoT, near-OpenAI-o1 quality at fraction of cost |
| High-volume code generation | DeepSeek V3 | Strong coding, lowest price uncached |
| Agentic loops with tool calling | Claude Sonnet 4.6 | Robust tool use; cache gives cost parity at scale |
| Document / image analysis | Claude Haiku 4.5 | DeepSeek lacks vision; 200K context handles large docs |
| EU/US data compliance required | Claude (any) | DeepSeek API routes through Chinese infrastructure |
| Batch offline reasoning tasks | DeepSeek R1 | Max cost savings when latency doesn't matter |
Frequently Asked Questions
How much does DeepSeek R1 API cost per token?
DeepSeek R1 costs $0.55 per million input tokens and $2.19 per million output tokens via the DeepSeek API. With context caching enabled, cache hits drop to $0.14/M — a 75% discount on repeated context. For reasoning-heavy workloads like math or code, R1 is roughly 10× cheaper than Claude Opus 4.7 sticker price.
Is DeepSeek cheaper than Claude Haiku?
At sticker price, yes: DeepSeek V3 ($0.27/M) is 3× cheaper than Claude Haiku 4.5 ($0.80/M). But Claude's 90% cache discount is more aggressive than DeepSeek's 75%. On cache-heavy pipelines (80%+ hit rate), Haiku's effective input cost drops to ~$0.16/M — nearly matching V3's cache price. Haiku also adds vision, 200K context, and robust tool use that DeepSeek lacks.
Can I run DeepSeek R1 on my own infrastructure?
Yes. DeepSeek R1 is open-weights (MIT license) and can be run on GPU clusters via Ollama, vLLM, or Together AI. Self-hosting eliminates per-token API costs entirely but requires GPU capacity: R1 (671B parameters) needs ~350GB VRAM for full precision, or ~180GB for 4-bit quantized. Smaller distilled versions (7B–70B) run on a single A100 or H100. For most companies, Together AI or Fireworks hosted inference at $0.50–0.80/M input is more practical than self-hosting.
What is the DeepSeek R1 context window?
DeepSeek R1 supports a 64K token context window via the DeepSeek API. This is significantly smaller than Claude's 200K context. For long-document analysis (PDFs, large codebases, extended conversations), Claude Sonnet or Haiku are better choices. For chain-of-thought reasoning tasks that fit in 64K, R1 remains extremely competitive on cost.
Does DeepSeek API have data privacy concerns?
DeepSeek is a Chinese company (High-Flyer Capital) and its API infrastructure routes through servers in China. This raises data residency concerns for organizations in regulated industries (healthcare, finance, government) or those subject to GDPR, CCPA, or US federal data handling requirements. Anthropic's Claude API is US-based with SOC 2 Type II certification and enterprise data agreements. For sensitive data, use Claude or run a self-hosted DeepSeek deployment on your own infrastructure.
How do I calculate my exact DeepSeek API costs?
Monthly cost = (input_tokens × $0.00000055) + (output_tokens × $0.00000219) for R1. At 100M input + 20M output tokens/month with 80% cache hit rate: (20M fresh × $0.55) + (80M cache × $0.14) + (20M output × $2.19) = $11 + $11.20 + $43.80 = $66/month. Use the Claude Cost Calculator to model your exact Claude workload alongside these DeepSeek figures for a side-by-side comparison.