Zurück zur Entwicklerzone

DeepSeek Modelle

Entdecken Sie alle 4 Modelle von DeepSeek mit detaillierten Preisen, Vor- und Nachteilen sowie Entwicklerempfehlungen.

4
Modelle
$0.140
Niedrigster Input
1M
Max. Kontext
2
Qualitätsstufen

Schnellempfehlungen

Bestes Preis-Leistungs-Verhältnis: DeepSeek V4 Flash ($0.140/1M)
Beste Qualität: DeepSeek V4 Flash
Am besten für logisches Denken: DeepSeek V4 Pro

DeepSeek V4 Pro

Reasoning

Deep reasoning, complex analysis

Offizielle Preise

Wann verwenden: Best budget reasoning model with massive output capacity for math, logic, and analysis.

Upgrade-Highlights

  • 384K max output — largest in the market (3x GPT-5.5's 128K)
  • 1M context window — 16x increase over DeepSeek R1's 65K
  • 99% cheaper than GPT-5.5 for reasoning ($0.435 vs $5/M input)
  • Cached input: $0.003625/M — 99.2% savings for repeated prefixes
  • Thinking mode: toggleable deep reasoning without extra cost
Input-Preis
$0.435
per 1M tokens
Output-Preis
$0.870
per 1M tokens
Cached Input
$0.0036
per 1M tokens
Batch-Input
per 1M tokens
Kontextfenster: 1M
Max. Output: 384,000 tokens
Wissensstand: 2025-06
VisionFunktionsaufrufFeinabstimmungJSON-Modus

Vorteile

  • 384K max output — largest available
  • 1M context window
  • 99% cheaper than GPT-5.5 for reasoning

Nachteile

  • No vision support
  • No fine-tuning
  • Thinking mode adds token overhead

Leistung

Ausgabegeschwindigkeit~30 tok/s
Rate-Limit2,000 RPM

Multimodal

BildeingabeBildausgabeAudioeingabeAudioausgabe

Benchmarks

MMLU
88.0%
AIME 2024
85.0%
SWE-bench Verified
71.0%
MATH
82.5%

DeepSeek V4 Flash

Flagship

Ultra-cheap coding, general tasks

Offizielle Preise

Wann verwenden: Best for high-volume coding and general tasks where cost is the primary concern.

Upgrade-Highlights

  • Cheapest LLM available: $0.14/M input, $0.28/M output
  • 1M context + 384K output — same capacity as V4 Pro at 1/3 the price
  • Dual mode: thinking and non-thinking — switch based on task complexity
  • Cached input: $0.0028/M — 98% savings for system prompts
  • No vision — text-only; use GPT-4.1 or Gemini for multimodal needs
Input-Preis
$0.140
per 1M tokens
Output-Preis
$0.280
per 1M tokens
Cached Input
$0.0028
per 1M tokens
Batch-Input
per 1M tokens
Kontextfenster: 1M
Max. Output: 384,000 tokens
Wissensstand: 2025-06
VisionFunktionsaufrufFeinabstimmungJSON-Modus

Vorteile

  • Cheapest LLM available at $0.14/M input
  • 1M context + 384K output
  • Supports both thinking and non-thinking modes

Nachteile

  • No vision support
  • Quality below GPT-5.4 for complex tasks
  • No fine-tuning

Leistung

Ausgabegeschwindigkeit~60 tok/s
Rate-Limit5,000 RPM

Multimodal

BildeingabeBildausgabeAudioeingabeAudioausgabe

Benchmarks

MMLU
84.5%
HumanEval
86.0%
SWE-bench Verified
55.0%

DeepSeek V3

Flagship

Coding, math, reasoning

Offizielle Preise

Wann verwenden: Best budget flagship for coding-heavy workloads. 10x cheaper than GPT-4.1 for comparable quality.

Upgrade-Highlights

  • 10x cheaper than GPT-4.1 ($0.27 vs $2/M input) for comparable coding
  • Fine-tuning available — customize for domain-specific coding tasks
  • Prompt caching: $0.07/M — 74% savings for repeated system prompts
  • 671B MoE params with 37B active — flagship intelligence at budget price
  • 65K context is small vs 1M competitors — upgrade to V4 for long context
Input-Preis
$0.270
per 1M tokens
Output-Preis
$1.10
per 1M tokens
Cached Input
$0.070
per 1M tokens
Batch-Input
per 1M tokens
Kontextfenster: 66K
Max. Output: 8,192 tokens
Wissensstand: 2024-07
VisionFunktionsaufrufFeinabstimmungJSON-Modus

Vorteile

  • Extremely cheap for flagship quality
  • Excellent coding & math
  • Fine-tuning + caching available

Nachteile

  • Only 65K context
  • No vision
  • Limited function calling reliability

Leistung

Ausgabegeschwindigkeit~50 tok/s
Rate-Limit3,000 RPM

Multimodal

BildeingabeBildausgabeAudioeingabeAudioausgabe

Benchmarks

MMLU
87.1%
SWE-bench Verified
50.0%
MATH
75.0%
HumanEval
89.0%

Agenten die dieses Modell verwenden

2

DeepSeek R1

Reasoning

Deep reasoning, math proofs

Offizielle Preise

Wann verwenden: Budget reasoning model — use when o3 cost is prohibitive for math/logic pipelines.

Upgrade-Highlights

  • 4x cheaper than o3 ($0.55 vs $2/M input) for comparable reasoning
  • AIME 2024: 79.8% — competitive with o1 on math benchmarks
  • Prompt caching: $0.14/M — 75% savings for repeated prefixes
  • Open-source reasoning model — self-host for unlimited usage
  • No function calling or vision — pure reasoning; upgrade to V4 Pro for tools
Input-Preis
$0.550
per 1M tokens
Output-Preis
$2.19
per 1M tokens
Cached Input
$0.140
per 1M tokens
Batch-Input
per 1M tokens
Kontextfenster: 66K
Max. Output: 8,192 tokens
Wissensstand: 2024-07
VisionFunktionsaufrufFeinabstimmungJSON-Modus

Vorteile

  • 4x cheaper than o3 for reasoning
  • Strong math & logic
  • Good prompt caching savings

Nachteile

  • No function calling
  • No vision or fine-tuning
  • Only 65K context

Leistung

Ausgabegeschwindigkeit~20 tok/s
Rate-Limit2,000 RPM

Multimodal

BildeingabeBildausgabeAudioeingabeAudioausgabe

Benchmarks

AIME 2024
79.8%
MATH
97.3%
GPQA
71.5%

Nebeneinander-Vergleich

ModellStufeInputOutputKontext
DeepSeek V4 ProReasoning$0.435$0.8701M
DeepSeek V4 FlashFlagship$0.140$0.2801M
DeepSeek V3Flagship$0.270$1.1066K
DeepSeek R1Reasoning$0.550$2.1966K