DeepSeek Makes Its 75% V4 Pro Discount Permanent
DeepSeek has made its 75% discount on DeepSeek V4 Pro permanent. When the promotion ends on May 31, the list price drops to a quarter of the launch price and stays there: $0.435 per million input tokens (cache miss) and $0.87 per million output, down from $1.74 and $3.48.
For a model that scores 9.0 on LMRank, that is 11x cheaper than GPT-5.5 on input and 34x cheaper on output. The pricing question for most teams is no longer whether a Chinese open-weight model is good enough. It is whether you trust this price to last, and whether your data can go to DeepSeek's API.
Update, September 2026: "permanent" lasted less than three months. With the official V4 release, DeepSeek introduced peak and off-peak billing on August 16 to 17. V4 Pro now lists at $0.66/$1.98 off-peak and $1.32/$3.96 at peak on DeepSeek's pricing page. See DeepSeek Peak Pricing.
The numbers
| Model | LMRank | Input $/M | Output $/M | Weights |
|---|---|---|---|---|
| Claude Opus 4.7 | 9.6 | $5.00 | $25.00 | Closed |
| GPT-5.5 | 9.4 | $5.00 | $30.00 | Closed |
| Qwen3.7 Max | 9.0 | $2.50 | $7.50 | Closed |
| GPT-5.4 | 8.9 | $2.50 | $15.00 | Closed |
| DeepSeek V4 Pro | 9.0 | $0.435 | $0.87 | MIT |
DeepSeek also cut cache-hit input prices to a tenth of their original level on April 26, which puts V4 Pro cache hits at about $0.0036 per million. For agents that resend the same system prompt and repository context every turn, that line can matter more than the headline input price.
What a month of traffic costs
Take a workload of 100M input and 20M output tokens per month, with no caching:
- GPT-5.5: $500 + $600 = $1,100
- Claude Opus 4.7: $500 + $500 = $1,000
- Qwen3.7 Max: $250 + $150 = $400
- DeepSeek V4 Pro: $43.50 + $17.40 = $60.90
The gap is large enough that even a 2x token-efficiency disadvantage on DeepSeek's side barely dents it.
Why the price may not hold
DeepSeek frames the cut as a structural efficiency gain from its sparse-attention MoE architecture, and the efficiency is real: V4 Pro needs a fraction of V3.2's per-token compute at long context. But DeepSeek's own V4 launch notes said Pro's throughput was limited by high-end compute, and that pricing would fall further once Huawei's Ascend 950 supernodes shipped in volume. A price that depends on future hardware deliveries is a commitment with conditions attached.
What you give up
- Quality at the top: 9.0 vs 9.6 for Opus 4.7. On the hardest agentic and reasoning tasks, the gap is visible.
- Instruction precision: strict formatting and structured output need more validation than with Claude or GPT.
- Jurisdiction: DeepSeek's API is operated from China. For many regulated workloads that settles the question before price comes up. Self-hosting the MIT weights is the alternative, but a 1.6T model needs a serious GPU cluster.
Takeaway
Move cost-sensitive, non-regulated traffic to V4 Pro and keep the integration portable. Put the model behind a router or a thin abstraction, keep your evals running against at least one alternative, and assume the price can change with a month's notice. Cheap tokens are worth using. Building around them as if they were guaranteed is the mistake.
Related: Cheapest Models · Best Open-Weight Models