DeepSeek V4.1 Flash is the model developers use most right now: over one week it carried one token in four in programming, according to OpenRouter. It holds a million-token context, calls tools and costs $0.12 per million input tokens and $0.48 per million output tokens on Kumo. Below: what each DeepSeek model in the catalogue does, what the benchmarks show, what it costs in practice and how to connect the DeepSeek API in five minutes.
Context
1M tokens
up to 384K tokens in one reply
Price on Kumo
$0.1 / $0.5
input / output per 1M, exactly $0.12 and $0.48
Cache reads
$0.002
per 1M tokens read from cache
Share of programming
27%
of tokens in one week, first place
Which DeepSeek models Kumo has
The catalogue carries three DeepSeek models, and all of them work through one OpenAI-compatible endpoint and one key. They differ less in level than in price and character: one fast and cheap for volume, one senior model for reasoning, one the cheapest in the whole catalogue.
DeepSeek V4.1 Flash (deepseek-v4-1-flash) shipped on September 10, 2026. It is a sparse mixture-of-experts model on a 552B-parameter backbone that activates about 8B parameters for input and 16B for output at each step. A compressed KV cache makes long agent sessions cheaper, and the million-token context fits an average repository whole. It is the family's main model and the subject of this article.
DeepSeek V4 Pro 0813 (deepseek-v4-pro-0813) is the August snapshot of the V4 line's larger model. It is slower and pricier than Flash, but stronger on science questions and in coding by Artificial Analysis's composite indexes. DeepSeek V4 Flash (deepseek-v4-flash) is the April snapshot 0423: 284B parameters with 13B active, replies of up to 900K tokens and the lowest output price of any model on Kumo.
All three support streaming and tool calling — confirmed by the gateway's probe requests. Structured output against a JSON schema is not confirmed for them yet, so check a strict response format on your side.
| Model | Identifier | Input, $ per 1M | Output, $ per 1M | Cache reads, $ per 1M | Context |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash | deepseek-v4-1-flash | $0.1 (0.12) | $0.5 (0.48) | $0.002 | 1M |
| DeepSeek V4 Pro 0813 | deepseek-v4-pro-0813 | $0.5 (0.528) | $1.6 (1.584) | $0.02 (0.017) | 1M |
| DeepSeek V4 Flash | deepseek-v4-flash | $0.07 (0.072) | $0.1 (0.145) | $0.01 (0.014) | 1M |
DeepSeek V4.1 Flash benchmarks
The model's strength is knowledge and long context. On GPQA Diamond, a set of graduate-level science questions, it answers 87.9% correctly, and on AA-LCR, which asks for reasoning over a very long document, it scores 84%. In the tool-using support-agent scenario (τ²-Bench Airline) it reaches 78%.
Its weakness is heavy agentic work in the terminal: on Terminal-Bench v4 it completes 26.8% of tasks. For comparison, Claude Sonnet 5.5 solves 63.6% of the same test and Claude Opus 5.5 59.6%. That is a fair price for a model of this class: V4.1 Flash is a good workhorse, but the hardest multi-step tasks in a repository are better given to a stronger model.
How the DeepSeek line grew
Artificial Analysis's Intelligence Index rolls about ten benchmarks into one number, and it shows how fast DeepSeek is moving: R1 from early 2025 scores 13.1, V3.2 from December 2025 21.5, and V4.1 Flash already 39.5. The junior Flash of the September generation even beats the senior V4 Pro from August (36.0) on the composite.
The index is not everything. V4 Pro 0813 scores higher on science questions (90.3% on GPQA Diamond) and has a published coding index of 68.8, which Artificial Analysis has not yet set for V4.1 Flash. If your work is long reasoning over code, run both models on your own examples.
Why DeepSeek V4.1 Flash is so popular
OpenRouter publishes which models carry the most tokens in each task category. In the week of October 5, DeepSeek V4.1 Flash came first by volume in all seven categories on the chart below: 27.2% of tokens in programming, 28.6% in finance, 24% in academia. No Claude or GPT model reaches the same volume in these categories.
The reason is simple: on volume work — log triage, code rewrites, classification, drafts — the quality gap to the top models is smaller than the price gap. When a mid-level model can do the job, developers take the cheapest one that is good enough. Popularity here is voting with money, not a quality ranking: for the hardest tasks the same developers switch to Claude and GPT.
What DeepSeek costs in practice
A million input tokens is several hundred thousand words — a few books or an average repository whole. For that much input V4.1 Flash charges $0.12, and $0.48 for a million tokens of reply. A typical working day of a coding agent that read 5M tokens and wrote 500K costs about 85 cents if nothing hit the cache.
In practice almost everything hits the cache: an agent rereads the same files and system prompt again and again. A cache read on V4.1 Flash costs $0.002 per million tokens — 60 times less than ordinary input. A hundred million tokens read from cache over a long session cost about 20 cents. That is why long agent sessions on DeepSeek cost noticeably less than the input price suggests.
