Last updated: October 29, 2026
DeepSeek V4.1 Flash 1017 is an October 17, 2026 update to DeepSeek’s V4.1 Flash model that improves coding, agent and visual benchmarks while producing shorter outputs. It is live in the DeepSeek app, website and API as deepseek-flash, from $0.15 per million input tokens, with MIT-licensed open weights.
DeepSeek calls the release “a Pareto improvement in performance and cost” over the previous version (DeepSeek release note). The same announcement opened its Search and Fetch APIs to all developers. For context on the policy debate around the company, see our explainer on why the US banned DeepSeek from government agencies.
Contents
- What’s new in DeepSeek V4.1 Flash 1017
- Specs and API limits
- How to call DeepSeek V4.1 Flash in the API
- DeepSeek V4.1 Flash pricing
- Using it in Claude Code
- Running the open weights
- How it compares
- Limitations to know
- Frequently asked questions
What’s new in DeepSeek V4.1 Flash 1017
The original V4.1 Flash launched on September 10, 2026 as a 552B-parameter mixture-of-experts model with native image understanding and 8B active parameters for input and 16B for output (V4.1 Flash announcement). DeepSeek said its KV cache needs one quarter of the HBM and one eighth of the SSD storage of the previous generation, which lowers cache-hit costs in long agent sessions.
The 1017 update keeps that design and focuses on results. The table uses DeepSeek’s own figures from the Hugging Face model card; they are vendor results, not independent tests.
| Benchmark | V4.1 Flash 1017 | V4.1 Flash |
|---|---|---|
| Terminal-Bench 4.0 | 42.3 | 31.2 |
| DeepSWE v1.1 | 77.0 | 74.2 |
| ProgramBench | 29.1 | 20.3 |
| HLE without tools | 42.6 | 36.8 |
| Agents’ Last Exam | 35.4 | 31.8 |
| Apex Bench v1.1 | 72.1 | 64.6 |
| Chartography (with tools) | 83.0 | 78.9 |
Beyond scores, DeepSeek says the update uses reinforcement learning with self-feedback to improve end-to-end project work, such as building games and 3D scenes from scratch. It is tuned for the DeepSeek Harness agent tool, and DeepSeek recommends Harness v0.2.1-rc.1 or later.
Specs and API limits
These values come from DeepSeek’s Models and Pricing page:
| Item | DeepSeek V4.1 Flash |
|---|---|
| API model name | deepseek-flash |
| Model version | DeepSeek-V4.1-Flash-1017 |
| Context length | 1M tokens |
| Maximum output | 384K tokens |
| Thinking mode | Thinking (default) and non-thinking |
| Features | JSON output, tool calls, Responses API, Anthropic API, vision, chat prefix completion (beta), FIM completion (beta, non-thinking only) |
| Concurrency limit | 2,500 |
How to call DeepSeek V4.1 Flash in the API
The API accepts both OpenAI and Anthropic request formats, so existing SDKs work with a new base URL. A minimal Chat Completions request:
curl https://api.deepseek.com/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
-d '{"model": "deepseek-flash",
"messages": [{"role": "user", "content": "Hello!"}],
"thinking": {"type": "enabled"},
"reasoning_effort": "high"}'Follow these steps, based on DeepSeek’s thinking mode guide:
- Set the base URL to
https://api.deepseek.com(OpenAI format) orhttps://api.deepseek.com/anthropic(Anthropic format). - Set
modeltodeepseek-flash. Old Flash model names are routed to the same model. - Choose thinking or non-thinking. Thinking is on by default at
higheffort. Send{"thinking": {"type": "disabled"}}to turn it off, or"reasoning": {"effort": "none"}in the Responses API. - Pick an effort level:
low,highormax. Other values are mapped; for example,mediumandxhighbecomehigh. - With the OpenAI Python SDK, pass
thinkinginsideextra_body. - In tool-calling loops, pass back
reasoning_contentfrom every previous assistant turn, or the API returns a 400 error.
In thinking mode, temperature, presence_penalty and frequency_penalty are accepted but ignored. For deepseek-flash, top_p only applies in thinking mode, and values below 0.95 are treated as 0.95.
DeepSeek V4.1 Flash pricing
DeepSeek uses peak and off-peak rates. Per its pricing page, peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours, including weekends, are off-peak at half price.
| Per 1M tokens | Peak | Off-peak |
|---|---|---|
| Input, cache hit | $0.006 | $0.003 |
| Input, cache miss | $0.30 | $0.15 |
| Output | $1.20 | $0.60 |
The Search and Fetch APIs each cost $0.75 per 1,000 successful requests. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free until billing starts at 16:00 UTC on October 31, 2026 (noon Eastern time), according to the release note. To save money, schedule batch jobs in off-peak hours and keep long prompt prefixes stable so they hit the cache.
