DeepSeek’s New API Pricing: What It Actually Means for Heavy Users

Estimated reading time at 200 wpm: 7 minutes

DeepSeek has announced new API pricing for V4 Pro and V4 Flash, effective 16:00 UTC on 16 August 2026. The headline is straightforward: rates have gone up. The detail is more interesting, and it matters more than the headline for anyone whose usage looks like sustained, document-heavy analysis rather than short, occasional queries.

What Changed

Previously, DeepSeek charged a single flat rate per token, regardless of time of day. That structure is gone. The new pricing introduces two changes at once: higher rates across the board, and a peak/off-peak split that didn’t exist before. But caution increased charges does not mean ‘you’ will automatically pay x 6 to x 12 more. Why? It’s about how tokens are consumed. Read on.

V4 Pro, per million tokens:

V4 Flash, per million tokens:

Even the cheapest new rate — off-peak cache hit on V4 Pro — is six times the old flat rate. At peak, cache-hit pricing is twelve times higher. Output pricing, the most expensive category on any token-based model, has risen between 2.3x and 4.7x depending on model and time of day.

Peak hours are defined as 01:00–04:00 and 06:00–10:00 UTC. Everything outside those windows is off-peak. For users in the UK, currently on British Summer Time (UTC+1), that translates to peak periods of 02:00–05:00 and 07:00–11:00 local time. Off-peak covers the rest of the day: midday, afternoon, evening, and the small hours after 11pm.

Why the Split Matters More Than the Headline

A flat percentage increase is easy to reason about. A peak/off-peak split is not, because its impact depends entirely on when and how a given user actually works. Two people with identical monthly token counts could end up with very different bills, depending on what time of day they’re active and how their prompts are structured.

This is where the real story sits, and it’s largely invisible in the headline pricing table.

The Two Variables That Actually Determine the Bill

1. Input-to-output ratio.

Every AI provider charges more for output tokens (what the model generates) than input tokens (what gets sent to it) — typically five to six times more. A user who pastes in large documents and asks for short, targeted answers has a very different cost profile from one having long, generative back-and-forth conversations. The former is input-heavy; the latter is output-heavy. Under the new DeepSeek pricing, this gap has widened rather than narrowed: output rates rose faster, proportionally, than input rates at every tier. Users whose work is dominated by reading and analysing large amounts of text, rather than generating it, are structurally better positioned under this pricing than users doing the reverse.

2. Cache-hit rate

DeepSeek’s API automatically caches the beginning of a prompt if it matches something processed in an earlier request. In practice, this means a long-running conversation — the same document, the same accumulated context, revisited turn after turn — gets progressively cheaper per turn, because most of what gets resent each time is unchanged and qualifies for the cache-hit rate rather than the full cache-miss rate. The difference between the two is not small: at off-peak rates, a cache hit costs roughly 3% of a cache miss. At peak, still around 3%.

This means session behaviour matters enormously. Short gaps between messages in an active thread preserve the cache and keep costs low. Long idle periods let the cache expire, forcing the next message to reprocess the full history at the expensive miss rate. Switching between different AI models mid-conversation has the same effect: a new model has no access to another model’s cache, so the first message on it after a switch is a guaranteed full-price miss on everything that came before.

What This Means in Practice

For a hypothetical heavy user — tens of millions of tokens a month, document-heavy work, most activity landing outside the new peak windows, and long-running sessions that keep the cache warm — the practical increase, while real, is considerably smaller than the headline rate hikes suggest. Applying the new rates to a workload of roughly 84 million tokens a month, split close to 98.5% input and 1.5% output with roughly half of input landing as cache hits, produces a bill in the region of 1.8 times the old cost — a meaningful rise, but nowhere near the six-to-twelve-times jump implied by the raw rate card.

By contrast, a user with a more balanced input/output split, working mostly during peak hours, or starting fresh conversations frequently rather than sustaining long threads, will feel the increase far more sharply — closer to, or exceeding, the headline multiples.

The Broader Lesson

Published API rate cards describe the ceiling and the floor of what a provider will charge, not what any individual user will actually pay. The gap between those two numbers is determined by usage pattern: how much is generated versus consumed, how sessions are structured, and when the work happens. A price increase that looks alarming in isolation can be substantially cushioned, or substantially amplified, by factors that have nothing to do with the sticker price itself.

For anyone evaluating a provider’s pricing change, the useful question isn’t “how much did the rate go up?” It’s “what does my own usage shape do to that number?” The two can produce very different answers.

Flowchart

The following flowchart or algo, is how I might think about it. It’s not perfect. It is not designed to serve ‘everybody’. Not everybody has multi-layered AI in OpenWebUI on Docker.

There is a wonderful thing called common sense that should be applied. Unfortunately it is not widely available and I’m not a supplier.

Glossary

API — Application Programming Interface. A set of rules letting one piece of software send requests to another and get results back, without needing to know how the other one works internally.

Token — The unit AI models process text in, roughly three-quarters of a word. Pricing, limits and speed are all measured in tokens rather than words or characters.

Cache hit — Text in your request the AI has already processed before, in this same conversation. Reused rather than reprocessed, and billed at a much lower rate.

Cache miss — Text the AI is processing for the first time in this exact form. No shortcut available, so it’s billed at the full, higher rate.

Input tokens — The text you send to the AI: your questions, documents, and prior conversation. Usually priced lower than the text the AI sends back to you.

Output tokens — The text the AI generates in reply. Almost always priced several times higher per token than the text you sent it.

Peak/off-peak pricing — Rates that shift by time of day. Using the service during busier peak hours costs more than the same usage during quieter off-peak hours.

UTC — Coordinated Universal Time. A fixed global reference time, unaffected by clock changes, from which local times like British Summer Time are calculated with an offset.