DeepSeek V4.1 Flash hands-on tests and comparisons

A Japanese user ran a Miura-ori shade modeling task in DeepSeek Harness: one-shot quality trails Opus 5.5 and feels like early 6 Sol or Opus 5, but modeling costs only ¥70 ($0.44) per hour, 70–80% cheaper than Opus.

On the same prompts, 1017 burned 17.2M tokens for $0.313 versus 784K tokens for $0.047 on the older Flash; 1017 worked longer, used up to 22x more tokens on one test, and showed better taste.

Long analysis of V4.1-Flash-1017: 77% DeepSWE success at ~$0.20/task vs GPT-6.1 Sol's ~75% at $0.65, 763B open weights (~510GB), $0.15/$0.3 pricing.

Having V4.1-Flash-1017 and Opus 5.5 audit V4.1 logs finds a huge share of session budget spent endlessly tweaking minor knobs plus errors, explaining slow progress.

The author generated a promo video for YOOAI.SITE's feature with the model, calling it decent but below Opus 5.5 quality.

The author says 1017 shipped with open weights; an independent bug test still puts Opus 5.5 ahead but by less than expected, so everyday coding may no longer need a frontier model.

The author notes DeepSeek is absent from the fingerprint tool's library so unregistered models map to the nearest match, and says V4.1 Flash was pretrained from scratch on 45T tokens with its own tokenizer and internal multi-teacher distillation.

Same Three.js two-gear meshing prompt: Gemini gets tooth geometry and gear ratio right, while DeepSeek runs but has poor tooth-gap math with overlap or floating.

In a same-prompt train-on-a-figure-8 animation test by @SPAC89, DeepSeek's metallic train and glowing sky city beat GPT-6.1 Sol Ultra's low-poly look at similar time and ~90% lower cost, and was 85%+ cheaper and twice as fast as Gemini 4 Argon.

An English post cites 552B MIT-licensed weights on Hugging Face, 77.0 on DeepSWE, and an effective input price near $0.017 per million tokens at a 91.5% cache hit rate.

A user asks how to choose between Claude Code, Codex and DeepSeek Harness when using DeepSeek-V4.1-Flash-1017.

Links to official release notes, the HuggingFace model card (Opus 5 still higher on HLE, ProgramBench) and pricing, with Search/Fetch API at $0.75 per 1k calls.