AIREITER
DeepseekText Chat

DeepSeek V4.1 Flash API — Below DeepSeek List, No Peak Surcharge

DeepSeek V4.1 Flash API below DeepSeek list, no peak-hour markup. 1M context, native image input, top agentic-coding scores. Try it or call the API.

InputOfficial $0.15 per 1M tokensAIReiter $0.1127 per 1M tokensOutputOfficial $0.60 per 1M tokensAIReiter $0.4507 per 1M tokensCache readOfficial $0.003 per 1M tokensAIReiter $0.0022 per 1M tokens
Run with API

INPUT

OUTPUT

Example
Generated in
9.3 seconds
Input tokens
184
Output tokens
1471
Tokens per second
158.17 tokens / second
Time to first token
-

Model details

Use the same model key in Playground, API requests, and internal workflows.

Model ID
deepseek-v4-1-flash
Provider
Deepseek
Protocol
OpenAI Chat Completions · Anthropic Messages
Context window
1,000,000 tokens
Max output
384,000 tokens
Input tokens
11.27 credits / 1M tokens
Output tokens
45.07 credits / 1M tokens
Cache read
0.225 credits / 1M tokens
Cache write
-

What changed in DeepSeek V4.1 Flash

V4.1 Flash is a new model family, not a re-tune of V4 Flash. DeepSeek rebuilt the architecture and trained it from scratch on 45 trillion multimodal tokens, and the Flash tier now outscores V4 Pro on most agent benchmarks.

Causal encoder-decoder MoE

A 552B mixture of experts split into a 20-layer encoder and a 20-layer decoder. Only 8B parameters are active while reading your prompt and 16B while writing the answer, which is why a model this large stays in the Flash price band.

Agentic coding above V4 Pro

DeepSeek reports Terminal-Bench 2.1 at 90.6 versus 82.7 for V4 Flash and 87.9 for V4 Pro, DeepSWE at 74.2 versus 54.4 and 62.7, and Codeforces at 3471. On Terminal-Bench 3.0 the jump is from 7.6 to 30.0.

Vision trained in, not bolted on

A new DeepSeek-ViT encoder was trained jointly with the text model from the first pre-training step. V4 Flash was text-only and its vision variant was an experiment. Screenshots, charts, and photos go in the same request as the prompt.

Quarter-size KV cache for 1M context

Compressed sparse attention and FP4 KV storage cut the global cache to about 890 bytes per token, roughly a quarter of V4 Flash. That is what lets a 1M-token window run at Flash cost and at 214 tokens per second in independent testing.

DeepSeek V4.1 Flash vs V4 Flash

At DeepSeek the migration already happened for you: the official id is now deepseek-flash, and requests that still say deepseek-v4-flash are served by V4.1. On AIReiter the two remain separate models so you can pin either one.

Fewer active parameters, higher scores

V4 Flash activates 13B parameters on every token. V4.1 Flash activates 8B on prefill and 16B on decode and still beats it on every agent benchmark DeepSeek published. Do not read the smaller prefill number as a downgrade.

Thinking is always on now

V4 Flash offered discrete thinking modes. V4.1 Flash uses a continuous reasoning effort and defaults to high. On this route a request that sends thinking disabled still returns reasoning; budget max_tokens accordingly.

Image input is part of the base model

If you were routing screenshots to a separate vision endpoint alongside V4 Flash, V4.1 Flash handles both in one call. Send images as base64 data URIs in the standard content array; this route does not fetch remote image URLs.

Output list price doubled, cache did not

Official output moved from $0.28 to $0.60 per million tokens. Cache-hit input stays near $0.003. Workloads with a long stable prefix and short answers cost about the same as before; chatty generation costs more.

Keep V4 Flash when

A pinned evaluation suite, a client that cannot handle reasoning_content in tool loops, or a strict per-token output budget is not ready to move. Otherwise start new work on V4.1 Flash.

DeepSeek V4.1 Flash API pricing

DeepSeek lists $0.15 input, $0.60 output, and $0.003 cache-hit input per million tokens during off-peak hours, and doubles all three on weekdays 01:00 to 04:00 and 06:00 to 10:00 UTC. AIReiter bills one flat rate around the clock: about 25% below the off-peak list and 62% below the peak list.
01

One flat rate, all day

The three token lines each bill at roughly three quarters of the DeepSeek off-peak rate. There is no peak window on this route, so a job that runs during Beijing office hours pays the same as one that runs at night. The live figures sit above the playground.

02

Where the saving actually lands

Against the official peak rate the AIReiter price is about 38%. Against off-peak it is about 75%. If your traffic is mostly European or US daytime, which is DeepSeek peak, the gap is larger than the headline suggests.

03

Reasoning tokens bill as output

Thinking cannot be turned off on this route and V4.1 Flash is verbose at high effort. Independent testing recorded roughly twice the output tokens of comparable models on the same task set. Set max_tokens for reasoning plus answer.

04

Cache hits are the cheapest line

A cache-hit input token costs about 2% of a cache-miss token. For agent loops that resend the same system prompt and tool schema on every turn, cache reads dominate the bill and are where V4.1 Flash is cheapest.

