---
title: "DeepSeek in 2026: from the $6 million shock to the open-weight V4"
description: "DeepSeek knocked 18% off Nvidia in a day in January 2025. In 2026 it ships the open-weight V4 family at a fraction of US prices. What it does well, what it costs, and the data question that has not gone away."
author: "Casper Larsen"
published: "2025-09-01"
updated: "2026-08-28T00:00:00+00:00"
canonical: "https://www.casperlarsen.dk/insights/deepseek-den-kinesiske-ai-revolution-der-rystede-silicon-valley"
---

# DeepSeek in 2026: from the $6 million shock to the open-weight V4

> **Key takeaway:** DeepSeek went from a $6 million shock in January 2025 to the open-weight V4 family in 2026: V4-Pro (1.6 trillion parameters, MIT licence) and V4-Flash at $0.14 per million input tokens. The quality gap to the US labs has mostly closed; the data-residency question (your data in China, unless you self-host) has not.

DeepSeek knocked 18% off Nvidia in a day in January 2025. In 2026 it ships the open-weight V4 family at a fraction of US prices. What it does well, what it costs, and the data question that has not gone away.

In January 2025 a Chinese lab few people had heard of released a model that cost about $6 million to train, matched the best from OpenAI, and knocked 18% off Nvidia's share price in a single day. Eighteen months later DeepSeek is a fixture, not a shock. Here is where it stands in August 2026.

## From V3 to V4

The V4 family arrived as an open-weight preview on 23 April 2026 under the MIT licence, then went to general availability over the summer: V4-Flash on 31 July and V4-Pro on 13 August. V4-Pro is a 1.6 trillion parameter mixture-of-experts model with a one million token context window; V4-Flash is the 284 billion parameter fast variant. Both are open weights, so anyone can inspect and run them.

There is no R2. The reasoning model is still R1 (updated May 2025), and the search volume around "R2" is anticipation, not a release.

DeepSeek also released a harness, dsh, under MIT, following the industry pattern that the system around the model matters as much as the model.

## Pricing

This is still the headline. Per million tokens, input/output:

- V4-Flash: $0.14 / $0.28
- V4-Pro: $0.435 / $0.87 at standard rates

For comparison, the US flagships sit at $2-5 input and $12-30 output. DeepSeek is roughly a tenth of the price for quality that independent benchmarks now put close to the leaders.

One change to watch: on 13 August 2026 DeepSeek introduced peak and off-peak pricing, with peak output rates several times the standard rate. If you run batch jobs, schedule them off-peak.

## Where it fits

- Coding and technical tasks, where it has been strong since V3.
- High-volume pipelines: classification, extraction, summarisation at scale.
- Self-hosting: the open weights mean a company can run V4-Flash on its own hardware with no data leaving the building.

## The question that has not gone away

Use DeepSeek's hosted API and your data goes to servers in China. For most Danish companies, under GDPR and with customer data in the prompts, that rules the hosted service out.

The open weights change the picture. Run V4 on European infrastructure (your own, or a European provider) and the data-residency problem disappears; you get the price without the geography. That is how most serious European deployments of DeepSeek work in 2026.

## What it means

The 2025 lesson held: top-tier models no longer require a US budget, and the gap between open and closed models has mostly closed. The 2026 lesson is the second half of that: the cheapest capable model is now open, and the decision about where it runs is yours.

*Updated August 2026. Originally published September 2025.*

---
Casper Larsen, Fractional Chief AI Officer, Denmark. Contact: https://www.casperlarsen.dk/contact · Free pre-flight check: https://www.casperlarsen.dk/audit