DeepSeek V4.1 Flash hands-on tests and comparisons

A user reports V4.1 Flash is fast and usually nails small requests in one shot, though vague prompts still go wrong; the real gain is cheap repeated rewrites.

The top five models by OpenRouter token usage over seven days are all Chinese, with DeepSeek at about 2.6x the runner-up; V4.1 Flash output costs $1.2 per million tokens versus $20 for Opus 5.5.

A user runs multi-subagent world-generation experiments with GLM5.3-Flash and DeepSeek-V4.1-1017, posting images but no comparative conclusions.

Official model card reports DeepSWE v1.1 at 77.0 vs Opus 5's 74.0 and Terminal-Bench 4.0 up to 42.3; Search and Fetch APIs cost $0.75 per 1,000 requests, billed from Nov 1.

On the same rocket launch simulation task, V4.1 Flash 1017 cost $0.78 and finished in 3 minutes versus GPT-6.1 Sol's $6.10 and 5 minutes, under one-eighth the cost with a coherent ignition-to-liftoff sequence.

The lead item of a daily AI news roundup says DeepSeek open-sourced V4.1-Flash-1017 under MIT, with faster shorter outputs, beating larger rivals on agentic tasks while sharply cutting per-task cost.

A short reaction post: the author is surprised by the strong 3D reconstruction capability of the open-source DeepSeek V4.1 Flash, though no specific test details or numbers are given.

A Japanese post summarizes the release: 763B parameters, MIT-licensed open weights, 77.0% on DeepSWE v1.1, claimed to beat GPT-6.1 Sol and Claude Opus 5 at about $0.20 per task.

The author asked DeepSeek V4.1 Flash 1017 and GPT-6.1 Sol to each make a 15-second ad video for themselves; both finished, but the styles differed sharply and the author preferred DeepSeek's, since ChatGPT keeps listing points one by one.

A user reports burning 150M tokens on V4.1-Flash-1017 in one morning for ¥6.92, with 99.8% cache hit rate and a goal of 99.9%.

A video post claims V4.1 Flash 1017 beats GPT-6.1 Sol Ultra, without naming the test or sharing numbers.

A user reports V4.1-Flash-1017 built a 3D architecture scene in one shot with pure Three.js, no tool calls or harness, noting better buildings and water but some small issues.