DeepSeek’s official API documentation says V4 Pro API service continues after September 14, 2026, with billing unchanged. The same pricing page gives DeepSeek V4.1 Flash a global U.S.-dollar schedule in which cache reuse matters more than the headline token price: cached input costs $0.003 per 1 million tokens off-peak, while uncached input costs $0.15.
That 50-fold gap is the practical news for developers building agents that repeatedly send the same instructions, files or tool context. Output is billed separately at $0.60 per 1 million tokens off-peak.
DeepSeek V4.1 Flash API pricing
The deepseek-flash identifier serves DeepSeek V4.1 Flash. Billing is based on total input and output tokens, with separate rates for cache hits, cache misses and generated output.
| Token category | Off-peak price per 1M tokens | Peak price per 1M tokens | Condition |
| Input, cache hit | $0.003 | $0.006 | Reused cached input tokens |
| Input, cache miss | $0.15 | $0.30 | Uncached input tokens |
| Output | $0.60 | $1.20 | Generated output tokens |
Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday. Every other period is off-peak, and each peak rate is exactly double its off-peak counterpart.
A simple workload with 1 million cached input tokens and 1 million output tokens costs $0.603 off-peak. Replacing the cache hit with 1 million uncached input tokens raises that combination to $0.75. At peak rates, the same totals become $1.206 and $1.50, respectively.
These are token charges rather than a complete application budget. Total spending depends on how many tokens an application sends and generates, how often its context qualifies for reuse, and when requests are processed.
Why cached input matters for agents
A KV cache stores intermediate information from processed context so the system can reuse it instead of repeating the same work. That makes cache reuse particularly relevant to agents, which may submit a long system prompt, documents and tool traces across multiple turns.
DeepSeek V4.1 Flash’s model card lists a global KV-cache footprint of approximately 890 bytes per token, roughly one quarter of the figure listed for DeepSeek V4 Flash. The model uses a 1-million-token context, so memory efficiency becomes especially relevant when applications keep large amounts of material active.
The price schedule reflects the same practical distinction. A cache hit is dramatically cheaper than a cache miss, but an application must actually reuse an eligible prefix or context to receive that rate. Sending similar information in a way that does not qualify as cached input is billed at the uncached-input price.
What the model offers
DeepSeek V4.1 Flash is a multimodal mixture-of-experts model with a 552-billion-parameter backbone. It activates 8 billion parameters per token during prefill, when it processes input, and 16 billion during decoding, when it generates output.
The model accepts text and images and produces autoregressive text. Its API supports tool calls, JSON output, the Responses API and Anthropic-format API access. The official Flash listing gives it a concurrency limit of 2,500 and a maximum output of 384,000 tokens.
| Specification | DeepSeek V4.1 Flash | Operational relevance |
| API identifier | deepseek-flash | The endpoint identifier for the Flash service |
| Context length | Up to 1 million tokens | Supports very large prompts and retained agent context |
| Backbone | 552B parameters | Describes total model capacity, not active computation for every token |
| Active parameters | 8B during prefill; 16B during decoding | Separates input processing from text generation |
| Input | Text and images | Enables multimodal prompts |
| Output | Autoregressive text | The model generates text responses |
| Maximum output | 384K tokens | Sets the documented output ceiling |
| API features | Tool calls, JSON output, Responses API and Anthropic API access | Supports structured and tool-using integrations |
| Concurrency limit | 2,500 | The documented limit for deepseek-flash |
| License | MIT for the repository and model weights | The license listed for the repository and weights |
The 8B and 16B active-parameter figures describe how computation is routed during different stages; they do not by themselves specify a universal local hardware configuration. The model card also lists a 40-layer Causal Encoder-Decoder architecture with 20 encoder layers and 20 decoder layers, plus Compressed Sparse Attention 2 and Engram conditional memory.
What the V4 Pro update changes
DeepSeek’s current documentation says V4 Pro API service continues after September 14, 2026, with billing unchanged. That supersedes the earlier description of a temporary switch to V4.1 Flash.
For developers, the distinction matters because an API identifier and the model serving a request are separate parts of an integration. The current pricing documentation preserves V4 Pro service rather than describing a temporary switch to Flash.
DeepSeek also says legacy V4 Flash names remain accepted but are served by V4.1 Flash and billed at the Flash price. The company reserves the right to adjust prices, so the published schedule can change.
