Call DeepSeek V4.1 Flash GA with deepseek-flash

Introduction

Search DeepSeek V4.1 Flash and still follow the 8 Sep beta window, and you hit two traps: the temp ID deepseek-v4.1-flash-expires-on-0910 has expired, and treating the press line “Pro will be routed to Flash” as current policy fights the live price table.

How to call it now: set model to deepseek-flash. On 10 Sep 2026 the official Change Log listed DeepSeek-V4.1-Flash as GA; the pricing page row is deepseek-flash (backend version DeepSeek-V4.1-Flash). This site is an independent community — model names, rates, and routing follow the official pages. For the expired eval notes see DeepSeek V4.1 Flash beta — do not put that temp ID in production.

I. What DeepSeek V4.1 Flash GA Is — and Is Not

Claim	Fact
You still need expires-on-0910 to reach V4.1	No. The shipping model name is deepseek-flash
You must change base_url	No. OpenAI-compatible remains https://api.deepseek.com
Old deepseek-v4-flash 404s immediately	No. It temporarily routes to V4.1 Flash at Flash rates; new code should not depend on the alias
Vision-Exp is still the vision SKU	No. deepseek-v4-flash-vision-exp is retired and temporarily routed; GA includes native multimodal
Pro is gone	No. After 14 Sep, deepseek-v4-pro continues with unchanged billing — Is DeepSeek V4 Pro still available?

In one line: daily work and vision go to deepseek-flash; flagship reasoning stays on deepseek-v4-pro; the temp beta string is dead. Key signup: DeepSeek web & API beginner guide; API overview: DeepSeek API.

Public architecture notes (vendor-side, check the model card — not an SLA from this site): 552B MoE, Causal Encoder–Decoder, about 8B active on prefill / 16B on decode, 1M context. Ship against the model name and the rate card, not paper numbers in a contract.

II. How to Call: Change Only the Model Name to deepseek-flash

base_url stays the same. Set model to:

deepseek-flash

Minimal example:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_KEY",
    base_url="https://api.deepseek.com",
)

response = client.chat.completions.create(
    model="deepseek-flash",
    messages=[{"role": "user", "content": "In three sentences, say which model tier you are."}],
)

print(response.choices[0].message.content)

Keys still come from the official platform. Thinking on/off, reasoning_effort, and 400 replay rules: DeepSeek thinking mode — GA supports thinking, but do not copy the beta-era “same Flash price, 20 concurrency” expectations.

Must do:

1. Config, gateway maps, and CI defaults all become deepseek-flash; delete expires-on-0910.
2. Vision requests should not still name Vision-Exp; use the current multimodal content blocks — step-by-step: How to send images with the DeepSeek API.
3. Reconcile bills against the new Flash rates, not the August “price hike” table.
4. Concurrency follows the Flash row on the pricing page (2,500 in the docs at writing time); saturation: DeepSeek API 429.

III. Migrating Old Names, the Beta ID, and Vision-Exp

Name in your config	What it hits at writing time	Do this
deepseek-flash	DeepSeek-V4.1-Flash	New code writes only this
deepseek-v4-flash	Temporarily routed to V4.1 Flash	Change it when you can; not a permanent alias
deepseek-v4-flash-vision-exp	Temporarily routed to V4.1 Flash	Vision now uses GA native multimodal
deepseek-v4.1-flash-expires-on-0910	Expired	Delete it; calls fail or return an unknown error
deepseek-v4-pro	DeepSeek-V4-Pro-0813	Keep it for the flagship tier; do not force-cut because a press note said so

Compatibility routing is a transition, not an SLA. Official copy says “temporarily” — aliases can disappear without a notice in your logs. Regression-test the same image+text and tool cases you used on old Flash / Vision-Exp, not just HTTP 200.

IV. Rates, Peak-Valley Pricing, and Cache

The day DeepSeek V4.1 Flash shipped, official Flash token prices dropped with the new architecture (effective 04:00 UTC on 10 Sep 2026). Pricing page at writing time (USD / million tokens, off-peak / peak):

Tier	Input · cache hit	Input · cache miss	Output	Doc concurrency cap
deepseek-flash	0.003 / 0.006	0.15 / 0.30	0.60 / 1.20	2500
deepseek-v4-pro	0.022 / 0.044	0.66 / 1.32	1.98 / 3.96	500

Peak hours are 01:00–04:00 and 06:00–10:00 UTC Monday–Friday (about 09:00–12:00 and 14:00–18:00 Beijing time), excluding Chinese public holidays; everything else is off-peak at half price. Prefix hits still follow disk cache rules: DeepSeek API context caching. The August “1100%” article described that hike — do not use it to estimate Flash bills after 10 Sep.

V. vs DeepSeek V4 Pro

The launch note said third-party tests put V4.1 Flash ahead of V4 Pro on performance, cost, speed, and total runtime — that is a routing hint, not “the Pro endpoint is closed.” Split traffic like this:

• Lower unit price, higher concurrency, native multimodal vision: deepseek-flash.
• Contracts/evals locked to the flagship, long agent chains: keep deepseek-v4-pro; official copy confirms it continues after 14 Sep with unchanged billing.
• V4.1 Pro: the press note mentioned it as the end of a transition. At writing time there is no model row and no date. Do not put an unreleased flagship in default routes.

Both rows document 1M context and 384K max output; Flash checks Vision, Pro does not. Capability edges come from your own regression set.

FAQ

Does the web app / App also say deepseek-flash?
This article covers DeepSeek API model names. Client labels follow the official product UI; this site’s entry: DeepSeek web.

Can we keep old names and rely on routing for now?
It works in the short term, billed at Flash rates. When aliases go away, unchanged configs fail at the worst moment. New repos and new environments should allow only deepseek-flash or an explicit Pro.

Do we still need a separate Vision model?
GA folds vision into deepseek-flash. Keeping Vision-Exp ties you to a retired alias.

Summary

Shortest DeepSeek V4.1 Flash GA path: set model to deepseek-flash → delete the beta ID and Vision-Exp defaults → check peak/off-peak and cache on the pricing page → leave flagship work on still-serving deepseek-v4-pro. Open the official model table first, then change config — do not keep forwarding expired beta tutorials.

Further reading

• DeepSeek API
• DeepSeek V4.1 Flash beta
• Is DeepSeek V4 Pro still available?
• What is DeepSeek V4 Pro
• DeepSeek API context caching
• DeepSeek thinking mode
• DeepSeek web & API beginner guide