DeepSeek V4 Flash
DeepSeek V4 Flash is a speed-oriented model in DeepSeek's V4 family, offering a 1,000,000-token context window for high-volume and long-document workloads.
API Pricing
Cheapest on OpenRouter — 74% below avg| Provider | Input / 1M | Output / 1M | Cached / 1M |
|---|---|---|---|
| $0.049 | $0.098 | $0.0098 | |
| $0.060 | $0.180 | $0.015 | |
| $0.091 | $0.182 | $0.018 | |
| $0.140 | $0.280 | $0.028 | |
| $0.140 | $0.280 | $0.0028 | |
| $0.140 | $0.280 | - | |
| $0.140 | $0.280 | $0.030 | |
| $0.140 | $0.300 | - | |
| $0.140 | $0.280 | $0.028 | |
| $0.149 | $0.322 | $0.074 | |
| $0.220 | $0.660 | - | |
| $0.250 | $0.300 | $0.060 | |
| $0.810 | $2.44 | - |
Prices updated daily. Last check: Sep 25, 2026
DeepSeek V4 Flash pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Reasoning & Knowledge
- GPQA Diamond71.6%
- Humanity's Last Exam7.8%
Agentic & Tool Use
- Terminal-Bench Hard34.1%
- τ²-bench94.4%
Instruction & Long Context
- IFBench47.2%
- Long-Context Reasoning41.7%
Benchmarks measured Sep 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- DeepSeek
- Family
- DeepSeek V4
- Context Window
- 1.0M
- Modalities
- Text
Capabilities
- Tool Calling
- No
- Open Source
- No
- Aliases
- deepseek-ai/DeepSeek-V4-Flash, DeepSeek-V4-Flash, deepseek-v4-flash
Strengths & Limitations
Strengths
- 1,000,000-token context window, enough to hold entire repositories or large document sets in one request
- Flash tier positioning targets low-latency, high-throughput serving rather than maximum per-query compute
- Very long context can reduce or remove the need for a separate chunking and retrieval layer in some pipelines
- Available under several standard aliases (deepseek-ai/DeepSeek-V4-Flash, deepseek-v4-flash), making multi-provider routing straightforward
- Part of the DeepSeek V4 family, so applications can potentially move between Flash and heavier V4 variants with similar prompting
- Long-context capacity suits agent loops where message history accumulates across many turns
Limitations
- We do not track published benchmark scores for this model, so capability claims cannot be verified from our data
- Throughput and time-to-first-token figures in our database are unpopulated, so real-world speed must be measured per provider
- Flash-tier models generally trade reasoning depth for speed compared with larger siblings in the same family
- Input modalities beyond text are not confirmed in our metadata — verify vision or audio support with your provider
- Using the full 1M-token context raises per-request cost and latency substantially compared with short prompts, regardless of tier
Key Features
About DeepSeek V4 Flash
Common Use Cases
DeepSeek V4 Flash fits workloads that combine large inputs with a need for fast, repeated inference: summarizing or classifying long documents at volume, answering questions across an entire code repository without building a retrieval index, processing meeting or support transcripts, and driving agent loops whose message history grows over many steps. The 1,000,000-token context is the main reason to select it over a shorter-context fast model, since it lets you skip chunking logic for inputs that would otherwise need it. For tasks that hinge on difficult multi-step reasoning, competitive-level math, or intricate refactoring, evaluate heavier models in the DeepSeek V4 family or dedicated reasoning models alongside it before committing.
Frequently Asked Questions
How much does DeepSeek V4 Flash cost to run?
Pricing depends on which inference provider serves the model and on the pricing type — input tokens, output tokens, cached input, and any batch or long-context surcharges are billed differently. Because providers change rates and add or drop models regularly, see the pricing table on this page for current per-provider figures. Note that requests that fill a large share of the 1M-token context will cost far more than short prompts, so estimate based on your actual average input length.
What is DeepSeek V4 Flash best used for?
It suits long-input, high-volume work: document and transcript summarization, repository-wide code questions, batch classification and extraction, and agent conversations with long histories. The 1,000,000-token context lets you pass large material directly instead of building a chunking and retrieval pipeline, while the Flash tier targets fast turnaround.
How does DeepSeek V4 Flash differ from other DeepSeek V4 models?
The Flash label marks it as the speed- and throughput-oriented member of the DeepSeek V4 family. Heavier variants in the family are generally aimed at deeper reasoning on harder problems, while Flash targets lower latency and cheaper high-volume serving. We do not track benchmark comparisons between the variants, so run your own evaluation on representative tasks if the difference matters to your application.
Can DeepSeek V4 Flash accept images or other non-text inputs?
Our database does not record input modalities beyond text for this model, which means we have not confirmed image or audio support either way. Check the model card published by the provider you plan to use before designing around multimodal input.
Does the 1,000,000-token context mean I should stop using retrieval?
Not necessarily. A 1M-token window makes it practical to skip chunking for many mid-sized corpora, but filling the context increases both cost and latency per request, and models can attend unevenly across very long inputs. A common pattern is to use retrieval to narrow the material and the long context to avoid aggressive truncation, rather than choosing one approach exclusively.