N
NexusRoute
Back to Models

DeepSeek R1

DeepSeekspecialized

DeepSeek's reasoning-specialized model built on the 671B MoE architecture. Rivals OpenAI's o1 on reasoning benchmarks at a tiny fraction of the cost. Uses visible chain-of-thought reasoning (unlike o1/o3 where reasoning is hidden). Open-weight and fully inspectable.

Released 2025-01-20Knowledge cutoff: 2025-01
Medium confidence|Updated 89d ago|88% source confidence

Specifications

Context Window

128K tokens

Max Output

32K tokens

Input Price

$0.550 / 1M tokens

Output Price

$2.19 / 1M tokens

Latency Tier

Slow (speed score: 4/10)

Capability Profile

Reasoning
9.5/10
Coding
9/10
Cost Efficiency
8.5/10
Factuality
8.5/10
Instruction Following
7/10
Structured Output
6.5/10
Long Context
6/10
Creativity
5.5/10
Tool Use
5.5/10
Safety & Enterprise
5/10
Conversational
4.5/10
Speed
4/10
Multimodal
1/10

Feature Support

Vision No
Audio In No
Audio Out No
Video No
Image Generation No
Image Editing No
Function Calling Yes
JSON Mode Yes
Structured Output No
Streaming Yes
Reasoning Yes
Realtime No
Computer Use No
Web Search No

Best Use Cases

Complex mathematical reasoning at a fraction of o3's cost
Science and logic problems requiring deep chain-of-thought
Coding tasks requiring careful multi-step analysis
Research problems where visible reasoning chains are valuable for verification
Budget-friendly reasoning for tasks that would otherwise need o3

Not Ideal For

Casual chat or conversational AI
Multimodal tasks (text only)
Enterprise deployments with safety/compliance requirements
Simple tasks where reasoning overhead is wasteful
Tasks requiring fast response times

Strengths

Reasoning quality rivaling o1/o3 at 5-20x lower cost
Visible chain-of-thought — you can inspect and verify the reasoning process
Open weights allow self-hosting and customization of the reasoning model
Strong on AIME, GPQA, and competition-level math benchmarks
Cost-effective alternative to o3 for most reasoning workloads

Weaknesses

Slow — reasoning chains can be very long and increase latency to 30-120s
Text-only with no multimodal capabilities
Safety alignment is minimal by Western enterprise standards
Chinese content censorship applies to the reasoning model too
Structured output compliance is weaker during reasoning mode
Conversational ability is poor — it's a reasoning specialist, not an assistant

Edge Cases & Notes

Visible reasoning tokens are part of the output and are billed accordingly
Reasoning chains can become very long on adversarial or ambiguous prompts
Distilled versions (R1-7B, R1-14B, R1-32B, R1-70B) are available for lighter deployments
The gap vs o3 is most visible on the hardest competition math problems

Provider Notes

Available through DeepSeek's API and third-party providers. Distilled versions available in multiple sizes. Self-hosting the full 671B model requires extensive GPU infrastructure. Data processing considerations apply.

Benchmarks

MMLU90.8%
HumanEval91.5%
Arena Elo1360
GSM8K97.5%

Benchmark Notes

GSM8K 97.5%. AIME ~79%. Rivaling o1 on most reasoning benchmarks. MMLU-Pro 90.8%. The most cost-efficient reasoning model available. Arena Elo reflects reasoning-specific evaluation.

Research Meta

Last Evaluated

2026-03-01

Source Confidence

88%

Evaluation Method

AIME, GPQA Diamond, GSM8K, MATH, LMSYS Arena (reasoning), cost-reasoning Pareto analysis

Needs Re-evaluation

No

Sources

  • DeepSeek R1 technical report
  • LMSYS Chatbot Arena
  • AIME evaluation results
  • Open LLM Leaderboard

Continue exploring

Route a prompt

See how DeepSeek R1 ranks

Compare models

Side-by-side analysis

Browse registry

Explore all 24 models