DeepSeek

DeepSeek V4 turns long context into a practical default

DeepSeek first became widely known for doing something the AI industry was not expecting: it made strong reasoning models dramatically cheaper and released open weights that developers could inspect, download and run. DeepSeek V4 pushes that argument further. The current model family combines a one-million-token context window, thinking and non-thinking modes, stronger agent skills, open weights and unusually low API prices.

The headline is not merely that V4 can hold more text. Many models now advertise enormous context windows. The important change is that DeepSeek treats one million tokens as the standard across its official services rather than a specialist tier. That shifts long-context work from a premium demonstration toward something developers can realistically test in coding agents, research tools, document systems and long-running workflows.

DeepSeek is also no longer one model with two confusing API aliases. The current lineup is clearer. DeepSeek-V4-Flash is the fast and economical model. DeepSeek-V4-Pro is the higher-capability option for complex reasoning and agent work. DeepSeek-V4-Flash-Vision-Exp adds image understanding as an experimental API model. The older deepseek-chat and deepseek-reasoner names were retired in July 2026, so integrations still relying on them need to move to the V4 model names.

V4 Flash and V4 Pro solve different cost problems

DeepSeek-V4-Flash is the default economic choice. DeepSeek describes it as a 284-billion-parameter mixture-of-experts model with 13 billion active parameters. It is designed to respond quickly while retaining reasoning that approaches V4 Pro on many tasks. DeepSeek says it performs on par with Pro on simpler agent work, making it the sensible starting point for high-volume coding assistance, extraction, classification, drafting and routine tool use.

DeepSeek-V4-Pro is much larger: 1.6 trillion total parameters with 49 billion active for a token. It is positioned for stronger world knowledge, math, science, coding and agentic execution. In the consumer app and web interface, users reach it through Expert Mode. In the API, the model name is deepseek-v4-pro.

The choice should not be reduced to cheap versus smart. A production team should build an evaluation set from its own tasks. Flash may be entirely sufficient for common requests and much less expensive at scale. Pro may reduce retries, supervision or multi-step failure enough to justify its higher token price. Routing simple requests to Flash and difficult ones to Pro can be more effective than choosing one model for everything.

Both models support thinking and non-thinking operation. Thinking is the default in the current API documentation, but developers can disable it for speed or control reasoning effort. DeepSeek describes low effort for simple tasks, high for routine agent workflows and max for difficult problems. This is useful because not every request deserves a long reasoning trace. A format conversion, short classification or obvious tool selection should not consume the same deliberation as a complex refactor or multi-document investigation.

One million tokens is capacity, not automatic understanding

A million-token context can hold a very large codebase, extensive research material or hundreds of ordinary documents. DeepSeek V4 also supports output up to a stated maximum of 384,000 tokens. Those numbers open interesting workflows: repository-scale review, long legal or policy comparison, persistent agent history, large document extraction and synthesis across many sources.

Capacity does not mean every token receives equal attention or that a model will reliably retrieve one subtle fact from a poorly organized mass of text. Long prompts increase latency, cost and the surface area for contradictory instructions. They can also make errors difficult to diagnose because the relevant evidence is buried among irrelevant material.

The best long-context workflow is selective. Remove duplicate files, generated artifacts and irrelevant history. Separate trusted instructions from untrusted source material. Add clear document boundaries and metadata. Ask the model to cite the exact source sections it used. For repeated requests over a stable prefix, DeepSeek's automatic disk-based context caching can reduce input costs when the beginning of a new request exactly matches a cached prefix.

Caching is one reason the pricing can become exceptionally low. A V4 Flash cache hit costs as little as $0.007 per million input tokens during off-peak hours. But a cache only helps when the prefix actually matches the stored unit. Minor changes near the start of a prompt can prevent a hit. Developers should keep stable system instructions and reference material at the front, then append changing user input later.

The API pricing rewards planned workloads

DeepSeek introduced separate peak and off-peak pricing with the V4 general release. Off-peak rates are half the peak price. Peak hours are 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, Monday through Friday. All other periods use the lower rate.

