DeepSeek V4.1 Flash hands-on tests and comparisons

V4.1 Flash 1017 ships MIT open weights and 1M context at $0.15 input / $0.60 output per million tokens off-peak, improves all 11 of its own benchmarks with shorter outputs, and is cheap enough to run agents all day.

Running the same prompt to build an interactive Gummy Bears webpage, Astra6 high cost $19.67 in 20 minutes while V4.1 Flash 1017 cost $0.9 in 58 minutes — roughly 20x cheaper but slower.

After DeepSeek shipped the faster, leaner V4.1-Flash-1017 over the weekend, Bloomberg's Asian chip index fell over 2% to a July low on Monday, led by TSMC and SK Hynix.

A link roundup of DeepSeek's official release note, the Search & Fetch API docs, and the V4.1-Flash-1017 weights on Hugging Face, plus a UTM-tagged agent tooling article.

V4.1-Flash-1017 is live with MIT weights, stronger agentic coding (DeepSWE 77.0, above Opus 5) and a big Terminal-Bench jump at unchanged price with shorter outputs; existing deepseek-flash calls auto-upgrade, and Search/Fetch APIs open at $0.75 per 1,000 calls.

A Japanese post says V4.1-Flash-1017 beats Claude Opus 5 on a self-reported agentic coding benchmark, with cost per task around $0.20, one third of the previous version.

Reports strong coding and agent results with a DeepSWE score of 77, slightly above Opus 5, at very low cost with open weights for self-hosting; complex reasoning still lags, making it suited to cost-sensitive developers and enterprise automation.

LobsterAI's September report says 73.6% of ordinary users rely on a single model and 77% issue AI tasks via WeChat, noting DeepSeek V4.1 Flash cached input costs just 4.6% of the original price.

On the same prompt in max mode, DeepSeek V4.1 Flash 1017 finished in 1h21m for $0.46 while Opus 5.5 took 2h09m for $55.59, roughly 120x cheaper.

Same one-shot prompt for four models: Opus 5.5 took $73.80 and 2h56m, DeepSeek Flash 1017 only $0.34 and 46 minutes, 217x cheaper with nearly matching quality.

The post says V4.1 Flash 1017 scores 1670 on Code Arena WebDev, first among open models, 4 points behind Qwen3.8-Max and 12 ahead of Opus 5 High, at $0.3/$1.2 with MIT open weights and nearly 10x cheaper than Kimi.

A hands-on test used V4.1 Flash 1017 to make a 32-second Blender rocket video in 24 hours (16 for rendering), costing about $2 in API usage across 221M tokens; the author calls it the open-weights king for Blender.