DeepSeek V4.1 Flash hands-on tests and comparisons

The quiet October 17 update reportedly brings clear benchmark gains and a new search engine capability, with API prices unchanged and shorter task and output times.

The open-weights release keeps flash(sept) pricing and upgrades existing apps automatically, with Search & Fetch calls through the DeepSeek API in Claude Code or Codex free until November 1.

A user says V4.1 Flash is free on OpenCode and near Opus 5 level, so they dropped their $120/month Claude Code and Codex subscriptions.

A user generated a snowy rollercoaster video with V4.1 Flash through DeepSeek Harness and Blender from a one- or two-sentence prompt, calling the result somewhat rough.

A user built an interactive 3D globe dashboard in DeepSeek Harness from one prompt plus a reference image, saying frontend design and detail approach Opus 5.5.

In a 'cyber paternity test' where the model is asked to identify itself, DeepSeek V4.1 Flash reportedly answered that it is GPT-6 Astra.

Same deepseek-flash model on 34 coding/shell/git tasks, two runs each: quality tied at 407/408, but DeepSeek Harness took 1.63x the time and 1.91x the output tokens of minimal Pi.

The author benchmarks DeepSeek V4.1 Flash against Nemotron 3.5 Super VL, finding throughput good but time to first token painful.

Japanese developer npaka published a note article on setting up and trying DeepSeek-V4.1-Flash-1017 inside DeepSeek Harness.

A relayed comparison finds both models take similar time and DeepSeek's output quality holds up, while the whole test costs only 0.44 cents in API usage.

Release roundup: the new build supports a 1M context window and image input, beats Opus 5 on some coding benchmarks, is free to try on chat.deepseek.com, and ships open weights on Hugging Face.

The poster says the new version is clearly optimized for DeepSeek Harness, enough to move from testing-only to regular use.