DEEPSEEK-V4

Winning the Price War

chapter 6 of 6 · about 6 minutes

In April 2024, DeepSeek-V2's $0.14/M price helped start the first large AI price war. In April 2026, DeepSeek-V4 repeats the challenge with an architecture built to lower long-context work. An API invoice has more than one number. Input is the text the service must read, output is the text it generates, and a cache hit is input whose earlier computation the service can reuse. A cache miss is new input that must be processed into a fresh key-value cache. Each rate below is priced per million tokens, where a token is the small unit of text a model reads and produces. The distinction matters because a long request has two kinds of work. Prefill reads the prompt and writes its memory state. Decode produces the response token by token from that state. Reusing a prefix can lower prefill work, but it does not remove the cost of new output. The earlier chapters explain why compressed attention, FP4 experts, 13B active parameters, and reusable on-disk caches can change those three lines at once.

How to read the price table

deepseek-v4-flash price history (official pages + press):
  Apr 2026 launch:    $0.14 in / $0.28 out          ($0.028 cache-hit)
  May 2026:           V4-Pro 75% discount → made permanent
  Jul 2026 (0731):    peak/valley pricing, off-peak = half price
  Aug 2026 official:  $0.22 in / $0.66 out peak · $0.11 / $0.33 off-peak
                      cache-hit $0.014 peak · $0.007 off-peak
  OpenRouter lists:   $0.089 / $0.177 (0423 build)

these snapshots combine a price cut and a price increase:
the off-peak tier is cheaper than launch; the peak tier funds it.

Why prices differ by time

The July change adds a time-of-day grid. Peak hours, when the service expects higher demand, pay the full rate. Off-peak hours pay half. This is a capacity-management tool borrowed from electricity markets. A batch agent that can wait until 3am, such as a long repository analysis, can move to less-contended hours. Work that needs a person's immediate response can pay to run during the busy period. In the table's snapshot, the peak input figure is $0.22, higher than the $0.14 launch figure, while the off-peak input figure is $0.11. The point of separating the rows is to keep the cluster busy across more of the week rather than treating every request as equally urgent. The official schedule defines peak as 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. That is 35 of 168 weekly hours, leaving 79% of the week at off-peak rates. The mechanism is ordinary capacity planning: a lower price invites delayable work into time that would otherwise leave accelerators underused. We also need to read the table as a set of snapshots rather than a single universal rate. The launch row, the later first-party row, and the OpenRouter listing refer to different dates, routes, and in some cases different releases. A provider can price the same open weights differently because it owns a different serving fleet, manages a different cache, or chooses a different margin. That is why cache-heavy and output-heavy workloads should not be compared with one input number alone. A repeated system prompt can benefit from a cache-hit rate. A request that emits a long answer still pays the output rate for every newly decoded token. The price question is therefore a workload question: how much of the prompt repeats, how much is new, how much output the task needs, and when the work can run.

Estimate an agent workload

Drag the token and cache sliders. Compare each estimated bill.

input tokens, M1
cached-input fraction0.8
output tokens, M0.2
DeepSeek-V4-Flash
$0.067
GLM-5.3-Flash (list)
$0.15 ·2×
Kimi K2.5
$0.60 ·9×
Gemini 3.7 Flash
$0.96 ·14×
Claude Opus 5
$6.40 ·95×
GPT-5.5
$7.40 ·110×

Prices are from OpenRouter on 2026-08-31. V4-Flash pricing awaits its final listing, as noted in the research file. Estimate = cached input × cache price + fresh input × input price + output × output price.

Change the cache-hit share and the output volume in the invoice builder, then compare which line dominates the bill. Start with a cold prompt, then raise the cache-hit share for a stable system prompt or a repeated document. Next, raise the output volume. The exercise makes a practical limit visible: a low cache-read price cannot cancel an output-heavy workload, and a low output price cannot erase the prefill cost of a continually changing prompt. The better choice depends on the shape of the agent's work rather than a single headline rate. DeepSeek-V4-Flash is not priced only to win benchmark tables. The Pro model targets that role in its Max mode, while Flash is positioned as the default economical choice for agent builders. The 2,500 concurrent-request limit per API key, versus Pro's 500, is a concrete capacity signal. Concurrency is the number of requests one key may run in parallel. It is not the same as tokens per second, but it determines how many simultaneous agent tasks an application can submit before it must queue work. The cheaper model's cap is 5× higher, which suits workloads with many independent calls. GLM-5.3-Flash and Qwen's Flash-Next compete for the same slot: many parallel tasks that need a capable model without reserving the most expensive capacity. The user-facing outcome is straightforward. Competing efficiency designs give builders more low-cost choices, but the right comparison still starts with a real request mix.
receipts, every claim in this chapter, checked 2026-08-31

Chapter sequence complete

Read another model, or revisit a chapter's receipts to compare its claims with the source material.