What Is DeepSeek V4 Pro? Flash vs Pro, Web vs API

Introduction

Search DeepSeek V4 Pro and many people treat it as another official site, a desktop client, or the same thing as V4 Flash. It is not.

DeepSeek V4 Pro is the flagship model in DeepSeek’s V4 line. The API ID is deepseek-v4-pro: about 1 million tokens of context, full vision, heavier reasoning. It is not a new domain, not Flash, and not the DeepSeek API itself. Web chat is still free to try; you only need the official open platform when you wire the model into your own product. This site is an independent community. Specs and prices follow the official API docs.

I. What DeepSeek V4 Pro is

Treat it as the heavy tier in the V4 family, not a second brand:

• Model ID: deepseek-v4-pro. Do not look for deepseek-v4 or deepseek-vision, and do not reuse retired names deepseek-chat / deepseek-reasoner.
• Size: the public 671B MoE flagship; context about 1M tokens; vision on the full path.
• This site’s features snapshot lists about 45 tok/s — an observation, not an SLA.
• Where weights are published, the license is MIT. On the hosted API you send the model string and a key from the official platform.

The current public build is often tagged DeepSeek-V4-Pro-0813 (mid-August 2026 capability update). The ID did not change; it is still deepseek-v4-pro. If you worry about a September 14 routing cutoff, see Is DeepSeek V4 Pro still available: the API continues with billing unchanged. Specs live on the DeepSeek V4 Pro product page.

Use it for long PDFs / multi-file repos, hard reasoning, charts and UI screenshots, and agent jobs you would map to an Opus-class tier. Everyday short Q&A and high-volume drafts should not default to the flagship.

II. How to choose versus V4 Flash

Same OpenAI-compatible endpoint https://api.deepseek.com/v1 — you only change the model name.

What you are doing	Which to use	Why
Daily chat, mail edits, small patches, high concurrency	V4 Flash (deepseek-v4-flash)	About 128K context, faster and cheaper, higher concurrency cap
Long documents, hard reasoning, full vision, long agent chains	DeepSeek V4 Pro	About 1M context, full vision, steadier reasoning
Check whether this model can do your job	DeepSeek web app	No model ID to type; chat stays free
Wire it into your own app or scripts	DeepSeek API docs	Keys only from the official platform

One line: Pro is the brain, Flash is daily throughput, the web app is the trial door, the API is the meter. If you see a time-boxed model ID (e.g. DeepSeek V4.1 Flash beta), treat it as a short eval endpoint — not a production default: DeepSeek V4.1 Flash beta. Sign-up and key steps are in the web app and API beginner guide — keep that tutorial separate from this chooser.

III. Web app, API, official site: three doors

On this query, SERPs mix the official site, .com.cn download properties, and explainer pages like this one. The wrong door feels like a different product:

1. Official site / open platform: source of the brand and of keys. API keys, quotas, and official pricing live only there.
2. DeepSeek web app: chat in the browser. Drop a real PDF, a real repo question, or a screenshot first. This site points to that entry; it is not the official DeepSeek site.
3. DeepSeek API: set model to deepseek-v4-pro or deepseek-v4-flash. Chat Completions hit https://api.deepseek.com; the Anthropic-compatible path is https://api.deepseek.com/anthropic.

Self-hosting and Hugging Face weights are a third track: you need the repo only for local inference, not as a substitute for the hosted API. If the agent must edit files and run a shell, that is DeepSeek Harness, not another model name.

IV. Three things to check after August 2026

1. Spell the ID. New projects use only deepseek-v4-pro / deepseek-v4-flash.
2. Web chat stays free; the API now bills peak vs off-peak. Beijing time 09:00–12:00 and 14:00–18:00 are peak; the rest is off-peak (half price), from 17 Aug 2026. The line with the largest percentage jump is Pro cache-hit input, not “the whole bill × 11”. Breakdown: DeepSeek API 17 Aug price change.
3. Shift what you can; cache what you can. Live support has to sit on peak. Nightly batch, evals, and cleanup belong off-peak. Reused long system prompts still win on cache, but do not budget the old 0.025 CNY rate.

If official numbers move again, the pricing page wins. This site does not issue keys and does not collect fees.

V. Common traps

• Treating “DeepSeek V4 Pro” as another official domain and hunting for an installer. You pick the model in the web app or the API — not by downloading another client.
• Sending real vision work to Flash, then blaming “V4” for unread charts. Full vision is on Pro.
• Copying ~45 tok/s into a contract. Measure your own traces. DSpark only accelerates the model you already called; it is not a third ID.
• Looking for an API key on this site. Keys are issued only on the official platform.

FAQ

Are DeepSeek V4 Pro and the DeepSeek web app the same thing? No. The web app is the door; DeepSeek V4 Pro is the flagship model behind it. Try one real task in the browser, then decide whether to hard-code deepseek-v4-pro.

Can I use DeepSeek V4 Pro without an API key? Yes, via free web chat. To call it from your own service, apply for a key on the official platform, then POST to https://api.deepseek.com/v1.

Is this official documentation? No. This is an independent community chooser. Model IDs, rate limits, and prices follow the official docs.

Summary

Four facts: DeepSeek V4 Pro is the deepseek-v4-pro flagship, not a new website; Flash differs on context, vision, and reasoning weight; the web app is still free to try; the API bills peak/off-peak from 17 Aug. Open the DeepSeek web app on a long job first, then lock the ID.

Further reading

• DeepSeek V4 Pro product page
• DeepSeek API docs
• DeepSeek API 17 Aug price change
• Web app and API beginner guide