DeepSeek V4.1 Flash Is Here — and You Can Use It From Claude Code Today

Sep 10, 2026 · #deepseek#claude#tutorial

If you follow Chinese AI at all, you know DeepSeek's pricing announcements have been a wild ride this quarter: peak/off-peak billing in August, a big Flash price cut in September, and now — literally today — V4.1 Flash is rolling out, with DeepSeek routing all deepseek-v4-pro traffic to the new model at the new (lower) rate.

Here's the short version of today's news, and then the part that matters for developers outside China: how to actually use it from your existing tools.

What changed today

V4.1 Flash officially launches (2026-09-10, 12:00 Beijing time).
Internally tested to beat V4 Pro on performance, speed, cost, and total response time.
Until V4.1 Pro ships, requests to deepseek-v4-pro are auto-routed to V4.1 Flash and billed at V4.1 Flash rates.
New pricing (per 1M tokens, off-peak / peak): input ¥1 / ¥2, output ¥4 / ¥8 — roughly 30% cheaper than V4 Flash, and far below V4 Pro's old ¥4.5 / ¥13.5.

That's a serious price-to-performance jump. The models are frontier-level, the context window is 1M tokens, and the API speaks both OpenAI and Anthropic dialects natively.

The one annoying thing

DeepSeek's own platform is built for the domestic market. If you're outside China, signing up means dealing with a Chinese phone number for verification, Chinese payment rails, and a docs experience that assumes you live there. A lot of us just want to point our existing clients at the model and go.

That's exactly the gap I built NovaAPI for: a gateway that gives you the same DeepSeek models (V4.1 line included) through a normal international flow — sign up with any email, pay with PayPal or USDT, no Chinese phone number anywhere. It also aggregates Claude-compatible endpoints, so Claude Code talks to it natively. (Full disclosure: this is my project — but the tutorial below works for any OpenAI/Anthropic-compatible gateway.)

5-minute setup: DeepSeek V4.1 from Claude Code

Step 1 — Create an account and a key

Sign up at api.lbase.com/register (any email works).
Top up with PayPal or USDT — the minimum is small, just enough to try it.
Go to API Keys and create a key. Pick the DeepSeek channel when asked.

You'll get something like sk-....

Step 2 — Point Claude Code at it

Claude Code respects the ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN environment variables:

export ANTHROPIC_BASE_URL="https://api.lbase.com"
export ANTHROPIC_AUTH_TOKEN="sk-your-novaapi-key"
claude

That's it. Claude Code will use the Anthropic-compatible endpoint, and NovaAPI maps Claude model names (claude-sonnet-4-6, claude-fable-5, etc.) onto the DeepSeek V4.1 line automatically.

Step 3 — Or use any OpenAI-compatible client (Cursor, Cherry Studio, code)

from openai import OpenAI

client = OpenAI(
    api_key="sk-your-novaapi-key",
    base_url="https://api.lbase.com/v1",
)

resp = client.chat.completions.create(
    model="deepseek-v4-flash",   # serving V4.1 Flash as of today
    messages=[{"role": "user", "content": "Write a quick Python snippet for a retry-with-backoff helper."}],
)
print(resp.choices[0].message.content)

Cursor: Settings → Models → OpenAI-compatible → Base URL https://api.lbase.com/v1.

Step 4 — Verify

curl https://api.lbase.com/v1/models \
  -H "Authorization: Bearer sk-your-novaapi-key"

You'll see the model list. Run a couple of prompts and check your usage page — billing is transparent, pay-as-you-go, in USD.

Why not just call DeepSeek directly?

If you already have a Chinese phone number and a way to pay in CNY, absolutely go direct — it'll be marginally cheaper. NovaAPI is for everyone else, and for people who want:

One key for multiple Chinese models (DeepSeek today, more families coming),
Claude Code / Anthropic-SDK compatibility without shims,
PayPal/USDT billing instead of Chinese payment apps,
No phone verification, no VPN, no WeChat.

A note on pricing and peak hours

DeepSeek bills peak/off-peak (peak = Mon–Fri 01:00–04:00 & 06:00–10:00 UTC). NovaAPI passes the official off-peak price × a transparent multiplier to you, and like DeepSeek, peak-hour requests cost more. The price table on novaapi.lbase.com is always current — check it before you commit a big batch job.

Wrapping up

V4.1 Flash looks like the best price-to-performance point in the DeepSeek lineup right now, and with the Pro auto-routing in place, "deepseek-v4-flash" is secretly a very fast, very cheap flagship. If you've been wanting to try Chinese frontier models but bounced off the phone-number wall — the door is open.

Questions, corrections, or war stories from wiring this into your stack? Drop a comment. I read all of them.