For V4 Flash and the experimental vision model, off-peak input costs $0.007 per million tokens on a cache hit and $0.22 on a cache miss. Output costs $0.66 per million tokens. Peak prices are $0.014 for cached input, $0.44 for uncached input and $1.32 for output.

V4 Pro costs more. Off-peak rates are $0.022 per million cached input tokens, $0.66 per million uncached input tokens and $1.98 per million output tokens. Peak prices rise to $0.044, $1.32 and $3.96 respectively.

These prices are low enough that token discipline can seem unimportant, but output and agent loops can still grow rapidly. Thinking mode, tool results, retry chains and large generated artifacts all add tokens. A runaway agent that repeatedly reads the same data or calls a failing tool can erase the advantage of a low model price. Set budgets, cap iterations, log cache-hit rates and measure successful task cost rather than price per million tokens alone.

Off-peak pricing is useful for batch work. Document indexing, nightly test generation, large-scale extraction and non-urgent code analysis can be scheduled outside peak windows. Interactive products cannot always wait, but background workloads often can. Teams operating across time zones should convert the published UTC windows carefully and avoid assuming their local night is automatically off-peak.

The API is designed to fit existing model clients

DeepSeek exposes an OpenAI-compatible base URL and an Anthropic-compatible endpoint. It supports Chat Completions, the Responses API, tool calls, JSON output, chat prefix completion and fill-in-the-middle completion for the text models in non-thinking mode. This lowers migration cost because many applications can switch providers by changing credentials, the base URL and model name.

Compatibility is not perfect interchangeability. Thinking mode has specific conversation rules. When tools are present, the reasoning_content from previous assistant turns must be passed back in subsequent requests, even if a turn did not call a tool. Omitting it can trigger a 400 error. JSON mode also needs explicit prompting, including the word json and an example of the desired structure, and the documentation warns that empty content can occasionally occur.

A migration should therefore run contract tests around streaming, tool schemas, structured output, error handling, token accounting and long conversations. A basic hello-world request proves authentication, not production compatibility. The more an application depends on provider-specific behavior, the more important it is to test actual workflows.

Agent work is now central to DeepSeek's positioning

DeepSeek V4 is built for agentic coding and multi-step tool use. Both Flash and Pro can reason through tool calls, and the official documentation provides integrations for agent and coding environments. DeepSeek also promotes DeepSeek Harness, its developer-preview environment for agent harness builders.

The one-million-token context is particularly relevant to coding agents. An agent can retain more repository context, tool output and work history before compaction. Pro's native Responses API support is promoted for Codex-style workflows, while Flash offers an economical backend for frequent coding tasks.

Agent quality still depends heavily on the harness. The model needs safe tools, clear permissions, reliable file operations, useful error messages and a policy for when to stop. A strong model inside a weak harness can loop, edit the wrong file or claim success without verification. For code, require tests and inspect diffs. For external systems, separate read access from write access and require approval for consequential actions.

Vision arrives as an experimental model

DeepSeek-V4-Flash-Vision-Exp adds image input while retaining V4 Flash's text capabilities. It can describe images, read screenshots and analyze charts. The API accepts JPEG, PNG, GIF and WebP files through base64 content, public URLs or the Files API. Images are converted into tokens and billed as input at V4 Flash rates.

The experimental label matters. Developers should expect behavior, limits or model naming to change. Vision-sensitive tasks also require their own evaluation. A model that describes a photograph well may still misread a dense chart, overlook a subtle interface state or confuse small text. Do not use it as the sole authority for medical images, identity, legal evidence or other high-stakes interpretation.

The Files API allows images to be reused without sending them in every request. Individual uploads can be as large as 64 MiB, and users can set expiration between one hour and 30 days or omit an expiry for permanent storage. Permanent should not be the default for sensitive screenshots. Set the shortest useful retention, delete test files after evaluation and avoid uploading information the service is not intended to process.

The consumer app is useful, but its privacy tradeoff is material

DeepSeek Web and the mobile app provide a free consumer assistant for coding, content creation, document reading and general questions. Instant Mode uses the faster path, while Expert Mode exposes V4 Pro. The app is the easiest way to experience the models without configuring an API or local runtime.

