DeepSeek V4.1 Flash hands-on tests and comparisons

A Japanese post points to the origin of the urban legend that DeepSeek V4.1 Flash 1017 distilled GPT-6 Astra, noting ~84% output similarity while joking it's coincidence.

The author compares Elorian, DeepSeek V4.1 (1071 checkpoint on Opencode) and Argon 4 on a stickman test, finding DeepSeek slower than expected but very cheap.

A 72-hour tech news roundup confirming DeepSeek-V4.1-Flash-1017 launched on 10/17 with faster shorter outputs, HF weights, and Search/Fetch API at $0.75 per 1k calls.

Gigazine reports that DeepSeek-V4.1-Flash-1017 tops GPT-6 Astra on coding benchmarks while being extremely cheap.

A user asks whether anyone has tried both DeepSeek-V4.1-Flash-1017 and Gemini 4 Argon and which feels better in practice, without offering any test results.

A 25k-star GitHub plugin adds public free channels to DeepSeek Harness, letting users run DeepSeek V4.1 Flash and Kimi K3 without signup or API keys, plus a local usage dashboard.

A hands-on test says DeepSeek V4.1 Flash 1017 Max is about 90% cheaper than GPT-6.1 Sol Ultra with similar runtime (mostly rendering), costing just $0.0044 in API fees while delivering surprisingly good quality.

The post says V4.1 Flash-1017 ranks first among open models and 12th overall on Arena WebDev, ahead of Claude Opus 5 (High) and about 10 points behind GPT-6 Sol (Max), at $0.3 input and $1.2 output per million tokens, roughly a tenth of Kimi K3 (Max)'s cost.

A setup guide recommends official API or Workbuddy plans, Harness/Codex as the harness, Opus 5.5 for planning and GPT 6.1 Sol for review, judging it clearly better than the prior Flash but no match for 5T-class models on complex tasks.

A roundup covering the ~5-hour Oct 15 outage, the Harness team folding into a new product department, the Oct 17 V4.1-Flash-1017 update, and Search/Fetch API pricing at ¥5 per 1,000 calls from Nov 1.

A Japanese note lists $0.15 per million input tokens, MIT-licensed open weights, first place among open models on Code Arena WebDev, and a DeepSWE score of 77.0 beating GPT-5.6 Sol, suggesting teams shift routine tasks to the cheaper model and compare quality.

A user asks whether the latest DeepSeek-V4.1-Flash-1017 is really better than GPT-6 and invites commenters to share their experience, with no concrete findings in the post.