DeepSeek V4.1 Flash hands-on tests and comparisons

A demo of a Can Gun shooter made with Gemini 4 Argon, where every skin and the whole structure are pure JavaScript, with a DeepSeek V4.1 Flash test teased.

The author cites a benchmark chart showing V4.1-Flash-1017 at 77% on DeepSWE v1.1 for about $0.20 per task versus Opus 5 at 74% and about $12, arguing cheap runs make parallel agent experimentation affordable.

The release touts faster performance, better project building and shorter cheaper outputs, plus Search & Fetch APIs open to all from $0.75 per 1,000 calls.

Off-peak pricing per 1M tokens is $0.15 input, $0.60 output and $0.003 cached input; peak rates double and weekends are fully off-peak.

The open-weight V4.1 Flash 1017 improves reasoning, coding, agent and vision benchmarks while cutting task time and output length for lower cost.

The author argues DeepSeek's shorter outputs and faster task completion directly cut latency and API cost without sacrificing benchmark reasoning.

The poster claims V4.1 Flash 1017 beat GPT-6.1 Sol xHigh at roughly the price of GPT-6 Luna and Haiku 5.5 while delivering more, with video evidence.

The post says the new model scores 77.0 on DeepSWE, leads on Terminal-Bench, uses fewer tokens, ships open weights and opens a $0.75/1K-call search API to Claude Code and Codex.

The post lists V4.1 Flash 1017 benchmarks and specs: 77 on DeepSWE beating Opus 5, 42.3 on Terminal Bench still behind, 1M context, MIT-licensed free weights, live on app and web.

The author is waiting for DeepSeek V4.1 Flash 1017 on Ollama Cloud and currently maintains open-source projects via the official DeepSeek API, with no concrete data.

A user notes that V4.1 Flash 1017 uses sparse attention attending only to top-k tokens per block instead of full context, cutting compute while keeping coherence.

Versus the previous version, coding, terminal use, agent execution and complex tasks improved notably, beating GPT-5.6-Sol on several specialized benchmarks, though Claude Opus 5 still leads on general reasoning and hard coding.