overfeed.news

DeepSeek adjusts API prices across V4 series on OpenRouter

399 palavras

Esta notícia ainda só está disponível em inglês.

DeepSeek changed API pricing for three models in its V4 series on OpenRouter effective October 29, cutting input and output rates for DeepSeek V4 Flash 0423 by 35 percent, raising both rates for DeepSeek V4 Pro 0423 by 9.1 percent, and moving DeepSeek V4.1 Flash input down 67.9 percent while output rose 19 percent.

Price changes

DeepSeek V4 Flash 0423 now costs $0.091 per million input tokens and $0.182 per million output tokens, down from $0.14 and $0.28 respectively on October 28. DeepSeek V4 Pro 0423 moved to $0.9553 per million input tokens and $1.9105 per million output tokens, up from $0.8754 and $1.7508. DeepSeek V4.1 Flash shifted to $0.045 per million input tokens and $0.50 per million output tokens, compared with $0.14 and $0.42 the previous day.

Model details

All three models provide a one-million-token context window and have open weights. DeepSeek V4 Flash 0423 has 290.9 billion parameters and was released April 24, 2026. DeepSeek V4 Pro 0423 has 1,598.8 billion parameters and the same release date. DeepSeek V4.1 Flash has 763.2 billion parameters and was released September 10, 2026.

Official specifications

According to deepseek-ai, the DeepSeek-V4 series includes DeepSeek-V4-Pro with 1.6 trillion total parameters and 49 billion activated parameters, and DeepSeek-V4-Flash with 284 billion total parameters and 13 billion activated parameters. Both models support a context length of one million tokens. The series incorporates a hybrid attention mechanism combining Compressed Sparse Attention and Heavily Compressed Attention to improve long-context efficiency. According to deepseek-ai, in the one-million-token context setting, DeepSeek-V4-Pro requires only 27 percent of single-token inference FLOPs and 10 percent of KV cache compared with DeepSeek-V3.2. The series also employs Manifold-Constrained Hyper-Connections to strengthen residual connections and enhance signal propagation stability, and uses the Muon optimizer for faster convergence and greater training stability. Both models are pre-trained on more than 32 trillion diverse and high-quality tokens. Post-training features a two-stage paradigm: independent cultivation of domain-specific experts through SFT and RL with GRPO, followed by unified model consolidation via on-policy distillation. Models are available for download on Hugging Face and ModelScope under the MIT License.

  • DeepSeek V4 Flash 0423 and V4.1 Flash input prices dropped sharply, while V4 Pro 0423 prices increased modestly.
  • All three models maintain a one-million-token context window and open weights.
  • The V4 series uses hybrid attention, mHC, and the Muon optimizer according to deepseek-ai.

Publicado por deepseek-ai. Leia na fonte.

deepseek-ai/DeepSeek-V4-Flash

DeepSeek adjusts API prices across V4 series on OpenRouter —…