Comparison11 min readPublished: 2026-09-13Updated: 2026-09-13

DeepSeek V4.1 Flash API Guide: Pricing, Benchmarks & Migration

DeepSeek V4.1 Flash adds native vision, a 1M-token context window, lower API rates, and a new cache-efficient architecture. See what changes for V4 Flash and V4 Pro users before the September 14 routing switch.

By Heizi· Founder & Editor· Published: 2026-09-13Official sources checked · 2026-09-13

Verification: Release details, model IDs, architecture, pricing, migration dates, and benchmark settings were checked against DeepSeek's announcement, API documentation, technical report, and model card. Benchmark results are vendor-reported; Get Model Key has not independently benchmarked the model.

Create a DeepSeek API key

Use the canonical model ID deepseek-flash for new integrations. Verify live pricing before moving production traffic.

What DeepSeek released on September 10

DeepSeek released V4.1 Flash on September 10, 2026. It is a new multimodal Mixture-of-Experts model rather than another V4 Flash post-training update. The model accepts text and images, supports contexts up to one million tokens, and is available through the DeepSeek API under the canonical ID deepseek-flash.

The release changes existing integrations too. The retired deepseek-v4-flash and deepseek-v4-flash-vision-exp names temporarily route to V4.1 Flash. At 04:00 UTC on September 14, 2026 (12:00 Beijing time), deepseek-v4-pro will also route to V4.1 Flash and use Flash pricing until V4.1 Pro launches. DeepSeek has not announced a V4.1 Pro release date.

Release date: September 10, 2026
Canonical API model ID: deepseek-flash
Input: text and images; output: text
Context window: up to 1M tokens; maximum output: 384K tokens
Open weights: MIT license on Hugging Face

Why the new architecture matters

V4.1 Flash has a 552B-parameter backbone, but its Causal Encoder-Decoder architecture activates only 8B parameters per token during input prefill and 16B during output decoding. DeepSeek says the design reduces compute and memory pressure for input-heavy agent workflows without shrinking the available knowledge capacity to the same degree.

The practical number is the KV cache footprint: 890 bytes per token, about one quarter of V4 Flash. Persistent sliding-window states are reduced further through bounded replay. A smaller cache does not make every answer better, but it can lower the serving cost of long repository sessions, repeated tool calls, and document-heavy agents.

Architecture figures reported by DeepSeek
Specification	DeepSeek V4.1 Flash
Backbone parameters	552B MoE
Active parameters	8B prefill / 16B decode
Global KV cache	890 bytes per token
Training	45T multimodal tokens; context extended to 1M
Reasoning control	Continuous effort setting from 1 to 100

Benchmarks: strong agent results, not a clean sweep

DeepSeek reports 74.2 on DeepSWE v1.1, 90.6 on Terminal-Bench 2.1, 54.8 on AutomationBench, and 31.8 on Agent's Last Exam. In the same vendor-run table, V4.1 Flash scores above V4 Pro on those tests. That supports testing it for coding agents and long tool-driven workflows.

The same report shows clear limits. V4 Pro remains ahead on GPQA Diamond and the comparable text-only HLE subset, while Opus 5 and GPT-5.6 Sol lead V4.1 Flash on several harder terminal, security, and programming tests. Scores also change with the agent scaffold: DeepSeek evaluated code tasks with different harnesses and maximum reasoning effort. Treat these numbers as a shortlist for evaluation, not a production verdict.

Selected vendor-reported results at maximum reasoning effort
Benchmark	V4.1 Flash	V4 Pro	V4 Flash
DeepSWE v1.1	74.2	62.7	54.4
Terminal-Bench 2.1	90.6	87.9	82.7
AutomationBench	54.8	43.2	37.7
Agent's Last Exam	31.8	25.7	25.2
GPQA Diamond	90.9	92.4	89.9

DeepSeek V4.1 Flash API pricing

DeepSeek charges separate rates for cached input, uncached input, and output. Off-peak rates are half the weekday peak rates. The prices below are for the official direct API and took effect at 04:00 UTC on September 10, 2026.

