DeepSeek V4.1 Flash is an open-weight Mixture-of-Experts model from DeepSeek with a 1,048,576-token context window. You can call it through AssemblyAI LLM Gateway as deepseek-v4.1-flash for $0.30 per million input tokens and $1.20 per million output tokens, using the OpenAI SDK and the same API key you use for speech-to-text.
On hour-long meeting transcripts it posts the top summarization score of any model on the Gateway in Voice Standard Bench, for under a cent per hour of audio. It's also slow on long inputs, it runs only on opt-in providers, and it trails other models on transcript correction. This guide covers all of it: the specs, both model IDs, pricing next to DeepSeek's own API, the Data Controls step most quickstarts skip, working code in Python, JavaScript and curl, and where the benchmark says to use something else.
DeepSeek V4.1 Flash at a glance
DeepSeek released V4.1 Flash on September 10, 2026, and describes it as the smallest model in its new architecture family, with native image understanding. Per the Hugging Face model card, it has a 552B-parameter backbone that activates 8B parameters per token during prefill and 16B during decode, and the weights are MIT-licensed. An update, V4.1-Flash-1017, followed on October 17, 2026.
Here's how it's listed in the LLM Gateway model catalog as of November 17, 2026:
| Spec | DeepSeek V4.1 Flash on LLM Gateway |
|---|---|
| Model IDs | deepseek-v4.1-flash and deepseek-v4.1-flash-1017 |
| Maker | DeepSeek (open weights, MIT license) |
| Context window | 1,048,576 tokens |
| Max output | 393,216 tokens |
| Input price | $0.30 per 1M tokens |
| Output price | $1.20 per 1M tokens |
| Cached input (read) | $0.006 per 1M tokens |
| Regional surcharge | None listed |
| Region | US endpoint (llm-gateway.assemblyai.com); not on the EU endpoint |
| Serving providers | DigitalOcean (default) and Fireworks for deepseek-v4.1-flash; Fireworks only for deepseek-v4.1-flash-1017 |
| Opt-in required | Yes. DigitalOcean and Fireworks are both opt-in providers |
| Supported parameters | max_tokens, temperature, tools, tool_choice, response_format, reasoning_effort, stream |
The Gateway accepts text input only, so DeepSeek's image understanding isn't available through it. If you need to send images to the model, call DeepSeek's API directly.
Two model IDs: deepseek-v4.1-flash and deepseek-v4.1-flash-1017
The catalog lists two versions at the same price. They behave differently on voice data, so pick deliberately.
| deepseek-v4.1-flash | deepseek-v4.1-flash-1017 | |
|---|---|---|
| DeepSeek release | DeepSeek-V4.1-Flash, Sep 10, 2026 | DeepSeek-V4.1-Flash-1017, Oct 17, 2026 |
| Providers on LLM Gateway | DigitalOcean (default), Fireworks | Fireworks |
| Price per 1M tokens (in / out) | $0.30 / $1.20 | $0.30 / $1.20 |
| Median latency, 2-minute transcripts | 19.7 s | 6.6 s |
| Summarization, 1-hour transcripts | 96 | 88 |
| Transcript correction, 1-hour transcripts | 78 | 52 |
| Pick it for | Batch summaries of long recordings | Short calls and extraction where response time matters |
Scores are from Voice Standard Bench, as of the October 29, 2026 run. The 1017 update answers two to three times faster on short transcripts and holds up better on extraction at 15 minutes (85 against 81). The original release writes better long summaries and is much stronger at transcript correction on long inputs. If you need a model served by a provider with a signed BAA, deepseek-v4.1-flash-1017 runs only on Fireworks; more on that in the data section below.
DeepSeek V4.1 Flash API pricing
LLM Gateway charges $0.30 per million input tokens, $1.20 per million output tokens and $0.006 per million cached input tokens for both model IDs, with no regional surcharge listed. Here's how that lines up with DeepSeek's own API, using DeepSeek's pricing page on November 17, 2026:
| Where you call it | Model name | Input per 1M | Cached input per 1M | Output per 1M | Notes |
|---|---|---|---|---|---|
| AssemblyAI LLM Gateway | deepseek-v4.1-flash or deepseek-v4.1-flash-1017 |
$0.30 | $0.006 | $1.20 | One rate at all hours; same key and invoice as Speech-to-Text |
| DeepSeek API, peak hours | deepseek-flash |
$0.30 | $0.006 | $1.20 | Peak: 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday |
| DeepSeek API, off-peak hours | deepseek-flash |
$0.15 | $0.003 | $0.60 | Off-peak rates are 50% of peak |
At DeepSeek's peak rates the per-token prices are the same. Off-peak, DeepSeek's own API charges half. DeepSeek's pricing page maps deepseek-flash to V4.1-Flash-1017 and lists no separate name for the September release.
