DeepSeek V4 Flash And Pro Pricing: What The Numbers Actually Mean
DeepSeek's current API lineup is two models: V4 Flash (`deepseek-v4-flash`, version DeepSeek-V4-Flash-0731) and V4 Pro (`deepseek-v4-pro`, version DeepSeek-V4-Pro-0813). Both take a one-million-token context, both cap output at 384K tokens, and both default to thinking mode. The commercial story is the price list, and the footnote under it.
Official list prices, per one million tokens, as published on DeepSeek's pricing page on 13 August 2026:
| Model | Cache-hit input | Cache-miss input | Output | Concurrency |
| --- | --- | --- | --- | --- |
| V4 Flash | $0.0028 | $0.14 | $0.28 | 2,500 |
| V4 Pro | $0.003625 | $0.435 | $0.87 | 500 |
The footnote is not small print. DeepSeek says it plans to raise overall API pricing soon, "with a significant increase expected." Treat today's rates as a window, not a permanent floor. If a workflow only works because Flash is fourteen cents, it does not work yet.
Flash Vs Pro Is A Routing Decision, Not A Brand Preference
Flash is the volume engine: 284 billion parameters with about 13 billion active. Pro is the flagship: about 1.6 trillion parameters with about 49 billion active. Both speak OpenAI and Anthropic API shapes, both do JSON, tool calls, and the Responses API. FIM completion is non-thinking only.
Flash is the default for classification, extraction, tagging, routing, drafting, and most agent loops where the next step is cheap and frequent. The 2,500-request concurrency cap is the other reason it is the production workhorse: Pro is capped at 500.
Pro earns the extra money on the hard step: a plan, a review, a migration, a judgment call where one wrong answer is expensive. It is not the model you put on every webhook.
The pattern that works is the same one we use on every other family: cheap tier by default, escalate on low confidence, human review on the decisions that matter. Choosing a Claude model per step is the same decision with different names.
The Number Most Teams Will Miss
Cache-hit input is two to three tenths of a cent per million tokens. That is a 98% discount versus a cache miss on Flash, and a 99% discount on Pro.
If your system prompt, tools schema, and retrieved context are stable, most of the input bill disappears. If every request ships a freshly assembled 20K-token blob, you pay the miss price every time and the "DeepSeek is cheap" story shrinks by an order of magnitude.
This is why we treat prompt caching as a design constraint, not an optimization ticket. Controlling LLM costs in production ranks model right-sizing first and caching second; Flash just makes the second lever unusually large.
A Worked Example Against This Week's Other Models
Same support-triage shape we use elsewhere: 6,000 input tokens and 800 output tokens per email, cache miss, 3,000 emails a month. Official list prices as of 13 August 2026. Read the ratios, then verify on your own traces.
| Model | Cost per email | Monthly at 3,000 |
| --- | --- | --- |
| DeepSeek V4 Flash | ~$0.00106 | ~$3.20 |
| DeepSeek V4 Pro | ~$0.00331 | ~$9.90 |
| Grok 4.6 (prompt under 200K) | ~$0.0168 | ~$50 |
| GPT-5.6 Luna | ~$0.0108 | ~$32 |
| GPT-5.6 Sol | ~$0.054 | ~$162 |
| Claude Opus 5 | ~$0.050 | ~$150 |
Flash is roughly 16 times cheaper than Grok 4.6 on this mix, and about 50 times cheaper than Sol or Opus 5. That is not a reason to move everything to Flash. It is a reason to stop sending classification and extraction to a frontier model.
Grok 4.6 is the other model that shipped this week. SpaceXAI is pricing it at $2 input and $6 output per million tokens under 200K prompt tokens, doubling both rates once the prompt hits 200K, for the entire request. The product wrapped around it is Grok Bot at x.ai/bot: agents with their own computer, not another metered chat box. Different job, different bill.
What This Does Not Change
A cheap token is still a token you should not have spent. Deterministic transforms, date parsing, lookups, and schema mapping belong in ordinary code. Repeated answers belong in a cache or a rules table. A high-volume classifier with stable labels belongs on a small model, or not on a model at all.
And a price this low is a vendor decision, not a law of physics. DeepSeek has already told you the next move is up. Build the thin layer that lets you swap the model ID without rewriting the product. That layer is a week of work. Re-pricing the whole workflow because one vendor changed a cell in a table is a quarter.
If you want the routing and the evaluation set done properly the first time, a paid DX audit is the scoped way to do it.
---
Sources: DeepSeek Models & Pricing · Grok 4.6 announcement · xAI model docs