DeepSeek updates V4.1 Flash: better at coding and cheaper per task
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
DeepSeek has put out DeepSeek-V4.1-Flash-1017, a refresh of its lightweight model that, by its own measurements, does better at coding and spends less on each job. On DeepSWE v1.1 it climbs from 74.2 to 77.0 and beats Claude Opus 5, which still leads on other tests. The weights can be downloaded under the MIT license, and the web search and reading APIs are now open at $0.75 for every 1,000 calls.
Chinese company DeepSeek released DeepSeek-V4.1-Flash-1017 on Saturday, October 17, a minor update to V4.1 Flash, the lightweight model it launched in September. According to the company, the new version improves on reasoning, coding, agent tasks (AI that does work on its own) and image tests, and it also gets tasks done in less time and with shorter answers, which brings down the cost of each one.
The model is already live on DeepSeek's website, app and API (the connection for developers). The weights can be downloaded from Hugging Face, the platform where AI models are shared, and from ModelScope, under the MIT license, so anyone can get it and use it. The company says it has tuned it to work with its DeepSeek Harness tool, for which it recommends version 0.2.1-rc.1 or later, and with Tencent CodeBuddy and WorkBuddy, among others. In the announcement it shows complete projects built with it, such as a browser racing game or a recreation of Portal 2 made with the Godot engine.
In DeepSeek's results table, the model scores 77.0 on DeepSWE v1.1, a coding test, against 74.2 for the previous version and 74.0 for Claude Opus 5, from Anthropic. On Automation-Bench, an automation test, it reaches 60.5, against 54.8 for the previous version and 50.3 for Claude Opus 5. On other tests in the same table, such as Terminal-Bench 4.0 (51.8 against 42.3) or HLE without tools, Claude Opus 5 is still ahead. These are the company's own evaluations.
The announcement also includes a cost chart. In it, each DeepSWE task comes to about $0.20 on average at its maximum reasoning level, less than a tenth of what Claude Opus 5 costs, which lands a few points lower in that chart. The OpenAI model that comes closest, GPT-6.1 Sol, stays under one dollar per task with a slightly lower result. For the other models, the chart takes its data from the public leaderboard for that test and, for OpenAI's models, from that company's official chart.
DeepSeek is also opening its web search and page reading APIs (Search and Fetch), which run on its own search engine, to every developer. Each one costs $0.75 per 1,000 calls. When they are used through the server-side tools in DeepSeek Harness, in Claude Code, from Anthropic, or in Codex, from OpenAI, they are free for now for its API users; charging will begin on October 31 at 16:00 UTC.
Why it matters · analysis and opinion
The comparison that matters most is not in the table but in the cost chart: an open-weights model that matches the best at coding for a fraction of the price puts pressure on anyone paying for closed models to automate long tasks. For a development team, the question stops being which model wins and becomes which one is enough for each job, and there the cheap option gains ground. It should be read with caution: all the measurements are DeepSeek's, and on several tests Anthropic's model is still ahead. Also, being able to download the weights does not make it free to use: it takes hardware capable of running it, or relying on the company's own API. The sensible move is to try it on real tasks before switching providers.
Official source: DeepSeek · Written with the help of AI: how we make the news
