On October 17, DeepSeek released its AI model “DeepSeek-V4.1-Flash-1017.”
The company says it has strengthened the model’s end-to-end project development,
from game creation to 3D modeling.
The update has been applied to the official website, app, and API,
and the model weights have also been released.
What Changes: Better Capabilities and Efficiency Through to Completion
The model card describes this version as a minor update.
DeepSeek says it used reinforcement learning with self-feedback
to improve project development capabilities.
The announcement showcased examples such as a 3D model of the Forbidden City,
a racing game that runs in the browser,
and a recreation of “Portal 2” using Godot.
The company also says the new version reduces task completion time
and output length compared with the previous version.
The official model card reports that the DeepSWE v1.1 score
rose from 74.2 to 77.0, and Terminal-Bench 4.0 from 31.2 to 42.3.
The evaluations used the Mini-SWE-V2 and Claude Code harnesses, respectively.
These figures are results under the evaluation conditions presented by the company.
API Calling Method and Pricing Unchanged; Weights Released Under the MIT License
According to the official change log,
the API calling method and pricing remain unchanged.
Specifying “deepseek-flash” gives access to the new version.
On Hugging Face, the model weights and repository
are released under the MIT License.
Search and Fetch APIs Also Available
On the same day, DeepSeek also released its “Search & Fetch” API,
powered by its in-house search engine.
According to the official guide, search and fetch each cost
$0.75 per 1,000 successful requests.
Failed requests, empty search results, and timeouts are not billed.
Searches and fetches made by DeepSeek API users
through the server-side tools of DeepSeek Harness, Claude Code, or Codex
are currently free.
Billing for these calls is scheduled to begin at 16:00 UTC on October 31
(00:00 Beijing time on November 1).
The unchanged model API pricing and the Search & Fetch pricing
are separate announcements.
LC Signal’s Perspective: Evaluating the Work Through to Completion, Not Just the Answer
Along with game and 3D examples, this update cites improvements
in development task scores, completion time, and output length.
Looking not only at what can be built but also at the time
and processing needed to finish it
seems likely to inform decisions about building AI into development work.
Key Sources
-
DeepSeek official announcement, October 17, 2026:
DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real
-
DeepSeek official change log:
Change Log
-
DeepSeek official model card:
DeepSeek-V4.1-Flash-1017
-
DeepSeek official guide:
Search & Fetch
This article was prepared with AI assistance using publicly available information
as of October 18, 2026.
Announcement dates follow the dates shown on the official pages.
Performance and efficiency improvements are based on the company’s announcement
and evaluation results.
Pricing for the model API and the Search & Fetch API,
and the conditions of the limited-time free access,
are described separately.
“LC Signal’s Perspective” reflects this publication’s interpretation
of the available information.