What the Gateway adds is everything around the model: one key and one invoice for transcription and LLM calls, transcript injection by ID, automatic fallbacks to another model, and zero data retention for prompts and completions. For workloads built on recordings, the more useful number is cost per hour of audio. Voice Standard Bench estimates that summarizing one hour of audio with deepseek-v4.1-flash costs about $0.0092 at Gateway prices.
Step 1: Get an API key and turn on the opt-in providers
- Sign up for a free AssemblyAI account and copy your API key from the dashboard.
- Open the Data Controls page and turn on DigitalOcean and Fireworks.
That second step is the one that trips people up. In the words of the Providers docs: "Providers marked Opt-in are disabled by default. To use them, you need to manually turn on their toggles on the Data Controls page in your dashboard." DigitalOcean is the default provider for deepseek-v4.1-flash and Fireworks is the second one listed; deepseek-v4.1-flash-1017 runs on Fireworks only. Turn the provider on before you send your first request.
One AssemblyAI key covers LLM Gateway and Speech-to-Text. Sign up, turn on the DigitalOcean and Fireworks providers in Data Controls, and send your first request in a few minutes.
Step 2: Call the DeepSeek V4.1 Flash API
LLM Gateway is OpenAI-compatible, so you keep the OpenAI SDK and change two things: the base URL and the model ID. Set your key first:
export ASSEMBLYAI_API_KEY="your_api_key_here"Python
pip install openaiimport os
from openai import OpenAI
client = OpenAI(
base_url="https://llm-gateway.assemblyai.com/v1",
api_key=os.environ["ASSEMBLYAI_API_KEY"],
)
response = client.chat.completions.create(
model="deepseek-v4.1-flash",
messages=[
{"role": "system", "content": "You summarize customer calls in three sentences."},
{
"role": "user",
"content": "Agent: Thanks for calling. Caller: My invoice shows two charges "
"for October. Agent: I see the duplicate and I've refunded it; it posts in 3-5 days.",
},
],
max_tokens=500,
)
print(response.choices[0].message.content)JavaScript
npm install openaiimport OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://llm-gateway.assemblyai.com/v1",
apiKey: process.env.ASSEMBLYAI_API_KEY,
});
const response = await client.chat.completions.create({
model: "deepseek-v4.1-flash",
messages: [
{ role: "system", content: "You summarize customer calls in three sentences." },
{
role: "user",
content:
"Agent: Thanks for calling. Caller: My invoice shows two charges for October. " +
"Agent: I see the duplicate and I've refunded it; it posts in 3-5 days.",
},
],
max_tokens: 500,
});
console.log(response.choices[0].message.content);Save it as quickstart.mjs so Node treats the top-level await as an ES module.
curl
curl https://llm-gateway.assemblyai.com/v1/chat/completions \
-H "Authorization: $ASSEMBLYAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4.1-flash",
"messages": [
{"role": "system", "content": "You summarize customer calls in three sentences."},
{"role": "user", "content": "Agent: Thanks for calling. Caller: My invoice shows two charges for October. Agent: I see the duplicate and I have refunded it; it posts in 3-5 days."}
],
"max_tokens": 500
}'To try the 1017 update, change the model string to deepseek-v4.1-flash-1017. Tool calling, structured outputs through response_format, reasoning_effort and streaming are all on the supported-parameters list, so the usual OpenAI request shapes work.
Step 3: Run DeepSeek V4.1 Flash on a recording
The case where this model stands out is long audio, so here's the full loop: transcribe a meeting with AssemblyAI-STT-3.7, then summarize it with DeepSeek V4.1 Flash. Pass transcript_id and put the literal tag {{ transcript }} in your message, and the Gateway fetches the transcript text for you. The same request sets a fallback model, so a failed DeepSeek call is retried on claude-opus-5-5, the next-highest 1-hour summarization score in the benchmark.
pip install assemblyai openaiimport os
import sys
import assemblyai as aai
from openai import OpenAI
API_KEY = os.environ["ASSEMBLYAI_API_KEY"]
SUMMARY_PROMPT = (
"Summarize this meeting for someone who missed it. List the decisions, the action "
"items with their owners, and any open questions. Use only what the transcript says.\n\n"
"Transcript:\n{{ transcript }}"
)
# 1. Transcribe with AssemblyAI-STT-3.7 and speaker labels
aai.settings.api_key = API_KEY
config = aai.TranscriptionConfig(speech_models=["assemblyai-stt-3-7"], speaker_labels=True)
transcript = aai.Transcriber().transcribe(sys.argv[1], config=config)
if transcript.status == aai.TranscriptStatus.error:
raise RuntimeError(f"Transcription failed: {transcript.error}")
# 2. Summarize with DeepSeek V4.1 Flash; the Gateway injects the transcript by ID
client = OpenAI(
base_url="https://llm-gateway.assemblyai.com/v1",
api_key=API_KEY,
timeout=600, # hour-long inputs can take a minute or more on this model
)
response = client.chat.completions.create(
model="deepseek-v4.1-flash",
messages=[{"role": "user", "content": SUMMARY_PROMPT}],
max_tokens=2000,
extra_body={
"transcript_id": transcript.id,
"fallbacks": [{"model": "claude-opus-5-5"}],
},
)
print(response.choices[0].message.content)Run it with python summarize_recording.py meeting.mp3. transcript_id and fallbacks are Gateway fields the OpenAI SDK doesn't know about, which is why they go in extra_body. The tag has to be exactly {{ transcript }}, with the spaces. For the transcription side, see the Speech-to-Text API. For a longer walkthrough of chunking, judging and batch jobs, see how to summarize meetings with LLMs in Python.
