DeepSeek V4 Pro: 75% Price Cut Becomes Permanent
On 22 May 2026, DeepSeek turned the 75% promotional discount on its flagship V4-Pro into a permanent list price. The discount was originally time-limited and set to expire on 31 May 2026 — instead, it now applies indefinitely. As a result, one million output tokens on V4-Pro cost $0.87 instead of the previous $3.48. The model itself — launched on 24 April 2026 with 1.6 trillion parameters (49 billion active), a Mixture-of-Experts architecture and roughly 1 million tokens of context — ships as open weights under an MIT license on Hugging Face. The combination of frontier ambition, open weights and an output price in the cent range is the real story here.
What actually changed
- The 75% discount on V4-Pro is permanent. What started as a time-limited promotion ending 31 May 2026 has been the regular list price since 22 May 2026 (Engadget).
- V4-Pro now costs $0.435 per million cache-miss input tokens, $0.003625 per million cached input tokens and $0.87 per million output tokens — previously four times as much.
- V4-Flash is unchanged at $0.14 input, $0.0028 cached input and $0.28 output per million tokens.
- Open weights under MIT license:
deepseek-ai/DeepSeek-V4-Proanddeepseek-ai/DeepSeek-V4-Flashare on Hugging Face — self-hostable, commercially usable.
What applied before
When DeepSeek V4 launched on 24 April 2026, the official list price for V4-Pro stood at $1.74 cache-miss input, $0.145 cached input and $3.48 output per million tokens. At launch, DeepSeek added a 75% promotional discount that pushed effective prices down to a quarter — limited until 31 May 2026. Such introductory discounts are common in the industry and are planned for: they pull early volume onto the new model, then expire, and the list price returns to normal.
That exact expectation — discount now, full price from June — did not materialize. Anyone who had budgeted around the expiry has to rethink: the promo price is the price.
What applies now
With the 22 May 2026 announcement, DeepSeek locked in three things:
1. The reduced price is the list price. The official pricing page now permanently lists V4-Pro at $0.435 cache-miss input, $0.003625 cached input and $0.87 output per million tokens. That is exactly a quarter of the original list prices. The cached-input price of $0.003625 is the real outlier on the low end: anyone using prompt caching pays almost nothing for recurring context.
2. V4-Flash stays the cheap all-rounder. The smaller, faster variant was left untouched and still sits at $0.14 input, $0.0028 cached input and $0.28 output per million tokens. For simple, high-volume tasks, Flash is therefore even cheaper than the already-discounted Pro model.
3. The gap to Western frontier models widens. According to Codersera, V4-Pro is roughly 11.5x cheaper than GPT-5.5 on input and roughly 34.5x cheaper on output. These multiples are a vendor/third-party comparison, not a benchmark result — but the point holds regardless of the exact figure: there are orders of magnitude between them, not percentages.
Context
A promotional discount becoming permanent is more telling, in business terms, than it sounds. A provider only freezes a 75% discount if the margin still holds at the low price. The implication: DeepSeek apparently assumes that V4-Pro inference at scale is cheap enough to run that even $0.87 per million output tokens leaves a profit. That is less a marketing gesture than a signal about the actual inference costs of the MoE architecture.
For Western frontier providers, this raises pressure at a sensitive point. As long as Chinese models were “cheap but weaker,” the price gap could be justified with quality. V4-Pro is an open model with frontier ambition — and independent verification of those claims is still outstanding; the benchmarks remain vendor figures for now. The trend is unmistakable nonetheless: with Qwen and MiniMax, further Chinese labs are pushing into the same segment with open weights. Open weights under permissive licenses are no longer the exception but are becoming the default expectation in this part of the market.
For an agency, the relevant question is not “best model” but “which model for which task at which price.” In that calculation, a factor of 11 to 34 moves the line considerably: tasks that were previously too expensive to automate become viable.
What you can do now
If you have high-volume, fault-tolerant tasks: Trial-move bulk work — classification, summarization, first drafts, data enrichment — to V4-Pro or V4-Flash and weigh output quality against the cost savings. At output prices of this magnitude, even a second correction pass often still costs less than a single run on a Western frontier model.
If privacy or client data is involved: Use the MIT license. The open weights allow self-hosting — sensitive data then never has to leave your own system. This is exactly the case where a hosted US model is ruled out and an open model makes the difference.
If you advise clients on “GEO” or AI budgets: Treat model choice as a moving target, not a one-time decision. Prices are currently falling in jumps, not small steps — an architecture that keeps the model swappable per task (rather than hard-wiring against one provider) is the more robust bet.
Background: How AI pricing works
Why input and output cost differently, what prompt caching buys you and how to compare token prices at all — the basics here: → AI Pricing Explained
Entdecke mehr
Usage-based instead of flat-rate: The AI cost turn was predictable — and here is how to absorb it
AI providers are pushing flat rates toward metered billing. Why it had to happen, what it costs and three levers that soften the shift.
Glossar$/MTok (Cost per Million Tokens)
Standard pricing unit for AI APIs — cost in US dollars per one million processed tokens. Listed separately for input, output and sometimes cache.
LexikonComparing AI models — who builds what and how to choose
The major model families in 2026 at a glance. Who builds Claude, GPT, Gemini, Llama, Mistral, DeepSeek, Qwen — and which model to pick when.