AI agents · OpenClaw · self-hosting · automation

Quick Answer

DeepSeek V4 Peak-Valley Pricing Explained (July 21, 2026)

Published:

DeepSeek V4 Peak-Valley Pricing Explained (July 21, 2026)

DeepSeek V4 shipped GA on July 19, 2026 with a first-of-its-kind pricing model for frontier AI: peak-valley rates. Off-peak API calls cost roughly half what peak calls cost — a mechanism borrowed from electricity grid economics that’s never been applied to LLM APIs at this scale before.

For high-volume workloads, this pricing change alone is more consequential than the model quality improvements between V4 Preview and GA. Batch inference workloads can now be scheduled around off-peak windows for effective 50% cost reductions on top of DeepSeek’s already-dramatic pricing advantage over GPT-5.6 Sol and Claude Fable 5.

Last verified: July 21, 2026

The Peak-Valley Pricing Structure

DeepSeek V4-Pro (1.6T total / 49B active params)

Token typeOff-peak (normal)Peak
Output tokens$0.87/MTok$1.74/MTok
Input tokens (cache miss)$0.435/MTok$0.87/MTok
Input tokens (cache hit)$0.043/MTok (est.)$0.087/MTok (est.)

DeepSeek V4-Flash (284B / 13B active params)

Token typeOff-peak (normal)Peak
Output tokens$0.28/MTok$0.56/MTok
Input tokens (cache miss)$0.14/MTok$0.28/MTok
Input tokens (cache hit)$0.0028/MTok$0.0056/MTok

Peak windows: 1:00-4:00 AM UTC and 6:00-10:00 AM UTC (aligned with China business day).

Off-peak windows: All other times — including US business hours 13:00-21:00 UTC (9 AM-5 PM Eastern).

Why This Matters

The Cache-Hit Number Is Absurd

V4-Flash cache-hit input at $0.0028/MTok is the cheapest frontier-adjacent input pricing that has ever existed on a public API. For workloads with high prompt reuse (RAG systems querying the same document corpus, agent workflows re-reading the same system prompt on every turn), this pricing collapses input costs to essentially zero.

Real example: A production RAG system doing 10M cache-hit input tokens per day costs $28/day at V4-Flash off-peak vs $150/day on Claude Fable 5 with prompt caching enabled — a 5x delta on top of the base pricing advantage.

Batch Workloads Save 50%

For workloads that can be scheduled (embedding backfills, batch summarization, agent test suite runs, overnight report generation), moving from peak to off-peak windows delivers a straight 50% cost reduction. Combined with V4’s already-dramatic pricing vs GPT-5.6 Sol and Claude Fable 5, this makes DeepSeek V4 the default choice for high-volume production inference.

Interactive Traffic Is Off-Peak by Default

US business hours (9 AM-5 PM Eastern = 13:00-21:00 UTC) are entirely off-peak. This means US-based interactive workloads (chat interfaces, coding assistants, agent loops) get the cheaper rate automatically without any scheduling work.

European business hours (9 AM-5 PM CET = 8:00-16:00 UTC) partially overlap with the 6:00-10:00 UTC peak window. Interactive traffic during European mornings hits the peak rate. This may be intentional cross-subsidy — Chinese and East Asian traffic pays the peak rate during their business day, subsidizing off-peak US traffic.

Migrating from Legacy Models

Deadline: July 24, 2026 — DeepSeek-Chat and DeepSeek-Reasoner model names retire.

Migration:

  • deepseek-chat → deepseek-v4-flash-nonthinking (or deepseek-v4-flash depending on client library)
  • deepseek-reasoner → deepseek-v4-flash-thinking

For higher capability, migrate to deepseek-v4-pro (thinking or non-thinking modes) — same API surface, more capability at higher price point.

Action items before July 24:

  1. Update model name strings in all production code.
  2. Re-benchmark critical prompts on V4-Flash vs V4-Pro (Pro is 3x pricier but meaningfully more capable on hard tasks).
  3. Enable prompt caching if not already on — the cache-hit pricing on V4-Flash is too cheap to ignore.
  4. Instrument peak vs off-peak spend to understand your natural distribution.

How V4 Pricing Compares to Frontier Alternatives

Output token cost per MTok (July 21, 2026)

ModelInputOutputNotes
DeepSeek V4-Flash (off-peak)$0.14$0.28Cheapest by 10x
DeepSeek V4-Flash (peak)$0.28$0.56Still cheapest
Gemini 3.5 Flash$0.30$2.50Google’s default
DeepSeek V4-Pro (off-peak)$0.435$0.87Frontier-adjacent at Flash-tier pricing
DeepSeek V4-Pro (peak)$0.87$1.74Still 5-10x cheaper than Sol/Fable
Claude Sonnet 5$2$10Anthropic’s mid-tier
GPT-5.6 Luna$3$12OpenAI’s mid-tier
Kimi K3$3$15Frontier-tier open weights
GPT-5.6 Sol$5$15OpenAI flagship
Claude Fable 5$10$50Anthropic flagship (Max/Team only after July 20)

Key ratios:

  • V4-Pro off-peak vs Sol: 17x cheaper output.
  • V4-Flash off-peak vs Sol: 54x cheaper output.
  • V4-Flash cache-hit input vs Fable 5 input: ~3500x cheaper.

Who Should Switch

Switch to V4-Flash now if:

  • Your workload is mostly summarization, extraction, or straightforward Q&A.
  • You’re doing high-volume batch inference.
  • Prompt caching is achievable (RAG, agent loops, repeated system prompts).
  • Cost matters more than absolute-best benchmark scores.

Switch to V4-Pro now if:

  • You want frontier-adjacent capability at ~10x lower cost than Sol.
  • MIT-licensed open-weight fallback matters for compliance.
  • Your workload includes complex reasoning or coding.
  • You want to hedge against Sol/Fable pricing increases.

Stay on Sol/Fable 5 if:

  • Your workload hits the very hardest reasoning tasks where V4 still trails.
  • You’re already paying for Max/Team access with predictable quotas.
  • Model behavior consistency matters more than cost (Sol/Fable have more mature safety tuning).

Run both in parallel if:

  • You have $50K+ monthly AI spend — the cost savings pay for the routing logic within a week.

The Bigger Picture: Pricing Innovation as Competitive Weapon

DeepSeek is running the same playbook that TSMC ran against Intel and BYD ran against legacy automakers: compete on cost structure innovation, not just cost per unit.

Peak-valley pricing:

  1. Aligns DeepSeek’s revenue with GPU utilization — off-peak discounts encourage batch workloads that fill GPU capacity that would otherwise be idle.
  2. Trains customers to think about scheduling — the discount is large enough (2x) that customers will re-architect workflows around it, deepening lock-in.
  3. Signals engineering maturity — running frontier models at this pricing profitably requires very sophisticated infrastructure. It’s a subtle competitive flex.

Expect Kimi (Moonshot) and other Chinese labs to copy this within 60 days. Expect Anthropic and OpenAI to not copy it — their infrastructure economics don’t support it (they run at capacity already).

Bottom Line

DeepSeek V4’s peak-valley pricing is the biggest structural change to LLM API economics since the launch of prompt caching. Combined with V4’s already-dramatic pricing advantage over Sol and Fable 5, it makes DeepSeek the default choice for cost-conscious production inference in Q3 2026.

For most engineering teams evaluating this week: migrate a non-critical batch workload to V4-Flash off-peak (Tue/Wed overnight, non-peak windows). Compare quality and total cost. If it holds, expand. The 10-50x cost reduction is real and durable — this isn’t a promotional rate.

Sources