Backend

DeepSeek's New Peak Pricing Hits Cache-Heavy Workloads with a 12× Increase

By 码事漫谈 · Aug 13, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A 12× cost spike on cache hits rewrites the economics of prompt-caching strategies that many RAG and agent architectures depend on. Teams running DeepSeek in production have two days to re-audit their prompt design and call scheduling before the new rates lock in.

Summary

DeepSeek's V4-Flash and newly semi-launched V4-Pro models will switch to a two-tier pricing structure on August 17, 2026, with peak rates applying during Beijing working hours. The most dramatic change is a 12× increase for V4-Pro cache hits during peak times, jumping from 0.025 to 0.30 CNY per million tokens. Input cache misses see a modest 1.5× increase, but output costs rise by at least 2.25× across both models.

A typical 1M-input, 1M-output call on V4-Pro will cost 36 CNY during peak hours, four times the old price of 9 CNY. The pricing shift effectively ends Flash's reputation as a dirt-cheap option, with its peak cache-hit price rising from 0.02 to 0.10 CNY. RAG pipelines, agent orchestrations, and any workflow relying on long prefix caching will feel the cost pressure most acutely.

The update also leaves concurrency limits unstated, where Flash previously allowed 2,500 and Pro allowed 500. If those caps are tightened alongside the price hike, high-throughput applications face a double squeeze. The time-of-use model signals a broader industry move toward charging more during working hours, pushing batch processing toward overnight schedules.

Takeaways

— V4-Pro cache hits during peak hours jump from 0.025 to 0.30 CNY per million tokens, a 12× increase.

— V4-Flash peak cache hits rise from 0.02 to 0.10 CNY, erasing its ultra-low-cost positioning.

— Output costs increase by at least 2.25× for both models, hitting long-reasoning and batch-generation workloads hardest.

— Input cache misses see the smallest increase at 1.5× across both Flash and Pro.

— A 1M-input, 1M-output V4-Pro call during peak hours costs 36 CNY, up from 9 CNY.

— Peak hours are defined as Beijing time 9:00–12:00 and 14:00–18:00, covering the standard workday.

— The new pricing table omits concurrency limits, which were previously 2,500 for Flash and 500 for Pro.

— Batch-processing scripts should be rescheduled to overnight idle periods to avoid peak rates.

Conclusions

DeepSeek is using cache-hit pricing as the sharpest lever to segment customers: high-value agent and RAG workloads pay a premium, while simple one-shot calls are nudged toward free-tier queuing.

The decision to leave concurrency caps unstated on the new pricing table creates uncertainty that may be intentional, letting DeepSeek throttle heavy users without committing to a public number.

Time-of-use API pricing, already standard in cloud infrastructure, is now reaching the model layer. Other providers will likely follow, making overnight batch windows a standard cost-optimization tactic.

Concepts & terms

Cache Hit / Cache Miss

In LLM APIs, a cache hit occurs when part of the input prompt matches a previously computed prefix, allowing the provider to skip recomputation and charge a lower rate. A cache miss means the input must be processed from scratch at the full input rate.

Prefix Caching

A technique where a long, static prefix (such as a system prompt or few-shot examples) is reused across many API calls. The provider stores the computed state for that prefix and only charges the lower cache-hit rate for subsequent calls that share it.

Time-of-Use Pricing

A billing model where rates vary by time of day, charging higher prices during peak demand periods and lower prices during off-peak hours. Common in electricity markets and cloud computing, it is now appearing in LLM API pricing.

From the discussion

The pricing change caught developers off guard, with expectations of a moderate increase shattered by the actual 12× jump. One pragmatic response is simply to reduce usage, especially for low-quality or unreadable code that gets run anyway.

— A 12× price increase far exceeded the anticipated 70% hike.

— Cutting back on usage is a practical reaction, particularly for code that is already of poor quality.

Featured comments

abc897

I was thinking at most a 70% increase, didn't expect it to jump more than tenfold.

See top comments, translated →

Source: juejin.cn ↗ Google Translate ↗ Backup ↗