ToolStep AI tools, ranked & compared
Explore Tools
AI Pricing

DeepSeek Flash Price Cut: New Lower API Pricing Starts September 10, 2026

Published September 9, 2026 · ToolStep Editorial Team · Based on DeepSeek's official pricing announcement

DeepSeek is cutting Flash-series API prices effective September 10, 2026 at 12:00 Beijing time (UTC+8). The adjustment applies to the Flash series — including deepseek-v4-flash and deepseek-v4-flash-vision-exp — and it is a meaningful one: cache-hit input drops 60%, cache-miss input drops roughly a third, and output gets cheaper too. This is not a site-wide DeepSeek price cut; it is specific to the Flash tier.

What changed, in one line: from September 10, 2026, running the Flash models off-peak costs ¥1 per million cache-miss input tokens and ¥4 per million output tokens — with cache hits at just ¥0.02 per million tokens.

The Exact New Pricing (Off-Peak)

Per the official announcement, the new off-peak prices per one million tokens are:

Token typeOld priceNew price (from Sep 10)Change
Cache hit input¥0.05¥0.02−60%
Cache miss input¥1.5¥1−33.3%
Output¥4.5¥4−11.1%

Peak Hours: The 2x Multiplier

During peak hours, all three new prices are doubled. That means cache-hit input at ¥0.04, cache-miss input at ¥2, and output at ¥8 per million tokens during peak windows. This peak/off-peak split is the most important caveat in the announcement: the headline prices are the floor, not the ceiling. Batch jobs, scheduled evaluations, and background processing should target off-peak windows to actually realize the savings.

Why the Cache-Hit Price Cut Matters Most

The 60% cut on cache-hit input is the standout change, and it is aimed directly at how modern AI applications actually consume tokens. Any workload that re-sends similar context — multi-turn chat, agent loops that carry system prompts and tool results forward, RAG pipelines with stable knowledge bases, coding assistants working inside one repository — hits the prompt cache constantly. On those workloads, cache-hit tokens can make up the majority of billed input, so a 60% reduction on that line changes the effective cost per request far more than the headline percentages suggest.

Concretely: an agent loop that sends a 50,000-token system-plus-history context and gets an 80% cache-hit rate now bills 40,000 tokens at the ¥0.02 rate instead of ¥0.05. At volume, that is the difference between a marginal agent prototype and one that fits a budget.

What This Means for API Users and Developers

Three practical implications from the announced pricing:

1. Flash-tier workloads get cheaper without code changes. If your application already calls deepseek-v4-flash or deepseek-v4-flash-vision-exp, the new prices apply automatically from September 10 — no migration, no API version bump.

2. Cache-friendly architecture is now worth more. Structuring prompts so that stable content (system prompts, documents, tool definitions) sits early in the context maximizes cache hits, which are now 2.5x cheaper than cache misses. This rewards RAG designs and long-running agent sessions over stateless one-shot calls.

3. Off-peak scheduling becomes a real lever. With peak pricing at 2x, moving nightly batch runs, evaluation suites, and non-urgent generation to off-peak windows roughly halves the bill compared with running the same jobs at peak.

Who Benefits Most

Agent and assistant builders benefit most, because their workloads are cache-heavy and run continuously. RAG pipelines with stable document sets come next. High-volume summarization and classification jobs also gain directly from the cheaper input side. Buyers who should look elsewhere: workloads that need the strongest reasoning quality regardless of price — the Flash tier is a speed-and-cost play, and this adjustment does not change that positioning.

Should You Change Model or Usage Strategy?

If you are already on the Flash series: stay, and tune for cache hits and off-peak scheduling to capture the full savings. If you are on a pricier tier for cost reasons alone: the new Flash pricing makes a downgrade worth testing on your real traffic. If you need frontier reasoning quality for hard problems: this announcement does not change that calculus — price cuts on a flash-tier model do not make it a frontier model. And if you were waiting on pricing before building on DeepSeek's flash tier: September 10, 2026 is the date the economics improve.

Caveats to Keep in Mind

ToolStep Editorial Takeaway

This is a competitive, usage-focused price cut aimed at the workloads that are actually growing: agentic and context-heavy applications. The cache-hit reduction is the strategically interesting part — it prices in the reality that repeated context is the dominant cost driver for agents — and the off-peak discount gives operators a genuine cost lever. For developers on the DeepSeek Flash tier, the practical move is simple: keep building, structure prompts for cache hits, and schedule what you can off-peak.

Sources

Sources:

This article is based on DeepSeek's official announcement and published pricing documentation. ToolStep did not receive compensation for this coverage, and no hands-on benchmarking was performed — this is a pricing and strategy analysis, not a model performance test.

Related on ToolStep