DeepSeek V4 Pro pricing: fourfold peak off peak analysis

Cheap API pricing is a wonderful customer acquisition tactic until the invoice starts sounding like a smoke alarm. DeepSeek V4 Pro has reached that awkward grown up moment: the API is no longer just being sold as cheap access to a capable model, it is being priced like shared infrastructure with rush hour. The fourfold increase is the headline. The peak and off peak split is the product lesson.

The launch is a price architecture Reuters reported that DeepSeek raised

API pricing for its V4 models, and Channel Insider described the V4 API increase as more than fourfold. According to DeepSeek API Docs, V4 Pro pricing is now split across off peak and peak rates, with off peak hours costing half the peak hour price. DeepSeek lists V4 Pro at $0.022 for cache hit input, $0.66 for cache miss input, and $1.98 for output per 1M tokens in off peak pricing, rising to $0.044, $1.32, and $3.96 in peak pricing. V4 Flash follows the same structure at $0.007, $0.22, and $0.66 off peak, then $0.014, $0.44, and $1.32 at peak. DeepSeek API Docs also names two V4 API models: deepseek-v4-flash serving DeepSeek-V4-Flash-0731, and deepseek-v4-pro serving DeepSeek-V4-Pro-0813. Both use a 1M context length and a maximum output of 384K, according to the same documentation. The interesting split is concurrency: DeepSeek lists a concurrency limit of 2500 for V4 Flash and 500 for V4 Pro. That is the product lineup in miniature, one lane for throughput, another for heavier capability. That concurrency gap is the tell. Pricing pages usually pretend to be accounting documents, but good ones are traffic control systems. This one tells developers that premium capacity is not infinite and that the company wants usage to move toward cheaper hours when it can.

Why V4 Pro can ask for more Artificial Analysis describes DeepSeek

V4 Pro (Reasoning, Max Effort) as an open weights model released in April 2026. The model scored 45 on the Artificial Analysis Intelligence Index, compared with a median of 26 among comparable models, and the site says it supports text input, text output, and a 1M token context window. That helps explain why DeepSeek can test a higher price ceiling without abandoning its challenger posture entirely. If the model is materially stronger, the pricing conversation shifts from cheapest token to useful token. The catch is verbosity. Artificial Analysis says DeepSeek V4 Pro generated 180M tokens in its Intelligence Index evaluation, compared with a median of 99M, and labels the model very verbose. In API economics, that is not a footnote, it is the meter spinning. A model that reasons well but talks a lot can turn generous output limits into a cost surprise for customers who do not instrument prompts, completions, and retries.

The second order effect is behavioral DeepSeek

API Docs says billing is based on the total number of input and output tokens, with input split into cache hit and cache miss categories. That distinction matters more under the new structure because the expensive path is not just using V4 Pro, it is using V4 Pro at peak, missing cache, and generating long responses. This pricing page is a Choose Your Own Adventure where the costly endings begin with uncached prompts and verbose outputs. The smart move for builders is to treat caching and response length as product requirements, not cleanup tasks. Peak and off peak pricing also changes the customer conversation. A flat API price invites teams to ask whether a model is cheap enough. A time based price asks whether the workload is urgent enough. Batch evaluations, internal analysis, and non interactive jobs can often move to cheaper windows, while user facing flows may justify peak rates if latency matters. That is a more honest model for scarce infrastructure than pretending every token arrives with the same cost profile.

What builders should copy, and what customers

should watch Reuters and Channel Insider frame the change as a price increase, which is true, but the more useful read is that DeepSeek is moving from adoption pricing to resource aware pricing. For founders and product leaders, the lesson is simple: if low pricing is your wedge, decide early what happens when usage becomes real. If the answer is a sudden fourfold bill shock, customers will feel ambushed. If the answer is a visible ladder of cache, timing, and output controls, customers can plan. For customers, the next sprint should include token level cost observability. Track cache hit rates, cache miss rates, output length, peak usage, and which features actually need V4 Pro instead of V4 Flash. The next logical move from DeepSeek is not necessarily another price change. Watch whether the official docs evolve toward clearer budget controls, caching guidance, and scheduling patterns that make the new price architecture easier to operationalize.