Contents

DeepSeek has released DeepSeek-V4.1-Flash-1017, reporting improvements in reasoning, coding, agent tasks, and visual recognition, along with shorter outputs and faster task completion. The Search and Fetch APIs are also available to all users, with each endpoint priced at $0.75 per 1,000 calls.

Key points#

  • DeepSeek-V4.1-Flash-1017 was released on October 17, 2026, and is available through DeepSeek’s website, app, and API.
  • The model weights are available on Hugging Face and ModelScope under the MIT license.
  • According to DeepSeek, the model improves on its predecessor’s benchmark results while reducing task completion time and output length.
  • Search returns structured JSON results, while Fetch retrieves a single web page and returns its contents as Markdown.
  • Eligible Search and Fetch calls through DeepSeek Harness, Claude Code, and Codex are free until October 31, 2026, at 16:00 UTC.

What the model update changes#

DeepSeek-V4.1-Flash-1017 is an update to DeepSeek-V4.1-Flash, the lightweight multimodal mixture-of-experts model introduced in September 2026. It targets improvements in reasoning, coding, agent tasks, and visual recognition.

In practical terms, DeepSeek says the model can complete tasks more accurately in less time and with shorter responses. Applications that call the model repeatedly could see reduced latency and output token costs. The supplied announcement materials do not quantify these reductions, so they need to be verified against actual workloads.

The hosted model is available through DeepSeek’s website, app, and API. Teams considering self-hosting can obtain the weights released under the MIT license from Hugging Face or ModelScope, but will need to assess hardware requirements and operating costs separately.

What the benchmarks show#

According to DeepSeek, the updated model outperformed GPT-5.6 Sol, Claude Opus 5, and Kimi K3 on several agent benchmarks. The evaluations cited include DeepSWE v1.1 for coding agents, AutomationBench v1.0.6 for workflow automation, and Apex Bench v1.1 for long-running agent tasks.

These are vendor-reported results. The supplied materials do not include specific scores, detailed test conditions, or independent reproductions. They provide a basis for evaluating this release, but do not demonstrate that it outperforms the competing models across all use cases.

For a useful comparison, measure both task success rates and total execution costs. Shorter responses help only if they preserve the information and actions needed for the task.

What Search and Fetch return#

The two newly available APIs serve different purposes. Search finds pages matching a query through DeepSeek’s search engine and returns structured JSON results. Fetch retrieves a single web page and returns its contents as Markdown.

Agents can use Search to find relevant sources, then use Fetch to retrieve selected pages for reading or information extraction. This lets developers bring current web content directly into agent workflows and retrieval-augmented generation applications.

The output formats also shape integration. JSON can feed source-selection logic, while Markdown provides page content in a model-friendly format. Before connecting either endpoint to an existing workflow, check the API documentation for request parameters, response schemas, and limits.

Pricing and free access period#

Each endpoint costs $0.75 per 1,000 calls. At this rate, 10,000 Search calls cost $7.50, and 10,000 Fetch calls cost another $7.50. A workflow that calls each endpoint once per task therefore incurs $15 in endpoint charges per 10,000 tasks, excluding model usage fees.

DeepSeek API users can temporarily use Search and Fetch for free when calling them through DeepSeek Harness, Claude Code, or Codex. The announced deadline is October 31, 2026, at 16:00 UTC, or November 1 at 1:00 a.m. Japan time.

Not all integration methods necessarily qualify for the promotion. Before running large batches, check DeepSeek’s documentation for eligible calling methods and billing treatment. Model inference pricing also needs to be checked separately from Search and Fetch pricing.

Next steps#

Run one representative task with DeepSeek-V4.1-Flash-1017 and record whether it succeeds, how long it takes, its output token count, and the number of Search and Fetch calls. Use those measurements to decide whether to expand the trial.

Source: https://gihyo.jp/article/2026/10/deepseek-v4-1-flash-1017

Contents