The companion repo has these scripts plus bench_lookup.py, which prints the model's live benchmark results from the public models endpoint, no API key needed.
How DeepSeek V4.1 Flash scores on voice tasks
Voice Standard Bench is AssemblyAI's open-source benchmark for LLMs on transcript work. It runs four tasks on real call transcripts at four lengths, and two LLM judges grade summaries, extraction and correction against the source transcript, while dictation is scored against an answer key. The methodology post explains the scoring in detail. Here are the results for deepseek-v4.1-flash from the October 29, 2026 run (scores out of 100):
| Transcript length | Overall | Summarization | Extraction | Correction | Dictation | Cost per audio hour | Latency p50 / p95 |
|---|---|---|---|---|---|---|---|
| 30 seconds | 95 | 95 | 90 | 96 | 97 | $0.0169 | 6.2 s / 36.0 s |
| 2 minutes | 92 | 97 | 88 | 85 | 97 | $0.0194 | 19.7 s / 94.9 s |
| 15 minutes | 83 | 93 | 81 | 75 | Not measured | $0.0192 | 92.7 s / 261.2 s |
| 1 hour | 83 | 96 | 76 | 78 | Not measured | $0.0117 | 73.3 s / 234.9 s |
Cost per audio hour is the bench's estimate of running its tasks on one hour of audio at global pricing. Latency is end-to-end request time. Three things stand out.
It leads on long summaries. At 1 hour, its summarization score of 96 is the highest of the 64 model and reasoning-effort combinations on the bench. claude-opus-5-5 and gpt-6-astra follow at 95, at $0.1631 and $0.1961 per audio hour for that task against $0.0092 for DeepSeek V4.1 Flash. The Choose a model docs list it as both the top score and the lowest-cost model near the top for summarizing 1-hour meetings.
It's slow on long inputs, and it errored often in the 1-hour runs. Median latency at 1 hour is 73.3 seconds, with a p95 of 234.9 seconds. In that bucket the bench logged 35 completed requests and 25 errors, and at 15 minutes it logged 50 completed and 10 errors. Errored requests are left out of the score, so 96 describes the calls that finished. The 1017 update logged 2 errors at 1 hour, at a median of 38.9 seconds. Plan for retries and a fallback before you put the original release on hour-long jobs.
It trails on everything but summaries at meeting length. Its overall score at 1 hour is 83, and 11 model and effort combinations score higher. gpt-6-astra leads at 88. On extraction from 1-hour transcripts it scores 76 against gpt-6.1-sol at 85, and on correction it scores 78 against glm-5.3 at 88. On short audio it's competitive across the board: 95 overall at 30 seconds, one point behind the top score of 96.
These numbers come from the bench's fixed prompts on call audio. Treat them as a shortlist, then test on your own transcripts. The full results and harness are in the GitHub repo.
Upload a recording in the Playground, transcribe it, and send the same summary prompt to DeepSeek V4.1 Flash and the models it competes with. No code needed.
When to use DeepSeek V4.1 Flash, and when to pick something else
Use DeepSeek V4.1 Flash for batch summaries of long recordings, such as meeting notes, recorded calls and lectures, where nobody is waiting on the response and the 1M-token window means you can skip chunking. Use the 1017 update for short calls and extraction where response time matters.
