DeepSeek V4 Pro launches with fourfold price increase as AI demand strains infrastructure capacity

DeepSeek V4 Pro Launch Marks End of Ultra-Low Pricing Era

DeepSeek has officially launched DeepSeek-V4-Pro-0813, ending the promotional pricing that made the Chinese AI startup a disruptive force in the market1

1

. The general availability version supersedes the preview model released in April and introduces substantial changes to the company's pricing structure3

3

. Starting at 16:00 UTC on August 16, API pricing for the V4 model family will increase by notable margins, with some rates rising more than 1,100%1

1

. The V4 Pro will cost $3.96 for 1 million output tokens at peak hours, more than four times the current rate of $0.872

2

. Input tokens are priced at $0.435 per million4

4

. DeepSeek is introducing peak and off-peak rates to allocate resources more reasonably as AI demand strains infrastructure capacity2

2

.

Peak and Off-Peak Rates Offer Pricing Flexibility

The new pricing structure includes off-peak rates at half the peak-hour price, encouraging developers toward more flexible workload scheduling1

1

. During off-peak hours, V4 Pro output tokens will cost $1.98 per million, while the V4-Flash model will be available for $0.66 per million output tokens compared to $1.32 during peak hours2

2

. The current API pricing was originally supposed to be a promotion ending on May 31, but DeepSeek announced that month it would make the discounted prices permanent before ultimately deciding to proceed with the price hikes2

2

. DeepSeek had offered a 75% promotional discount on V4 Pro through May 5, making the pricing shift a notable reversal of the company's earlier strategy3

3

.

Cost-Effective AI Models Still Undercut Premium Competitors

Despite the fourfold increase, DeepSeek remains significantly cheaper than many competitors in the AI model pricing landscape2

2

. Kimi K3, developed by Chinese company Moonshot, costs $15 per 1 million output tokens, while Anthropic's Fable 5 charges $50 per million output tokens3

3

. OpenAI's most advanced model, GPT-5.6 Sol, costs $30 for 1 million output tokens2

2

. DeepSeek V4 Pro is up to 57 times cheaper than Fable 5 while delivering nearly identical results in key benchmark scores4

4

. OpenAI's low-cost offering, GPT-5.6 Luna, costs $1.20, which is less than the DeepSeek V4-Flash at peak hours but more expensive than the off-peak rate2

2

.

Agentic AI Performance Shows Significant Improvements

The V4-Pro-0813 focuses on agent capabilities, where AI systems use tools, execute code, and complete multi-step workflows without human intervention3

3

. Benchmark scores demonstrate substantial improvements over the preview version across agentic AI performance metrics5

5

. The model scored 87.9 on Terminal Bench 2.1 compared to 72.1 for V4 Pro Preview, while achieving 62.7 on DeepSWE and 61.5 on NL2Repo3

3

5

5

. On the CyberGym benchmark, it achieved 83.3, narrowly surpassing Fable 5's score of 83.14

4

. The model also scored 74.1 on Toolathlon-Verified, 25.7 on Agents' Last Exam, and 31.8 on AutomationBench, outperforming Fable 5's 29.14

4

5

5

.

Enhanced Features Include OpenAI Responses API Support and Codex Integration

The V4 Pro API has been updated to work with the OpenAI Responses API format out of the box, providing native compatibility for developers3

3

. Built-in Codex support allows for one-click setup, optimizing the model for coding workflows5

5

. Thinking effort levels for both V4 Pro and V4-Flash have been expanded to three settings: low for simple tasks, high for daily agent workflows, and max for complex tasks requiring more deliberation3

3

5

5

. The model is capable of handling a context window of up to 1 million tokens and can produce outputs as long as 384,000 tokens, with the option to run in either thinking or non-thinking mode3

3

.

DSpark Speculative Decoding Enhances Performance

DeepSeek-V4-Pro-0813 includes DSpark speculative decoding, a module built on the V4 Pro Preview architecture that improves inference efficiency5

5

. Developers can enable DSpark with vLLM using seven speculative tokens with greedy draft sampling, or through SGLang without specifying a separate draft model path since target and draft weights come from the same checkpoint5

5

. For local deployment, DeepSeek recommends temperature settings of 1.0 with top-p at 0.95 for agentic scenarios and top-p at 1.0 for other scenarios, with maximum output length of 384K tokens for high and max reasoning modes5

5

.

Market Competition Intensifies as Rivals Close Gap

DeepSeek's R1 model went viral in early 2025 and briefly gave the company a commanding position in the AI race, but competitors including Moonshot AI, Alibaba, and ByteDance have since closed the gap with their own releases3

3

. The launch comes after the cheaper V4-Flash model performed above expectations in independent tests, in some cases outpacing the April preview of V4 Pro, drawing attention given that Pro is positioned as the more capable product3

3

. Models like SpaceX's Grock 4.6, priced at $2 per million input tokens and $6 per million output tokens, offer competitive performance and match GPT-5.6 Sol with a score of 61 on the Artificial Analysis Index4

4

. The company secured roughly $7.4 billion in its first round of outside capital in June, and was reportedly in discussions for an additional round that would value it at around $74 billion3

3

. The model weights are released under the MIT License, maintaining accessibility for developers5

5

.