DeepSeek V4.1 Flash Update: Better at Building Games and 3D Scenes, Faster, and Cheaper

DeepSeek has updated V4.1 Flash to V4.1-Flash-1017. The new version is already available on its official website and mobile app.

This update introduces three main improvements:

• Better at building complete mini-projects: Constructing mini-games, 3D scenes, and interactive environments into visible, playable experiences from start to finish.
• Faster and more cost-effective: DeepSeek says tasks take less time and produce less text; its published tests also show lower average costs.
• Opening proprietary search to external software: The technology DeepSeek chat uses to find and read webpages is now available as a separate service that other software can use.

Better at Creating Mini-Games and 3D Scenes

Building a mini-game requires creating visuals, controls, and rules together. A 3D car model alone isn't a race—you still need tracks, driving mechanics, a speedometer, and a finish-line leaderboard. The new Flash version specifically improves connecting all these moving parts into fully functional, runnable mini-projects.

DeepSeek showcased four projects built with the new model, displaying how they perform in action:

3D Forbidden City Architectural Walkthrough: The footage starts with a high-angle aerial view of undulating palace roofs, then swoops down through crimson courtyards and gates to reveal the main hall's structure, snowy courtyards, and light-and-shadow transitions. This type of demonstration serves as a viable prototype for cultural exhibits or interactive 3D spaces.

Web-Based Night Racing Game (Neon Velocity): Rather than generating a static car image, it constructed an entire mini-game loop—from the vehicle showcase on the start screen and clicking to play, to the countdown, accelerating down a neon track while dodging rival cars, live dashboard and minimap displays, and reaching the finish line with a final standings screen.

Portal 2 Mechanics Demo: Built using the Godot game engine, this project re-creates a classic first-person test chamber. Holding a portal device, the player can shoot blue and orange portals onto walls, stepping into one to immediately exit through the other, with an elevator sequence at the end. In game development, creating an initial prototype like this helps validate the core mechanic—stepping in one door and exiting another—before designing full levels.

Valley of Ages: A top-down historical evolution scene. As era tags transition from the Stone Age to Classical Antiquity, the Middle Ages, and the Modern Era, the buildings along the same river valley evolve from primitive thatched huts into classical pillared temples, medieval towns, and finally modern skyscrapers and roads.

For creators building mini-games, interactive exhibits, or educational showcases, the primary value of these capabilities lies in turning an initial concept into a visible, playable prototype that others can test before refining. The official demos showcase selected examples made with development tools; they do not guarantee that any arbitrary prompt will succeed in a single shot, nor does demonstrating basic gameplay mean a complete commercial game is finished.

Running Projects on Your Computer Requires Companion Tools

Brainstorming ideas in a chat window is very different from having AI run a project on your computer. To create files and launch games locally, you need to pair the AI with tools capable of executing these operations and grant them appropriate system permissions.

The complete projects in the demonstrations rely on a division of labor between two roles:

1. The model provides the thinking: It analyzes requirements, designs the necessary file structure, writes the code, and figures out how to troubleshoot execution errors.
2. The companion tool provides the action on your computer: It actually creates code files on the local drive, runs programs, and captures error logs to feed back to the model for adjustments.

When you submit a request, you might only type a sentence or two on the surface, but behind the scenes, the model and companion tool may repeat a cycle of "write code, run it, catch errors, revise, and retry" in an effort to get the program running. The officially recommended companion tool is called DeepSeek Harness, which creates files and executes commands under authorized permissions, allowing written code to run in practice.

How Does It Compare Against Other Models?

According to the benchmark table released by DeepSeek, the new Flash is competitive in coding and automated task evaluations, though it still trails Claude Opus 5 on difficult general questions and some image-analysis tests. Strengths and weaknesses vary across different tasks.

Coding: Wins some, loses some against Claude. In one coding evaluation, Flash scored 77.0, surpassing Claude Opus 5's 74.0, GPT-5.6 Sol's 73.0, Kimi K3's 67.5, and GLM 5.3's 66.9. On another coding test, however, Flash scored 29.1 against Opus 5's 37.0. While competitive in specific programming tasks, it cannot be said to outperform Claude across the board.

Automated tasks: A notable highlight in the table. In an automated task benchmark, Flash achieved 60.5, ahead of Opus 5's 50.3, with the listed versions of GPT, Kimi, and GLM also falling behind. Paired with companion tools, this capability enables AI to execute sequential actions toward a goal and drive a project forward, rather than merely offering suggestions. Real-world performance still depends on the specific task.

