Introduction
DeepSeek announced on May 22, 2026 that the promotional pricing for its flagship DeepSeek-V4-Pro will not expire. The discount had been running since the model's April 24 launch and was scheduled to end on May 31; instead, the company posted that it was "making our discount permanent."
This is worth stating precisely, because it is easy to report as a price cut and it is not one. Nothing got cheaper on May 22. What changed is that a temporary rate became the standing rate — which matters more than a headline reduction, since teams can now build cost models on it.
Detailed Pricing Breakdown
Per million tokens, from the DeepSeek pricing page:
- Input (cache miss): $0.435 — 75% below the $1.74 list price
- Input (cache hit): $0.003625 — roughly 120x cheaper than a miss
- Output: $0.87 — 75% below the $3.48 list price
Both rates are down by the same 75%, despite the discount being widely reported as steeper on output than on input — $3.48 to $0.87 is exactly three-quarters, the same proportion as $1.74 to $0.435.
The cache-hit rate is the number that changes agent economics. A loop that replays a long system prompt or a repository snapshot pays almost nothing for the repeated context, so the effective cost of a multi-turn coding session sits far below the headline input price.
How It Compares
Against the frontier models available at the time of publication, per million tokens:
| Model | Input | Output |
|---|---|---|
| DeepSeek-V4-Pro | $0.435 | $0.87 |
| Gemini 3.5 | $1.50 | $9.00 |
| GPT-5.5 | $5.00 | $30.00 |
That is roughly 3 to 11 times cheaper on input and 10 to 34 times on output — a wider margin than the "several times cheaper" framing the comparison usually gets. Competitor pricing moves, so treat these as a May 2026 snapshot and check the vendors' current rate cards before modelling spend.
What This Does Not Tell You
Price is the easy half of the comparison, and the announcement says nothing about the harder half.
- No benchmark claim is being made here. DeepSeek's own V4 release publishes evaluation numbers; a pricing announcement does not, and cost-per-token says nothing about how many tokens a model needs to finish your task. A cheaper model that reasons twice as long is not cheaper.
- Permanent is a statement of intent, not a contract. It removes an expiry date. It does not prevent repricing later.
- Peak-hour pricing is a live question. Reporting in June suggested DeepSeek was considering time-of-day rates for the V4 generation. None appears on the pricing page as of publication, but flat-rate assumptions are worth re-checking.
Conclusion
The substance of the May 22 announcement is narrow — an expiry date was removed — but the effect is not. Frontier-adjacent intelligence at $0.435 and $0.87 per million tokens, with a cache hit costing a fraction of a cent, is now DeepSeek's standing offer rather than a countdown. For anyone building agent loops, where token spend compounds across tool calls and retries, a rate you can plan around is worth more than a rate that is merely low.
Whether V4-Pro is the right model for your workload is a separate question, and one the pricing page cannot answer. Measure it on your own tasks.
Sources
- DeepSeek on X — the May 22 announcement making the discount permanent
- DeepSeek API pricing — current rates, including cache-hit and cache-miss tiers
- DeepSeek API changelog — the April 24 V4-Pro and V4-Flash launch entry