On Kumo you pay as you go from one balance, topped up with a card, SBP or cryptocurrency. There is no subscription and no weekly quota: top up and work, and set spending limits yourself on every key.
Level against price: what to compare DeepSeek with
The honest comparison is not with Claude Opus but with models of the same class. On the chart below each dot is a model from the Kumo catalogue: intelligence index up, the price of a million output tokens across on a log scale. Higher and further left is better value.
The picture: DeepSeek V4 Flash is the cheapest model on the chart, V4.1 Flash sits in the middle of its class, and it has strong rivals nearby. Claude Haiku 5.5 and GPT-6 Luna cost $0.3 per million output tokens on Kumo with indexes of 43.4 and 38.1. So choose by your own tasks: DeepSeek wins on very long context, cheap cache and wherever open weights matter, while on the pure ratio of index to output price it has competition.
What DeepSeek V4.1 Flash is good for
Use it where there are many tokens and medium difficulty. When a task is bound by volume rather than by peak quality, V4.1 Flash does it for pennies.
- Coding agents on routine work: edits from a description, tests, refactoring, migrations, documentation.
- Long documents and logs — a million-token context means you do not have to cut the input into pieces.
- Bulk processing: classification, data extraction, translation, text drafts.
- Chatbots and assistants where latency and price matter: the model writes about 220 tokens per second, the first token arrives in about 0.7 seconds.
- A pair of models: DeepSeek for volume, Claude Sonnet or GPT-6.1 Sol for the hard steps and for checking the result.
How to connect the DeepSeek API in five minutes
- Sign up in the Kumo console and top up the balance — by card, SBP or cryptocurrency.
- Create an API key and give it a spending limit right away: a separate key per project shows where the money went.
- Set the gateway address https://api.kumorouter.com/v1 in any OpenAI-compatible client or SDK.
- Choose the model deepseek-v4-1-flash (or deepseek-v4-pro-0813, deepseek-v4-flash) and send the first request.
- Check in the console's request log that the charge went to the right model.
curl -s https://api.kumorouter.com/v1/chat/completions \
-H "Authorization: Bearer $KUMO_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4-1-flash",
"messages": [
{"role": "user", "content": "Write a Python function that validates an IBAN."}
],
"max_tokens": 2048
}'
A Python example
The official OpenAI SDK works unchanged — just swap the address and the key. Streaming and tool calling are switched on the same way as for OpenAI's models.
from openai import OpenAI
client = OpenAI(
base_url="https://api.kumorouter.com/v1",
api_key="KUMO_API_KEY",
)
stream = client.chat.completions.create(
model="deepseek-v4-1-flash",
messages=[{"role": "user", "content": "Explain what this SQL query does: ..."}],
stream=True,
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")
DeepSeek in Cursor, Cline and other editors
Any tool with an "OpenAI Compatible" setting connects DeepSeek the same way: the address https://api.kumorouter.com/v1, your Kumo key and the model name deepseek-v4-1-flash. That is how it works in Cline, Roo Code, Continue, OpenCode and in Cursor with your own key. Step-by-step settings are in the guides to Cline and VS Code, OpenCode and Cursor.
DeepSeek models on Kumo speak the Chat Completions protocol. Claude Code speaks Anthropic's protocol, so use Claude models there — the setup is in the Claude Code guide.
Frequently asked questions
- Does the DeepSeek API work from Russia?
- Through Kumo, yes: the gateway is reachable without a VPN, and the balance is topped up with a Russian card, SBP or cryptocurrency. You need neither a DeepSeek account nor a foreign card.
- How much does DeepSeek V4.1 Flash cost via the API?
- On Kumo, $0.12 per million input tokens and $0.48 per million output tokens. A cache read is $0.002 per million. The current rates are always on the models page.
- How is DeepSeek V4.1 Flash different from V4 Pro?
- V4.1 Flash is newer, faster and cheaper, and higher on Artificial Analysis's composite index (39.5 against 36.0). V4 Pro 0813 is stronger on science questions and about three times the price. Take Flash for volume work; for long reasoning, run both on your own examples.
- DeepSeek or Claude for coding?
- For routine code, tests and documentation DeepSeek V4.1 Flash gives good results at a small fraction of the price. For complex agentic work and large refactors the leaders of the rankings — Claude and GPT — are clearly stronger. A common setup is to keep both models on one key and switch by task.
- What context does DeepSeek V4.1 Flash have?
- A million tokens of input and up to 384 thousand tokens in one reply. That is enough to pass an average repository or a large stack of documents whole.
- Is it safe to send code to DeepSeek?
- Requests go through the Kumo gateway to the model's provider. How we handle data is described on the security page. For secrets, keys and personal data the general rule holds: do not send them to any cloud model.
Sources
- Benchmarks and indexes: the Kumo benchmarks page — Artificial Analysis and OpenRouter data, October 16–17, 2026.
- Popularity by category: OpenRouter rankings, week of October 5, 2026.
- Prices and model specs: the Kumo models page, price list v27 of October 17, 2026.
- DeepSeek documentation: api-docs.deepseek.com.