DeepSeek's privacy policy deserves a direct reading before anyone pastes confidential material into the hosted service. It says the service collects prompts, uploaded files, chat history, device and usage information, and may use personal data to develop and train its technology. Users can opt out of using personal data for model training, and can manage or delete chat history through settings.

The policy also states that personal data is directly collected, processed and stored in the People's Republic of China. It explicitly says the services are not designed to process sensitive personal data and asks users not to provide it. That is a significant consideration for government, health, legal, education and enterprise users with data-residency or confidentiality obligations.

Consumer convenience should not be confused with organizational approval. A company may permit public code questions but prohibit customer records, proprietary source or internal documents. Teams should establish a clear policy, review the service terms and choose local or differently governed deployment where hosted processing is unsuitable.

Open weights change the deployment decision

DeepSeek releases model weights and technical material, allowing qualified teams and hosting providers to run models outside the official service. This is one of the platform's most important differences from fully closed model vendors. It supports research, customization, private deployment and a broader ecosystem of third-party inference options.

Open weights do not make a model easy to self-host. V4 Pro and Flash are large mixture-of-experts systems that require serious infrastructure, optimized kernels and operational expertise. Most organizations will not run them on a workstation. Distilled or older DeepSeek models may fit smaller hardware, while full V4 deployment is more likely to happen through specialist infrastructure or cloud providers.

Self-hosting also transfers responsibility. The organization must secure the endpoint, patch dependencies, manage model and tokenizer versions, monitor abuse, scale inference and understand the applicable license. It can improve data control, but only if the whole system is operated well.

The model version should always be explicit

DeepSeek's rapid release cadence is a strength and a source of operational risk. V3, R1, V3.1, V3.2 and now V4 arrived in a relatively short period. API aliases can move to newer dated builds, and older identifiers can be retired. Results, tokenization and tool behavior may change even if the calling code looks similar.

Production teams should record the resolved model version with every evaluation and important output. Maintain regression tests for prompts, structured data and tool decisions. Watch the official changelog and pricing page rather than relying on old tutorials. Avoid hard-coding a retired name such as deepseek-chat or deepseek-reasoner into new software.

Where DeepSeek fits against other model platforms

DeepSeek V4 is a strong candidate for developers who need capable reasoning, coding, long context and agent support at a low API cost. It competes with OpenAI models, Anthropic Claude, Google Gemini, Qwen and other open-weight model families. Its combination of API compatibility and open weights gives teams more deployment options than a single closed endpoint.

It is not the automatic choice for every organization. Some buyers will prioritize data residency, enterprise contracts, mature administrative controls or a broader multimodal product. Others may prefer a smaller model that can be run locally with modest hardware. Consumer users should consider the hosted-service privacy terms, not only benchmark performance.

The right comparison uses a representative workload and a complete cost model. Measure answer quality, tool success, latency, cache behavior, output length, retries and human correction. Include privacy, compliance and infrastructure in the decision. A cheap token that produces an unusable result is expensive; an open model that requires an oversized operations team may not reduce total cost.

The WhatAI view

DeepSeek V4 is one of the most consequential model releases of 2026 because it combines three trends that are often sold separately: huge context, agent-ready reasoning and aggressively low inference prices. Flash makes high-volume experimentation accessible. Pro offers a stronger option for difficult work. Vision broadens the API, while open weights keep the ecosystem from depending entirely on one hosted service.

The strongest opportunity is not asking longer questions for the sake of it. It is building workflows that can bring a large, well-organized body of context into a reasoning and tool-use loop without making cost the dominant constraint. The main risk is treating capacity and low price as substitutes for architecture, evaluation or governance.

Use the app for ordinary low-sensitivity exploration. Use the API when you need controlled integration and measurement. Consider open deployment when data control or customization justifies the operational burden. Whichever route you choose, keep prompts selective, test the exact model version and place human review around decisions that matter.

Use Cases

Related Frontier LLMs & Chat Models Tools

Explore the Network