DeepSeek V4.1 Flash hands-on tests and comparisons

On the same Agents.md task, DeepSeek in DSH took ~35 min, 42M tokens and ~$0.33 versus Opus 5.5 in CC at ~25 min, 15.5M tokens and ~$0.25, yet the author preferred DeepSeek's result.

A user argues against calling a Claude subscription better value, saying they run DeepSeek V4.1 Flash for free or at dirt-cheap prices.

Coin Bureau reports the open-weight model beats its predecessor on nearly all benchmarks while finishing tasks faster with fewer output tokens, under an MIT license with image and text support.

On one prompt, Opus 5.5 finished in 20 minutes for $6 while DeepSeek V4.1 Flash-1017 took an hour for $0.23, roughly 26x cheaper; the author says Opus still wins on polish and control but DeepSeek comes close.

A user relays that V4.1 Flash is very fast with capability close to Opus 5.5, and asks whether it improved for design and whether smart subagents can be built this way.

A user says V4.1 Flash 1017 costs dozens of times less with output close to top models, but its long thinking makes it slower to finish tasks.

On the new Noahbench leaderboard, DeepSeek V4.1 Flash 1017 ties muse wonder for fifth place while costing only one-eighth as much.

In a 20-article Japanese writing test, V4.1 Flash cost ~¥3.3 each versus Opus's ¥37, but took 145s with 22 meaning errors and 12 speaker mix-ups, well above Opus 5.5.

A user had V4.1-Flash-1017 write a complete 3D game from scratch; it delivered bloom post-processing, custom shaders, a boss fight and synthesized BGM in a single 115KB file for about $0.05.

A user runs V4.1 Flash locally on four DGX Sparks at about 80 tok/s, far above 30 tok/s for GLM 5.3, but warns hardware prices have doubled.

The 1017 update rolls out to web, app and API with DeepSWE rising from 74.2% to 77.0% (vs Opus 5's 74.0%) and Terminal-Bench 4.0 at 42.3% vs 51.8%; Search/Fetch APIs open at ¥5 per 1,000 requests and MIT-licensed weights are out.

Post compares search/fetch API pricing: DeepSeek V4.1-Flash-1017 at $0.75 per 1,000 calls versus $10 for OpenAI and Anthropic, $5 for Grok and $14 for Gemini.