> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ruapi.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# DeepSeek vs Claude: what one day of work actually costs
> Take a real one-day Claude Code bill — 405M tokens, $167 — and reprice the exact same tokens on DeepSeek V4 Pro and Flash. Line by line, and where the 25–30× gap comes from.
The easiest way to compare models isn't the price list — it's a real invoice. Here's the usage export from one day of Claude Code: an ordinary day with an agent reading the repo, editing files and running tests. It added up to 405 million tokens and **\$167.24**:
| Billing line | Tokens for the day | Price | Cost |
| ------------------ | ------------------ | -------------- | ------------ |
| Input (cache hit) | 395,374,336 | \$0.30 / MTok | \$118.61 |
| Input (cache miss) | 8,509,057 | \$3.00 / MTok | \$25.53 |
| Output | 1,540,240 | \$15.00 / MTok | \$23.10 |
| **Total** | 405,423,633 | — | **\$167.24** |
One line is worth pausing on. Almost the entire volume — 395 million tokens out of 405 — is cache reads, not fresh input. Claude caches the repeated context: the system prompt, the contents of open files, the conversation history. Reading it costs \$0.30 per million instead of \$3.00 — ten times less. Without caching, the same day would have run about \$1,235. So \$167 is already a heavily discounted number, not the sticker price.
The question from here is simple: what do those exact same tokens cost on DeepSeek?
## The same day on DeepSeek V4 Pro
Same three lines, priced with `deepseek-v4-pro` — the top of the line, the closest match to Sonnet by class of work:
| Billing line | Tokens for the day | Price | Cost |
| ------------------ | ------------------ | -------------- | ---------- |
| Input (cache hit) | 395,374,336 | \$0.004 / MTok | \$1.58 |
| Input (cache miss) | 8,509,057 | \$0.435 / MTok | \$3.70 |
| Output | 1,540,240 | \$0.87 / MTok | \$1.34 |
| **Total** | 405,423,633 | — | **\$6.62** |
\$167.24 versus \$6.62 — for the same work, that's **25× cheaper**.
## The same day on DeepSeek V4 Flash
`deepseek-v4-flash` is the cheapest model in the line, and it's fast; it's what you reach for on a steady stream of simple tasks:
| Billing line | Tokens for the day | Price | Cost |
| ------------------ | ------------------ | ------------- | ---------- |
| Input (cache hit) | 395,374,336 | \$0.01 / MTok | \$3.95 |
| Input (cache miss) | 8,509,057 | \$0.14 / MTok | \$1.19 |
| Output | 1,540,240 | \$0.28 / MTok | \$0.43 |
| **Total** | 405,423,633 | — | **\$5.57** |
Here the gap is even wider — roughly **30×**.
## The result in one table
| Model | Cost of the day | Difference |
| --------------------- | --------------- | ------------- |
| Claude Sonnet | \$167.24 | — |
| **DeepSeek V4 Pro** | **\$6.62** | \~25× cheaper |
| **DeepSeek V4 Flash** | **\$5.57** | \~30× cheaper |
## Where the gap comes from
It comes from two things, and both favor DeepSeek.
**Price per token.** Fresh input on Sonnet is \$3.00 per million; on Pro it's \$0.435 — nearly seven times less. Output is \$15.00 against \$0.87, a gap of more than seventeen times. Output is where the saving shows most: the longer the model's answers, the wider the invoices spread apart.
**Price of a cache read.** Cache carries the load here — 97.5% of the tokens. And even on cache reads DeepSeek is cheaper: \$0.004 per million on Pro against \$0.30 on Sonnet. For this day, cache reads cost \$118.61 on Claude and \$1.58 on Pro.
Stack one on top of the other, and on an identical set of tokens the gap reaches 25–30×.
Prices are current as of testing and the vendors can change them. Current per-token rates are on the [Pricing](/en/pricing) page and at [www.ruapi.ai](https://www.ruapi.ai). What matters here is the order of the difference between the models, not the exact dollar figures.
## Which one to use
Cheaper isn't the same as "always better." It depends on the task.
* **`deepseek-v4-pro`** — the everyday workhorse: code, refactoring, reasoning, agents. Quality stays close to the top models while costing a fraction. If Claude Code or a similar agent is running for you right now, this is the first thing to swap in.
* **`deepseek-v4-flash`** — for volume: classification, tagging, short replies, rough drafts. Where price and speed matter more than top quality.
* **Claude** — when you need to read images (DeepSeek has no vision) or top quality on the hardest tasks. One RuAPI token covers everything, so the models happily coexist in a single project.
DeepSeek models are **text-only**: they don't read images, diagrams or screenshots. If you need vision, look at [Claude](/en/claude-api), [Gemini](/en/gemini-api) or GLM-5V. More on the line itself is on the [DeepSeek API](/en/deepseek-api) page.
## How to switch
If your project already talks to RuAPI over the OpenAI-compatible protocol, there's one line to change — the model ID. The `base_url` and key stay the same:
```python theme={null}
from openai import OpenAI
client = OpenAI(
api_key="YOUR_RUAPI_KEY",
base_url="https://www.ruapi.ai/v1",
)
resp = client.chat.completions.create(
model="deepseek-v4-pro", # instead of claude-sonnet-4-6
messages=[{"role": "user", "content": "Refactor this function and explain what you changed"}],
)
print(resp.choices[0].message.content)
```
Setup from scratch — the key and a first request in Python and curl — is in the [Quickstart](/en/quickstart).
## Next
Pro and Flash: how they differ and what they do.
Lines, versions, and when Claude earns its price.
Pay per token used, with a log of every request.
base\_url, key and your first request.