DeepSeek’s logo, as used on its model pages. Image: DeepSeek
DeepSeek has upgraded its cheapest AI model, and the Chinese lab says the new version now beats Anthropic’s Claude Opus 5 on several coding and agent tests while finishing jobs faster and with shorter answers. DeepSeek V4.1 Flash 1017, named for its October 17 date, went live on Saturday in DeepSeek’s API, website and app, and its weights are free to download on Hugging Face under the MIT licence.
DeepSeek calls it a “minor update” to V4.1 Flash, which it released on September 10. The numbers in its own model card are bigger than that label suggests.
What changed in DeepSeek V4.1 Flash 1017
DeepSeek says the refresh brings “substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks”, and that it “reduces task execution time and output length”, so the same job costs less. It says it tuned the model specifically for its own DeepSeek Harness coding agent and for Tencent’s CodeBuddy and WorkBuddy, and asks Harness users to update to version 0.2.1-rc.1 or later.
The biggest jumps, compared with the September version:
- Terminal-Bench 4.0: 42.3, up from 31.2 (run in the Claude Code harness).
- ProgramBench: 29.1, up from 20.3.
- Apex Bench v1.1: 72.1, up from 64.6.
- Humanity’s Last Exam (no tools): 42.6, up from 36.8.
- DeepSWE v1.1: 77.0, up from 74.2.
The model also accepts images as well as text, and DeepSeek says its scores on chart and visual reasoning tests rose too.
How it compares with Claude Opus 5 and GPT-5.6
DeepSeek’s table puts the new Flash ahead of Claude Opus 5 on DeepSWE (77.0 to 74.0), Automation-Bench (60.5 to 50.3), Agents’ Last Exam (35.4 to 28.6) and Apex Bench (72.1 to 65.8). It also beats OpenAI’s GPT-5.6 Sol on every coding and agent test in the table.
Opus 5 still leads clearly on Terminal-Bench 4.0 (51.8), ProgramBench (37.0), NL2Repo-Bench and Humanity’s Last Exam, where it scores 56.3 against Flash’s 42.6. These are DeepSeek’s own figures, measured with harnesses it chose, and haven’t been checked independently. Among other open models, it scores above Moonshot AI’s Kimi K3 and Z.ai’s GLM 5.3 on every test in the table except Humanity’s Last Exam, where Kimi K3 is slightly ahead.
Price and where you can use it
The price hasn’t changed, and the model name stays deepseek-flash, so existing apps get the upgrade automatically. On DeepSeek’s API it costs $0.30 per million input tokens and $1.20 per million output tokens at peak times, and half that off-peak: $0.15 and $0.60. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, so the whole US working day and most of Europe’s afternoon fall in the cheaper window. Repeated (cached) input costs $0.003 to $0.006 per million tokens.
The context window is 1 million tokens, with answers of up to 384,000 tokens. The API works with both OpenAI and Anthropic-style tools, which is how developers plug it into Claude Code, Codex and OpenCode. Outside DeepSeek, OpenRouter lists it at the off-peak rate of $0.15 and $0.60. If you’d rather run the September version free for testing, it is on NVIDIA’s free API catalog.
The weights can be downloaded anywhere Hugging Face is reachable, and the MIT licence allows commercial use. The model is large: the V4.1 Flash family is a mixture-of-experts design that NVIDIA lists at 552 billion parameters, with about 8 billion active at a time, so it needs serious server hardware rather than a laptop.
DeepSeek also opens its search engine
On the same day, DeepSeek opened its in-house web search engine to developers as a Search and Fetch API, which it says handles “hundreds of millions of search requests” a day and filters out SEO spam and low-quality AI-written pages. Both cost $0.75 per 1,000 successful requests. Search calls made through DeepSeek Harness, Claude Code or Codex are free until billing starts on November 1 at midnight Beijing time (16:00 UTC on October 31).
The update lands a day after Nvidia-backed Reflection released its own 501 billion-parameter open model aimed squarely at China’s best, and while DeepSeek is working on a much larger model trained on Huawei chips.
Why it matters
A free-to-download model that DeepSeek says matches or beats Opus 5 on several agent tests, at a fraction of the price, puts more pressure on the US labs’ pricing for coding agents. The claims are DeepSeek’s own for now, so independent tests over the coming days will show how much of the jump holds up.
Sources: DeepSeek API change log, DeepSeek-V4.1-Flash-1017 model card (Hugging Face), DeepSeek API models and pricing, OpenRouter.
Latest More Labs news
