New to this? Read it in simple words
- DeepSeek released a new version of its V4.1 Flash AI model on October 17.
- DeepSeek says it does coding and agent tasks better, faster and with shorter answers.
- Anyone can download the model for free, and the API price stays the same.
- All the test scores are DeepSeek’s own; nobody outside has checked them yet.
- Open-weight model
- An AI model whose trained numbers anyone can download, run and change on their own computers.
- MIT licence
- A short licence that lets anyone use, change and sell the software, including for business.
- API
- A way for other programs to use a company’s AI model over the internet, usually paid per use.
DeepSeek is a Chinese AI company known for open-weight models, which anyone can download and run. Its V4.1 Flash model came out on September 10. Apps that use the model through DeepSeek’s API now get the new version.
What DeepSeek says changed
DeepSeek calls the release a “minor update” to V4.1 Flash. It says the model scores higher on tests, finishes tasks faster and gives shorter answers. Its makers trained it further to build whole software projects, using the model’s own feedback as the reward.
In DeepSeek’s table, the score on the DeepSWE coding test rose from 74.2 to 77.0. Terminal-Bench 4.0, a test of work in a computer terminal, rose from 31.2 to 42.3. On Humanity’s Last Exam without tools, the score rose from 36.8 to 42.6.
The same table still puts Anthropic’s Claude Opus 5 ahead on 7 of the 11 tests, including Terminal-Bench, where it scores 51.8. Bloomberg, The Decoder and IT之家 reported the release on October 17.
From DeepSeek’s model card, announcement and price page of October 17, and The Decoder’s reading of DeepSeek’s cost chart. No independent test yet.
Cost, price and where to get it
DeepSeek’s chart puts one DeepSWE task at just under $0.20 on average, against about $0.50 for the old version, The Decoder reports. At a lower effort setting, the model still scores over 72 for about $0.10.
The API price is unchanged: $0.30 per million input tokens and $1.20 per million output tokens at peak times, and half that off-peak. The model reads text and images, and handles up to 1 million tokens of context.
The weights are on Hugging Face and ModelScope under the MIT licence, which allows commercial use. The model also runs on DeepSeek’s website and in its app.
Search tools, and what is not checked yet
DeepSeek also opened its Search and Fetch tools to all developers. They let AI agents search the web and read pages, using DeepSeek’s own search engine. Each tool costs $0.75 per 1,000 calls.
For DeepSeek API users, calls made through DeepSeek Harness, Claude Code and Codex are free until October 31 at 16:00 UTC. DeepSeek says the model is tuned for its own Harness tool and for Tencent’s CodeBuddy and WorkBuddy.
The Decoder notes that independent benchmarks have not come out yet. Until they do, the gains are DeepSeek’s own claims.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1Primary source · October 17, 2026DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real DeepSeek
- 2
- 3
- 4
- 5
- 6Research · October 17, 2026Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs The Decoder
- 7
Every score and the cost per task come from DeepSeek’s own tests. The outlets reported the release from DeepSeek’s announcement.
DeepSeek’s minor open update promises better coding scores and cheaper tasks at the same price, but nobody outside has tested it yet. How we rate the mood