Peak windows are Monday through Friday, 01:00–04:00 and 06:00–10:00 UTC. In Beijing time, that is 09:00–12:00 and 14:00–18:00. Other hours use the off-peak rate. A gateway or cloud reseller may use different prices.

Official direct API price per 1M tokens
Billable usage	Off-peak	Peak
Cached input	$0.003 / ¥0.02	$0.006 / ¥0.04
Uncached input	$0.15 / ¥1	$0.30 / ¥2
Output	$0.60 / ¥4	$1.20 / ¥8

Migration checklist before the V4 Pro switch

If your application already uses deepseek-v4-flash, it is probably receiving V4.1 Flash through the compatibility route. Move new configuration to deepseek-flash so the intended model is explicit. DeepSeek has not published an end date for the old aliases, so compatibility should not be treated as permanent.

V4 Pro users have a more important regression risk: the model behind the same deepseek-v4-pro string changes on September 14. Save accepted Pro outputs now, then replay representative prompts against deepseek-flash. Compare task success, tool-call arguments, latency, total output tokens, and image behavior. Keep the previous results as a baseline even if your code needs no endpoint change.

Use deepseek-flash in new deployments
Record V4 Pro baselines before September 14 at 04:00 UTC
Test thinking effort and output length with real prompts
Validate tool schemas and structured output
Update cost alerts to the new cached, uncached, and output rates
Check the live API response and documentation again on cutover day

Call DeepSeek V4.1 Flash with the OpenAI Python SDK

DeepSeek keeps an OpenAI-compatible API surface. Set the base URL to DeepSeek, provide a DeepSeek platform key, and use deepseek-flash as the model. Keep the key in an environment variable rather than placing it in source code.

python

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["DEEPSEEK_API_KEY"],
    base_url="https://api.deepseek.com",
)

response = client.chat.completions.create(
    model="deepseek-flash",
    messages=[
        {"role": "user", "content": "Review this API migration plan and list the two highest risks."}
    ],
)

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

Minimal text request using the canonical V4.1 Flash model ID.

Who should test V4.1 Flash first

V4.1 Flash is a sensible first test for long-context coding agents, repository analysis, document extraction, image-assisted debugging, and high-volume workflows where cached prompts repeat across many steps. Its API price and cache footprint are the clearest advantages.

Do not migrate purely from the headline that Flash beats Pro. Keep a stronger-model fallback for difficult research, security work, and tasks where one failed attempt costs more than the token savings. The right comparison is cost per accepted task, including retries and human review.

Open weights do not mean easy self-hosting

DeepSeek published V4.1 Flash weights under the MIT license and provides reference paths for vLLM and SGLang. That is useful for infrastructure teams that need deployment control, but the full model is still enormous. DeepSeek's own announcement asks large-scale deployers to discuss configurations around 2,000 GPUs plus a storage cluster.

Most teams should start with the hosted API, collect real token and latency data, and consider quantized or self-hosted deployment only after the workload is stable. Open weights improve choice; they do not remove hardware, storage, and operations costs.

FAQ

What is the DeepSeek V4.1 Flash API model ID?

Use deepseek-flash for new integrations. The retired deepseek-v4-flash and deepseek-v4-flash-vision-exp names temporarily route to V4.1 Flash for compatibility.

How much does the DeepSeek V4.1 Flash API cost?

Off-peak direct API rates per 1M tokens are $0.003 cached input, $0.15 uncached input, and $0.60 output. Weekday peak rates are twice those amounts. The Chinese pricing page lists the corresponding off-peak rates as ¥0.02, ¥1, and ¥4.

Does DeepSeek V4.1 Flash support images?

Yes. V4.1 Flash natively accepts text and image input and generates text. It replaces the separate V4 Flash Vision experimental route.

What happens to deepseek-v4-pro on September 14?

At 04:00 UTC on September 14, 2026, DeepSeek says requests using deepseek-v4-pro will route to V4.1 Flash and use Flash pricing until V4.1 Pro launches. The model string may keep working while the model behavior changes.

Is DeepSeek V4.1 Flash better than V4 Pro?

It scores higher on several DeepSeek-run coding and agent benchmarks, but V4 Pro remains ahead on some reasoning tests. Run a regression set from your own application before treating it as a universal upgrade.