DeepSeek V4 has just reached general availability, and its pricing changes the equation for any SME paying monthly AI API bills. The Chinese lab permanently retired its legacy models on July 24, 2026, in favor of a V4 family that is up to ten times cheaper than GPT-5.6 or Claude Sonnet 5 on some workloads. A real budget win, but one that comes with a question less often asked: can you legally send European customer data to it?
In short
- DeepSeek V4 has been generally available since July 20, 2026; the legacy identifiers
deepseek-chatanddeepseek-reasonerwere retired on July 24, 2026 at 15:59 UTC (source: official DeepSeek documentation). - Two versions are available: V4 Flash ($0.14 / $0.28 per million input/output tokens) and V4 Pro ($0.435 / $0.87), both with a 1-million-token context window.
- For comparison, Claude Sonnet 5 (Anthropic) costs $2 / $10 per million tokens at launch pricing, and GPT-5.6 Sol (OpenAI) costs $5 / $30 (sources: Anthropic, OpenAI).
- The real risk is not technical, it is legal: China has no adequacy agreement with the European Union, and several data protection authorities (Italy, France, Ireland) have opened inquiries into DeepSeek's practices.
- For an SME, the sensible approach is simple: test DeepSeek V4 on tasks with no personal data, never on customer or HR files until proper contractual safeguards are in place.
What exactly is DeepSeek V4?
DeepSeek V4 is the new generation of models from the Chinese lab DeepSeek, known for delivering performance close to top American models at a fraction of the price. The V4 family was first offered in preview starting in April 2026, before reaching general availability on July 20, 2026, according to the provider's official documentation.
Two variants coexist today:
- DeepSeek V4 Flash: the fast, economical model, built for volume (customer support, summarization, data extraction).
- DeepSeek V4 Pro: the most capable model, for complex reasoning and code generation.
Both share a 1-million-token context window (roughly several hundred pages of text in a single request) and a maximum output of 384,000 tokens. Since July 24, 2026, the legacy model names deepseek-chat and deepseek-reasoner, used by many third-party tools, no longer work: anyone still relying on them had to migrate to deepseek-v4-flash or deepseek-v4-pro before that date.
The real argument: price
What sets DeepSeek V4 apart is not so much raw power as its price-to-performance ratio. On coding and reasoning tasks, several independent analyses place V4 Pro roughly on par with GPT-5.6 Terra, at a fraction of the cost. Here is how the official rates of the three main providers compare on their mid-tier model, as of July 2026.
| Model | Input ($/M tokens) | Output ($/M tokens) | Context |
|---|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.28 | 1M tokens |
| DeepSeek V4 Pro | $0.435 | $0.87 | 1M tokens |
| Claude Sonnet 5 (launch) | $2 | $10 | Not specified |
| GPT-5.6 Luna | $1 | $6 | 1.05M tokens |
| GPT-5.6 Sol | $5 | $30 | 1.05M tokens |
Sources: official DeepSeek documentation, Anthropic (Claude Sonnet 5 announcement), OpenAI (API pricing page).
In practical terms, for an SME running a customer support agent that generates the equivalent of 5 million output tokens a month, the bill drops from around $150 with GPT-5.6 Sol to under $5 with DeepSeek V4 Pro. This kind of gap explains why more and more software vendors now offer DeepSeek as a "budget" option alongside GPT or Claude.
The risk the pricing tables do not show
DeepSeek is a Chinese company, hosted in China. That is where the real question lies for a European SME. The European Commission has never granted an adequacy decision to China: in theory, transferring the personal data of European residents to Chinese servers without proper contractual safeguards breaches Article 46 of the GDPR.
This is not a theoretical concern. Since early 2025, several data protection authorities have opened cases involving DeepSeek:
- Italy's authority (Garante) blocked access to the DeepSeek app in its territory, judging the provider's privacy policy insufficient regarding data processing.
- France's CNIL requested further information on how the model handles data.
- Ireland's authority (DPC) questioned the company about the collection of Irish residents' data.
Remember before migrating
Never send customer, employee, or partner data to DeepSeek's hosted API without validation from your data protection officer. The open-weight model itself can be self-hosted on European infrastructure: that is the only option that resolves the data-transfer issue at the source.
None of this diminishes DeepSeek V4's technical interest. It simply restates a rule that enthusiasm for low prices sometimes obscures: the cost of an AI model is not only its price per token. The cost of a compliance failure can be far higher than any API bill.
How to evaluate a low-cost AI model safely
- Separate use cases. Reserve DeepSeek V4 for tasks with no personal data: generic drafting, code, summaries of public documents.
- Check the actual hosting. An open-weight model (as DeepSeek V4 is) can run on a European server if a cloud provider offers it: the GDPR question then changes entirely.
- Demand a clear contract. No commercial agreement replaces a proper DPA (Data Processing Agreement) and GDPR-compliant standard contractual clauses.
- Compare, do not switch overnight. Test DeepSeek V4 alongside your current tool before any full migration, to compare quality, latency, and real cost on your own use cases.
FAQ
Is DeepSeek V4 really cheaper than ChatGPT or Claude?
Yes, based on official rates published in July 2026: DeepSeek V4 Pro costs about $0.87 per million output tokens, versus $10 for Claude Sonnet 5 and up to $30 for GPT-5.6 Sol. The gap narrows against entry-level models like GPT-5.6 Luna ($6), but remains significant.
Can an EU SME legally use DeepSeek V4?
Yes, but with caution. Using DeepSeek's hosted API for non-personal data poses no particular issue. Sending European customer or employee data without contractual safeguards exposes the company to a GDPR compliance risk, in a context where several European authorities have already opened inquiries into this provider.
What is an "open-weight" model?
It is a model whose trained parameters are published and freely downloadable, unlike a closed model accessible only via a remote API. An open-weight model can be installed on your own server or a European cloud provider's, letting you keep control over where data is hosted.
Should an SME rush to migrate to DeepSeek V4 to cut AI costs?
No, not hastily. The price advantage is real, but migration should happen gradually: start with use cases free of sensitive data, measure output quality on your own tasks, and validate any compliance question with your data protection officer before extending usage.
DeepSeek V4 confirms a defining trend of 2026: the cost of generative AI keeps falling, sometimes dramatically. For an SME, that is good budget news, provided data compliance is not treated as an afterthought. To go further on choosing the right AI model for your business, browse our other Mag resources.