Using it in Claude Code
DeepSeek publishes a Claude Code integration guide. On Linux or macOS:
export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=<your DeepSeek API key>
export ANTHROPIC_MODEL=deepseek-flash[1m]
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-flash[1m]
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-flash[1m]
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-flash
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-flash
export CLAUDE_CODE_EFFORT_LEVEL=max
export CLAUDE_CODE_AUTO_COMPACT_WINDOW=786432Then run claude in your project folder. Set the Opus variable explicitly: DeepSeek maps any model name starting with claude-opus to deepseek-v4-pro, which is billed at the higher V4 Pro price. Web search in Claude Code works through DeepSeek’s Search API and adds token and search charges. Our Claude Code power-user guide covers the tool itself.
Running the open weights
The weights are on Hugging Face and ModelScope under the MIT License. The model card lists 763B parameters and recommends temperature 1.0, top_p 1.0, a 1M-token context window and max_tokens of at least 256K. There is no Jinja chat template; the repository includes a Python reference encoder and a set of Rust libraries with Python bindings, called deepseek-recipe, for production prompt formatting.
How it compares
| V4.1 Flash | V4 Pro | |
|---|---|---|
| API model | deepseek-flash |
deepseek-v4-pro |
| Version | V4.1-Flash-1017 | V4-Pro-0813 |
| Vision | Yes | Not supported |
| Peak input / output per 1M tokens | $0.30 / $1.20 | $1.32 / $3.96 |
| Concurrency limit | 2,500 | 500 |
On DeepSeek’s model card, the 1017 update scores 42.3 on Terminal-Bench 4.0, ahead of GPT-5.6 Sol (39.9) but behind Claude Opus 5 (51.8). For a newer closed alternative, see our GPT-6.1 Terra guide, which lists its API price at $1 input and $5 output per million tokens, or our Gemini 4 Argon guide.
Limitations to know
- Benchmark figures are DeepSeek’s own and use specific harnesses, such as Claude Code for Terminal-Bench 4.0.
- FIM completion works only in non-thinking mode.
- Peak pricing doubles costs during set UTC hours on weekdays.
- Check your organization’s data and compliance policies before sending sensitive material to the hosted API. For background, read about US accusations of model distillation by Chinese AI firms.
Frequently asked questions
What is DeepSeek V4.1 Flash 1017?
It is a minor update to DeepSeek V4.1 Flash released on October 17, 2026. DeepSeek says it improves reasoning, coding, agentic and visual benchmarks while shortening outputs and task times. It now serves all API requests that use the deepseek-flash model name.
How much does DeepSeek V4.1 Flash cost?
DeepSeek’s pricing page lists $0.30 per million cache-miss input tokens, $0.006 per million cache-hit input tokens and $1.20 per million output tokens at peak times. Off-peak rates are half of those: $0.15, $0.003 and $0.60.
What model name do I use for DeepSeek V4.1 Flash in the API?
Use deepseek-flash. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but DeepSeek says those models are retired and the requests are served by V4.1-Flash-1017 at the Flash price.
What is the DeepSeek V4.1 Flash context window?
The API supports a 1M-token context length with a maximum output of 384K tokens. The Hugging Face model card recommends a 1M-token context window and max_tokens of at least 256K for local inference.
Can I run DeepSeek V4.1 Flash locally?
Yes. The weights are on Hugging Face and ModelScope under the MIT License. The model card lists 763B parameters in safetensors format and provides an inference folder with weight-conversion instructions. It does not include a Jinja chat template, so use the provided encoding code.
Can I use DeepSeek V4.1 Flash with Claude Code?
Yes. Point ANTHROPIC_BASE_URL to https://api.deepseek.com/anthropic, set your DeepSeek API key as ANTHROPIC_AUTH_TOKEN, and set the model variables to deepseek-flash. DeepSeek’s integration guide lists the full set of environment variables.
Sources
- DeepSeek API Docs: DeepSeek-V4.1-Flash-1017 release note
- DeepSeek API Docs: DeepSeek-V4.1-Flash announcement
- DeepSeek API Docs: Models and Pricing
- DeepSeek API Docs: Integrate with Claude Code
- Hugging Face: DeepSeek-V4.1-Flash-1017 model card
Image: DeepSeek logo, by DeepSeek, MIT License, via Wikimedia Commons. Placed on a white background and resized.
1 thought on “DeepSeek V4.1 Flash: Power-User Guide to the 1017 Update, API and Pricing”