DeepSeek V4.1 Flash 1017: API Pricing, Open Weights and New Search Tools

DeepSeek released V4.1 Flash 1017 on October 17, updating the model behind its website, app and Flash API. The same announcement opens its in-house web search engine to developers through standalone Search and Fetch APIs. Existing API users can keep calling deepseek-flash; DeepSeek says the calling method and token prices are unchanged, according to the official changelog.

The update gives developers two things to evaluate together: a revised model for completing tasks and a retrieval service for bringing current information into those tasks. DeepSeek says the model produces shorter outputs and finishes work faster, which could reduce the cost of an agent run even when its price per token stays fixed.

Those claims need testing on the actual workload, particularly if an application depends on lengthy tool sequences or repeated attempts. A useful evaluation would track completed tasks, elapsed time and total billed usage on the same set of jobs. Keep prompts and tool permissions consistent, and include jobs that previously needed human intervention.

What changed in Flash 1017?

The Hugging Face model card calls 1017 a minor update and publishes the weights under the MIT license. DeepSeek reports Terminal-Bench 4.0 rising from 31.2 to 42.3, DeepSWE v1.1 from 74.2 to 77.0, and ProgramBench from 20.3 to 29.1. Its HLE result without tools improves from 36.8 to 42.6.

These are company-reported results. Terminal-Bench uses the Claude Code harness, while DeepSWE uses Mini-SWE-V2, so the software surrounding the model is part of the comparison. The scores give teams a reason to run their own evaluation, without establishing how reliably it will handle a particular repository.

API context, modes and existing aliases

The official pricing page lists a one-million-token context window and a maximum output of 384,000 tokens. Thinking mode is the default, with non-thinking mode also supported. Flash accepts visual input and supports tool calls and the Responses API.

Legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names remain accepted, but now route to Flash 1017. That means an unchanged configuration can receive a changed model, making a regression test sensible for existing integrations.

DeepSeek Flash 1017 API pricing

Prices are in US dollars per million tokens, from the same pricing page:

Token type	Off-peak	Peak
Uncached input	$0.15	$0.30
Cached input	$0.003	$0.006
Output	$0.60	$1.20

Peak windows are 01:00–04:00 and 06:00–10:00 UTC on weekdays excluding Chinese public holidays. Everything else is off-peak, including weekends. Search and Fetch each cost $0.75 per 1,000 successful requests. Server-side Search and Fetch calls through DeepSeek Harness, Claude Code and Codex are currently free; the pricing page schedules billing for November 1 at 00:00 Beijing time.

How Search and Fetch work

The Search and Fetch guide describes two JSON endpoints using the existing DeepSeek API key. Search returns structured web results, while Fetch retrieves an individual page as Markdown. Search can return between one and 50 results, with optional page content. Each service has an independent limit of five requests per second per account. Fetch also has a per-domain rate limit. Failed calls, empty search results and timeouts are not billed, and requesting more results does not increase the per-request search charge.

For a research assistant, that creates a fairly direct integration path: find relevant pages, retrieve the useful ones and pass the evidence into the model. Developers still need to check whether search results are current and relevant for their own domains. The announced limits also deserve attention when several agents make requests simultaneously, since separate API keys do not create separate account quotas.

What self-hosting still requires

DeepSeek recommends Harness 0.2.1-rc.1 or later for the updated model. Self-hosting has an additional implementation detail. The release has no Jinja-format chat template. DeepSeek provides reference encoding code and deepseek-recipe, Rust libraries with Python bindings for converting requests, encoding prompts and parsing responses. The toolkit supports Messages, Chat Completions and Responses formats, but leaves inference, tool execution and HTTP transport to the application. Teams running the weights should verify prompt formatting alongside model quality before switching production traffic.

I'm Paras. I work at the intersection of AI, engineering and distribution.

If you're building an AI product, I help turn technical launches and product thinking into clear content developers and founders can use and share.

Book a 30-minute call.