DeepSeek

A closer look at DeepSeek

DeepSeek offers a chat product on the web and app, a hosted API for its V4 models — deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813) — and MIT-licensed open weights for DeepSeek-V4.1-Flash-1017 you can run yourself. The API is OpenAI/Anthropic-compatible and bills per token from a prepaid balance, with separate peak and off-peak rates and cheaper cached input; those per-token rates are not a price for the chat product.

Editorial guidance based on published documentation. The workflow below is a suggested starting point, not a measured product result.

When to choose DeepSeek

• A developer wants to call a hosted model from existing OpenAI or Anthropic SDK code by changing only the base URL and model name.
• A team wants usage-based pricing it can shape: schedule heavy calls off-peak, reuse cached input, and pick Flash for cheaper high-volume calls.
• A builder needs live web pages inside an agent or RAG workflow and can call the Search and Fetch endpoints for structured results and Markdown.

When to choose something else

• A user needs image (vision) input through the deepseek-v4-pro model: the API pricing table marks Vision as not supported for deepseek-v4-pro, while deepseek-flash supports it.
• A team needs API access with no prepaid balance and no per-token metering: the API deducts every call from a topped-up or granted balance, and calls fail with 402 Insufficient Balance when the balance runs out.
• A workload needs more than 2,500 concurrent deepseek-flash requests, or more than 500 concurrent deepseek-v4-pro requests, without filing the capacity-expansion request: requests above the account-level limit receive HTTP 429.

A practical starting workflow

1. Decide the surface first: use the web chat or app for one-off questions, the API for code, or the open weights for self-hosting.
2. For the API, create an account on the open platform, create an API key, and store it as DEEPSEEK_API_KEY.
3. Point an OpenAI-compatible client at https://api.deepseek.com (or the Anthropic base URL), choose deepseek-flash or deepseek-v4-pro, and send a first chat completion; thinking mode is on by default, so switch it off only when the task does not need it.
4. Estimate the cost of one call before scaling: count input and output tokens, apply the cache-hit or cache-miss rate, and check whether the call falls in peak hours (01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays).
5. Add live web only if the task needs it: call Search for ranked pages, Fetch for one page as Markdown, keep each endpoint within its QPS limit, and top up the balance before it reaches zero.

Check before committing

• API prices are per 1M tokens and vary by model, input/output, cache hit or miss, and peak or off-peak time; the page shows $ amounts and the currency is unconfirmed, and the page says prices may change.
• The chat product's price was not stated on the pages read: do not treat the API's per-token rates as a chat subscription price.
• The granted balance is mentioned on the pricing page without an amount or eligibility rule, so do not plan on a fixed free credit; check the account's own balance page.
• Vendor benchmark figures on the model card are the vendor's own comparisons and were not independently verified here.

Common questions

Compare approaches for this task

• ChatGPT — Turning a brief or source files into structured drafts, explanations, and working documents.
• Claude — Working through source material, developing structured drafts, and iterating on code or documents.
• Ollama — Developers building with open models who want local runs with zero data retention plus a cloud option.; Users who want to start with no subscription, running models locally and paying per token for cloud models.
• Together AI — Together AI is an AI-native cloud for running open-source models: serverless and dedicated inference, batch inference, GPU clusters, code sandboxes, managed storage and fine-tuning, from experimentation to production scale.