DeepSeek
DeepSeek V4.1 Flash
A new Flash generation for coding and long-context agent work. DeepSeek's asymmetric architecture uses 8B active parameters to read input and 16B to generate output.
Replaces DeepSeek V4 Flash 0731: input $0.0798 → $0.1575 per 1M (+97%), output $0.1607 → $0.63 per 1M (+292%), same context window.
Run this model
deepseek/deepseek-v4.1-flash
Use an endpoint supported by this model.
Get a key for this setup →const response = await fetch('https://api.minirouter.sh/v1/chat/completions', { method: 'POST', headers: { Authorization: `Bearer ${process.env.MINIROUTER_KEY}`, 'Content-Type': 'application/json', }, body: JSON.stringify({ model: 'deepseek/deepseek-v4.1-flash', messages: [{ role: 'user', content: 'Why is the sky blue?' }], }),}) const data = await response.json()console.log(data.choices[0].message.content)Use deepseek/deepseek-v4.1-flash for text requests. Image input is not supported on MiniRouter for this model. Start with the Chat Completions example; publisher client integrations do not establish MiniRouter client compatibility.
Model overview
- Input / 1M
- $0.1575
- All-in MiniRouter rate
- Output / 1M
- $0.63
- All-in MiniRouter rate
- Context
- 1M
- 1,000,000 tokens
- Max output
- 384K
- tokens
- Architecture
- 552B MoE
- 8B input / 16B output active
- DeepSWE v1.1
- 74.2%
- DeepSeek-reported, max effort
- Announced
- 10 Sept 2026
- publisher release
Publisher evidence
A step forward for coding agents.
DeepSeek reports 74.2% on DeepSWE v1.1 and 90.6% on Terminal Bench 2.1. The comparisons below come from one publisher table, with separate axes for each benchmark.
These evaluations use maximum effort (100), temperature 1.0 and top_p 0.95. Terminal Bench uses DeepSeek Harness Minimal; DeepSWE uses mini-SWE, with a 1M-token context. They are not MiniRouter performance measurements.
V4.1 vs V4
DeepSWE v1.1
Resolved repository tasks rise from 54.4% for V4 Flash and 62.7% for V4 Pro to 74.2%.
DeepSeek-reported results
0 to 100%
- V4.1 Flash74.2%
- V4 Pro62.7%
- V4 Flash54.4%
View chart data
| Model | Reported result |
|---|---|
| V4.1 Flash | 74.2% |
| V4 Pro | 62.7% |
| V4 Flash | 54.4% |
V4.1 vs V4
Terminal Bench 2.1
Terminal task Pass@1 rises from 82.7% for V4 Flash and 87.9% for V4 Pro to 90.6%.
DeepSeek-reported results
0 to 100%
- V4.1 Flash90.6%
- V4 Pro87.9%
- V4 Flash82.7%
View chart data
| Model | Reported result |
|---|---|
| V4.1 Flash | 90.6% |
| V4 Pro | 87.9% |
| V4 Flash | 82.7% |
Choosing the model
Evaluate it on the work you repeat.
The strongest published improvements are in agentic coding. Task quality, tool reliability, and total tokens matter alongside the headline score.
Compare coding benchmarks- 01
Repository work and terminal tasks
Try it on bounded fixes, code review, and tool-driven development. The published DeepSWE and Terminal Bench results make these useful starting points for your own evaluation.
- 02
Long documents and repeated context
Use the live context and output limits shown above. Repeated context may benefit from upstream caching; actual cache use and peak hours affect the final bill.
- 03
Check reasoning and visual requirements
The publisher reports GPQA Diamond at 90.9%, below V4 Pro's 92.4% in its table. The model has native vision, but MiniRouter currently supports text input for this route.
Live route pricing
Plan around peak hours and output tokens.
The price panel uses MiniRouter's catalog rates, including the 5% markup. DeepSeek's provider route charges 2× from 01:00–04:00 and 06:00–10:00 UTC, Monday–Friday; all other hours are off-peak.
Publisher pricing and schedule- 01
Separate list prices from your route price
DeepSeek lists $0.15/M uncached input and $0.60/M output off-peak, before MiniRouter's markup. Peak list prices are $0.30/M and $1.20/M respectively.
- 02
Compare the whole task
Reasoning and longer answers can increase output usage. Compare total task cost and success rate with V4 Flash or another coding model using the same tools and prompt.
From V4 and the beta
Use the permanent model ID.
The September 10 release is distinct from the temporary V4.1 Flash beta. Select the exact model ID in your API request and rerun your representative tasks.
Read the release and API changes- 01
Replace the beta ID
Change deepseek/deepseek-v4.1-flash-beta to deepseek/deepseek-v4.1-flash once the permanent model appears in MiniRouter's catalog. Keep your MiniRouter base URL and key.
- 02
Check aliases at the provider boundary
DeepSeek's direct API uses deepseek-flash and redirects older V4 aliases. Those publisher redirects do not guarantee that a MiniRouter V4 ID selects this release.
- 03
Keep evaluation results separate
The September 10 Artificial Analysis API refresh has no matching V4.1 Flash evaluation. Use the publisher results above as evidence to investigate; independent speed and quality measurements are not yet included here.
One route, standard clients.
Authenticate with a MiniRouter bearer key. Caller-provided upstream credentials are not used on this route.
Read model setup docs →https://api.minirouter.sh/v1deepseek/deepseek-v4.1-flash/v1/chat/completions/v1/messages/v1/responsesmax_tokens, temperature, stop, tools, tool_choice, reasoning, include_reasoning
Configured provider
Priority follows the live catalog order.
Input: text · Output: text
| Provider | Priority | Input / 1M | Output / 1M | Parameters | Access |
|---|---|---|---|---|---|
vercel | 1 | $0.1575 | $0.63 | max_tokens, temperature, stop, tools, tool_choice, reasoning, include_reasoning | MiniRouter balance |