Pick another model when the job is live, needs EU processing, or is transcript correction. Here's how the alternatives compare on 1-hour summarization:
| Model | 1-hour summarization score | Summarization cost per audio hour | Latency p50 (1 hour) | Provider status | EU endpoint |
|---|---|---|---|---|---|
deepseek-v4.1-flash |
96 | $0.0092 | 73.3 s | Opt-in (DigitalOcean, Fireworks) | No |
claude-opus-5-5 |
95 | $0.1631 | 31.5 s | On by default (AWS Bedrock) | Yes |
gpt-6-luna (high) |
92 | $0.0027 | 37.6 s | On by default (AWS Bedrock, OpenAI) | No |
minimax-m3 |
79 | $0.0084 | 14.3 s | Opt-in (Fireworks) | No |
claude-opus-5-5 is the pick if you need the EU endpoint or providers that are on by default. gpt-6-luna scores 4 points lower for less per audio hour and responds faster. minimax-m3 has the same token prices as DeepSeek V4.1 Flash but scores 17 points lower on this task. For live voice work, smaller models usually win on short transcripts, and our post on small vs. large LLMs covers the tradeoff. If you're picking a summarizer by transcript length, start with the best LLM for summarizing transcripts. Building a conversational agent on DeepSeek instead? The DeepSeek voice agent tutorial wires it to streaming speech-to-text and text-to-speech.
Regions, data retention and the BAA
Region. Both model IDs are listed for the US endpoint, https://llm-gateway.assemblyai.com/v1. The EU endpoint serves Claude and Gemini models only, so for EU data residency use a model such as claude-opus-5-5.
Retention. LLM Gateway offers zero data retention for inputs and outputs: it doesn't store the content of prompts or completions and keeps only usage metadata such as the model, token counts, status and request ID. AssemblyAI has opted out of model training with all LLM Gateway providers.
Providers. The two providers behind DeepSeek V4.1 Flash differ. The Providers page lists Fireworks with zero data retention and a signed BAA, and lists DigitalOcean's retention as "see provider", with no BAA. If you process protected health information, AssemblyAI is considered a business associate under HIPAA and offers a Business Associate Addendum (BAA); check which provider serves each model you plan to use and confirm coverage with our team before sending PHI.
Rate limits and fallbacks
LLM Gateway rate-limits requests per minute and tokens per minute, tracked per account and per model, provider and region over a rolling 60-second window. New accounts start at 30 requests per minute at Tier 0 and move up tiers automatically as spend grows. A request over either limit returns 429.
Given the error rate in the long-transcript runs, set a fallback on any production call to deepseek-v4.1-flash. The fallbacks array takes up to two backup models when you set fallback_config.depth to 2. Our guide to adding LLM fallbacks to a voice pipeline shows the patterns, including how to pick the backup model.
Final words
DeepSeek V4.1 Flash gives you a top-scoring long-meeting summarizer at token prices that keep the summary of an hour-long transcript under a cent. In exchange you accept slow responses on long inputs, two opt-in providers and US-only processing. Benchmarks for open models move with every update, as the gap between the September release and the 1017 update shows, so check the live numbers in the models endpoint before you commit, and pin the model ID you've tested.
Transcribe with AssemblyAI-STT-3.7 and summarize with DeepSeek V4.1 Flash on one API key, with a fallback model set in the same request. Start free.
Frequently Asked Questions (FAQs)
How do I get DeepSeek V4.1 Flash API access?
You can call DeepSeek V4.1 Flash through DeepSeek's own API as deepseek-flash, or through AssemblyAI LLM Gateway as deepseek-v4.1-flash with the OpenAI SDK and base URL https://llm-gateway.assemblyai.com/v1. On the Gateway, turn on the DigitalOcean and Fireworks providers in Data Controls first, because both are opt-in and off by default. deepseek-v4.1-flash defaults to DigitalOcean, and deepseek-v4.1-flash-1017 runs on Fireworks only.
What does the DeepSeek V4.1 Flash API cost?
On LLM Gateway it costs $0.30 per million input tokens, $1.20 per million output tokens and $0.006 per million cached input tokens, at all hours. DeepSeek's own API charges the same at peak hours and half that off-peak. Voice Standard Bench estimates about $0.0092 to summarize one hour of audio with it.
What is the DeepSeek V4.1 Flash context window?
DeepSeek V4.1 Flash has a 1,048,576-token context window on LLM Gateway, with up to 393,216 output tokens. That's enough to send a full hour-long meeting transcript in one request without chunking.
Is DeepSeek V4.1 Flash good for summarizing meetings?
Yes, for long recordings where response time isn't critical. As of the October 29, 2026 Voice Standard Bench run, it has the top summarization score on 1-hour transcripts (96), but its median response time at that length is 73.3 seconds and the bench logged 25 errored requests in that bucket, so set a fallback model.
Which providers offer DeepSeek V4.1 Flash with zero data retention?
On LLM Gateway, Fireworks is listed with zero data retention and a signed BAA, while DigitalOcean's retention is listed as "see provider". The deepseek-v4.1-flash-1017 model runs on Fireworks only. LLM Gateway itself doesn't store prompt or completion content.
Is DeepSeek V4.1 Flash available in the EU?
No. On LLM Gateway, DeepSeek V4.1 Flash is available through the US endpoint only. The EU endpoint serves Claude and Gemini models, so use one of those, such as claude-opus-5-5, if your data must stay in the EU.