deepseek-chat
deepseek-chat is a cost-effective, high-capacity conversational LLM API with a 1M-token context window, thinking/non-thinking modes, tool calls, and OpenAI-compatible endpoints.
Last verified:
What is deepseek-chat?
deepseek-chat is an AI-powered chat and code-assistant API built on DeepSeek's large-language models, designed for developers and product teams who want to embed conversational and reasoning capabilities into their applications. It offers an ultra-long context window of up to 1M tokens, allowing processing of very long documents, codebases, or complex conversations in a single request. The API supports both thinking (reasoning-intensive) and non-thinking modes, tool calls, JSON-mode output, and context-aware completions, making it suitable for chatbots, code generation, data analysis assistants, and complex workflow automation.
Key features include a massive 1M-token context window, built-in support for both quick, non-thinking interactions and deliberate reasoning paths, OpenAI-style endpoint compatibility for minimal integration changes, tool calls and JSON-mode responses, input caching at a drastically lower price than cache-miss input, and high concurrency limits — especially for the Flash variant — making it appropriate for real-time chat and high-volume applications. Pricing is computed per token on both input and output.
deepseek-chat is well suited for developers and technical teams processing long code files, documentation, or large datasets, as well as for building chat experiences where memory and context continuity are critical. It fits internal knowledge-base assistants, code-review bots, AI-driven customer support tools, and workloads that reuse common prompts or system messages. The OpenAI-compatible interface lets teams swap or dual-use models with minimal integration effort.
deepseek-chat pricing
Pricing model: Freemium
Deepseek Chat is billed per 1M tokens for both input and output; the expense is calculated as the number of tokens multiplied by the per‑token price and deducted from your topped‑up or granted balance with granted credits used first. For the deepseek‑v4‑flash model, the input cache‑hit price is $0.0028 per 1M tokens, the input cache‑miss price is $0.14 per 1M tokens, and the output price is $0.28 per 1M tokens. For the deepseek‑v4‑pro model, the input cache‑hit price is $0.003625 per 1M tokens (originally $0.0145, discounted 75%), the input cache‑miss price is $0.435 per 1M tokens (originally $1.74, discounted 75%), and the output price is $0.87 per 1M tokens (originally $3.48, discounted 75%). After the 75% discount promotion ends on May 31, 2026, the deepseek‑v4‑pro pricing will be permanently adjusted to one‑quarter of the original post‑discount rate. The models also have different concurrency limits: 2500 for deepseek‑v4‑flash and 500 for deepseek‑v4‑pro, with charges applied only when both input and output tokens are consumed.
deepseek-chat pros
- ultra‑long 1M‑token context window
- supports both thinking and non‑thinking modes
- OpenAI‑style API for easy integration
- tool calling support for complex workflows
- JSON‑mode output for structured responses
- chat‑prefix and FIM completion in beta
- input caching with 90% lower price than cache‑miss
- high concurrency limits (2500 for Flash)
- 500 concurrency limit for Pro which is still robust
- per‑token pricing that scales with usage
- compatibility with Anthropic‑style endpoints
- real‑time streaming response support
- built‑in rate‑limiting and user isolation
- broad feature parity with leading commercial models
- lower effective cost for repeated prompts via caching
deepseek-chat cons
- pricing can become expensive at high token volumes
- cache‑hit benefit only applies to repeated inputs
- thinking mode increases latency and cost
- model names like deepseek‑chat and deepseek‑reasoner are being deprecated
- API pricing is subject to periodic adjustment
- no separate free tier described; usage deducts from balance
- high concurrency still has hard limits per user
- complexity added by dual thinking/non‑thinking modes
- beta features like FIM and chat‑prefix may change
- requires careful token usage tracking to avoid overruns
Frequently asked questions about deepseek-chat
Is there a free tier for Deepseek Chat API usage?
There is no distinct free tier listed; instead, charges are deducted from your topped‑up or granted balance, with granted balance used first when both are available, so initial usage may be subsidized through promotional credits but ongoing usage is paid per token consumed. To avoid unexpected charges, you are advised to monitor your balance and top up according to your expected token usage.
How are tokens counted and billed in Deepseek Chat?
Tokens are counted for both input and output in a request, and the total expense is computed as the number of tokens multiplied by the applicable per‑million‑tokens price for that model and traffic pattern. The billable amount is then deducted from your account balance, with any available granted balance exhausted before touching your topped‑up funds.
What is the difference between cache‑hit and cache‑miss input pricing?
When the same input is reused and fully cached by the system, it is billed at the much lower cache‑hit price, which is about one‑tenth of the standard cache‑miss price for all models. Cache‑miss pricing applies when the input is new or changed enough to bypass the cache, so using invariant system prompts or long shared context can significantly reduce your effective input cost.
What is the context length supported by Deepseek Chat models?
The deepseek‑v4‑flash and deepseek‑v4‑pro models both support a 1M‑token context window, which is extremely large compared to many other commercial LLMs and allows the model to read and reason over full codebases, long documents, or multi‑file projects in a single request. This context length is shared between input and output, with a maximum output length of 384K tokens per response.
How do I integrate Deepseek Chat with existing OpenAI clients?
You can use standard OpenAI SDKs by pointing the base URL to https://api.deepseek.com and supplying your Deepseek API key, which lets you reuse most existing OpenAI‑style code without rewriting the core logic. The API also exposes an Anthropic‑style endpoint at https://api.deepseek.com/anthropic for users who prefer that format.
What is the concurrency limit for Deepseek Chat models?
Deepseek‑v4‑flash has a concurrency limit of 2500 requests per user, while deepseek‑v4‑pro has a 500‑request concurrency limit, both enforced through the user_id parameter to isolate and rate‑limit traffic. If you exceed these limits, your requests will be throttled or rejected according to the Rate Limit & Isolation rules.
When does the 75% discount on deepseek‑v4‑pro expire?
The 75% discount on deepseek‑v4‑pro pricing is scheduled to end at 2026‑05‑31 15:59 UTC, after which the official price will be one‑quarter of the original post‑discount rate rather than the heavily discounted promotional rate. You should check the pricing page periodically because DeepSeek reserves the right to adjust product prices.
Are deepseek‑chat and deepseek‑reasoner still supported?
The model names deepseek‑chat and deepseek‑reasoner are being deprecated and will eventually be removed, with deepseek‑v4‑flash non‑thinking mode and thinking mode serving as their replacements respectively. Existing integrations using those legacy names should plan a migration path to the new v4 model names to avoid future breaks.
Does Deepseek Chat support JSON and tool‑calling modes?
Yes, both deepseek‑v4‑flash and deepseek‑v4‑pro support JSON‑mode output and tool calls, which lets you drive structured workflows where the model returns typed JSON or invokes external functions without manual string parsing. These features are designed for building automation, plugins, and API‑driven assistants that need predictable, machine‑readable responses.
How can I reduce my Deepseek Chat API costs?
You can reduce costs by reusing invariant system prompts or long context to maximize cache‑hit input pricing, avoiding redundant or verbose inputs, and carefully limiting the number of output tokens when a shorter response is sufficient. Monitoring token usage through the API token‑usage endpoints and adjusting your balance top‑up strategy based on real‑world workloads will also help control expenses.