When to use DeepSeek V4.1 Flash — and when not to

DeepSeek now positions Flash ahead of V4 Pro on quality, cost, and speed. The remaining reasons to pick something else are about knowledge recall and client compatibility, not raw capability.
01

Use it for coding and terminal agents

Multi-file edits, test-fix loops, CI log triage, and computer-use tasks are where the published gains over V4 Flash and V4 Pro are largest. Long tool traces fit the 1M window without chunking.

02

Use it for screenshots and charts

OCR from a UI capture, reading a chart, or checking a rendered page can go straight into the same request as the code question. No second model, no second bill.

03

Use V4 Pro only for compatibility

DeepSeek kept V4 Pro serving after V4.1 Flash shipped because customers asked, not because it scores higher. Stay on Pro if a contract or a frozen eval requires it.

04

Do not use it as an encyclopedia

The base model trails V4 Pro on SimpleQA-Verified by about 13 points. For fact lookup without retrieval, ground the model with your own documents or pick a model tuned for recall.

05

Compare with GLM-5.3 Flash on price

GLM-5.3 Flash is the other multimodal flash model on AIReiter and lists cheaper on output. Pick V4.1 Flash when the task is an agent loop or repository work; pick GLM-5.3 Flash for high-volume extraction and classification.

Call the DeepSeek V4.1 Flash API

The playground on this page and the API share one model id. The one integration rule that breaks clients is how reasoning is handled inside tool loops.

01

Send a real agent task in the playground

Paste a failing log with the diff and a question. Watch output tokens, which include reasoning, and confirm the answer quality before you wire it in. Image input is available through the API.

02

POST /api/v1/chat/completions

Use model "deepseek-v4-1-flash", an AIReiter API key, and the standard Chat Completions body. Streaming works. The same id is also accepted on /api/v1/messages if your stack speaks Anthropic Messages.

03

Round-trip reasoning_content when tools are present

Whenever the request includes a tools array, every prior assistant turn must carry its reasoning_content back, including turns that made no tool call. Dropping it returns HTTP 400. This is the rule that broke several agent frameworks on V4 and it still applies.

04

Attach images in the content array (API)

Use an image_url part with a base64 data URI. PNG, JPEG, GIF, and WebP are accepted. Remote image URLs are not fetched on this route, so inline the bytes. Keep the text part first when the image is context for a question rather than the subject of it.

DeepSeek V4.1 Flash API questions

Model id, pricing, the V4 Flash migration, image input, and the reasoning rule that returns errors.

/ 01

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

On AIReiter use deepseek-v4-1-flash. The internal key is chat-deepseek-v4-1-flash. At DeepSeek directly the official id is deepseek-flash, and the older deepseek-v4-flash id is now served by V4.1 Flash as well.

/ 02

How is DeepSeek V4.1 Flash priced?

DeepSeek lists $0.15 input, $0.60 output, and $0.003 cache-hit input per million tokens off-peak, with all three doubling during weekday peak hours. AIReiter bills one flat rate at all hours, about 25% below the off-peak list and 62% below the peak list. Current route prices are shown above the playground.

/ 03

Is V4.1 Flash better than V4 Pro?

On the agent and coding benchmarks DeepSeek published, yes: Terminal-Bench 2.1 is 90.6 versus 87.9 and DeepSWE is 74.2 versus 62.7. V4 Pro still leads on GPQA Diamond and on knowledge recall. DeepSeek itself now routes new users to Flash.

/ 04

What is different from V4 Flash?

A new encoder-decoder architecture with 8B to 16B active parameters instead of 13B, native image input, always-on reasoning, a quarter-size KV cache for the 1M window, and large gains on every agent benchmark. Official output list price also rose from $0.28 to $0.60.

/ 05

Can I turn thinking off?

Not on this route. Requests that send thinking disabled still return reasoning_content, and reasoning_effort is not forwarded. Control cost with max_tokens and a tight prompt instead.

/ 06

Does it accept images?

Yes, through the API. Vision is native in V4.1 Flash and was verified on this route. Send PNG, JPEG, GIF, or WebP as a base64 data URI in an image_url part; remote URLs are not fetched. The playground on this page is text-only for now.

/ 07

Why does my tool-calling loop get HTTP 400?

You dropped reasoning_content from an earlier assistant message. When a tools array is present, DeepSeek requires the reasoning field on every prior assistant turn, even turns without a tool call. Store it and send it back unchanged.

/ 08

What context and output limits apply?

DeepSeek documents a 1M-token context window and 384K max output. The playground defaults to 8192 output tokens; raise it for long agent runs, and remember reasoning tokens count toward it.

/ 09

Which endpoint should I call?

POST https://aireiter.com/api/v1/chat/completions is the primary contract for this model. POST https://aireiter.com/api/v1/messages also works with the same id for Anthropic-style clients.

/ 10

Can I try it before integrating?

Yes. The playground on this page runs the same model id and route as the API, so token counts and behavior you see here are what the API returns.