Handling high-difficulty problems: Still a noticeable gap behind Opus 5. In evaluations testing complex reasoning without external tools, Flash scored 42.6 compared to Opus 5's 56.3; GPT-5.6 Sol (44.5) and Kimi K3 (43.5) also scored slightly above Flash. Strong coding benchmarks do not imply top-tier performance on every complex question.

Visual analysis: Competitive, but not yet surpassing Opus 5. In a tool-assisted chart analysis test, Flash scored 83.0, close to Opus 5's 84.0 and ahead of GPT-5.6 Sol's 79.9 and Kimi K3's 68.1. While capable of handling visual charts, this single result should not be taken as an overall ranking across all image-analysis tasks.

These benchmark numbers were published by DeepSeek; when choosing a model, consider your specific workload and real-world results.

Shorter Waits and Lower Costs: Efficiency Improvements

DeepSeek stated that the new version completes tasks faster and produces less text. The practical takeaway is less waiting and accomplishing tasks with leaner outputs. Exact savings depend on the nature of the task, and no single speedup percentage was given for all tasks.

For casual daily chat, the main focus is speed and response quality. However, for people paying to connect the model to their own software, the total cost of completing a task also matters.

In a standard coding benchmark released by DeepSeek, the previous model's top configuration had a pass rate of 74.2%, which rose to 77.0% in the new version. Expressed as an intuitive proportional conversion: for roughly every 100 benchmark questions, the new version solves about 77 correctly, compared to about 74 for the previous version.

Across this test set, the average cost per attempt also fell from approximately $0.49 to around $0.20 per question. In other words, accuracy went up while average spending dropped. Both cost figures are estimated values read from official charts; they reflect task costs within this specific benchmark and cannot be taken directly as general API unit pricing, nor do they guarantee identical percentage savings for every project.

Opening DeepSeek Search to Third-Party Software

DeepSeek's own chat app previously supported live web search. What's newly added is making the underlying search service available to other software: developers can integrate it into custom AI assistants, empowering them to discover up-to-date webpages and read content before responding. DeepSeek notes that its proprietary search engine already handles hundreds of millions of queries daily.

The service centers around two primary operations:

• Search (Find pages): Provide keywords, and it queries the web to return relevant webpage titles, URLs, and concise snippets.
• Fetch (Retrieve full text): If search snippets aren't detailed enough, it extracts the main body text of a specified webpage for the AI to analyze.

For instance, a developer building a customer support assistant for a brand's new product lineup could design the workflow like this:

1. When a user asks about a newly released feature, the assistant first invokes Search to locate relevant documentation pages on the official website.
2. If the search snippets already contain a clear answer, the assistant can answer right away, skipping an unnecessary full-page fetch.
3. If the question is complex, the assistant then uses Fetch to retrieve the full user manual text, allowing the model to answer based on the detailed content.

While search provides assistants with fresh information, web sources and AI responses can still contain inaccuracies; critical facts should always be verified against primary sources.

Both API endpoints are publicly priced at 5 CNY (5 yuan) per 1,000 requests:

A key distinction to note: "5 CNY per 1,000 requests" counts individual API calls, not serving 1,000 end-users or conducting 1,000 conversational sessions. For example, if an assistant performs one query by searching once (1 Search) and retrieving one webpage (1 Fetch), that counts as 2 calls—costing 1 fen (0.01 CNY) combined. The cost of the model reading text and generating answers is billed separately. Only successful requests returning results are billed; failed network attempts or empty results incur no charges.

This pay-per-call pricing applies specifically to third-party software accessing the search API, not conversational usage within the DeepSeek app. For API users already accessing DeepSeek's server-side search through DeepSeek Harness, Claude Code, or Codex, search and fetch fees are temporarily waived, with billing scheduled to take effect at 00:00 Beijing Time on November 1, 2026.

Where to Start if You Want to Try It

Users with different needs can explore the new release according to their workflow:

• Casual Chat Users: Simply open the official DeepSeek web interface or mobile app. The model has already been updated automatically, allowing you to gauge the differences in your everyday prompts.
• Testing Local Projects on Your Machine: A standard chat window cannot interact directly with your local files; you will need companion tools capable of running code locally.